CN104392240A - Parasite egg identification method based on multi-feature fusion - Google Patents

Abstract

A method for identifying parasite eggs based on multi-feature fusion, including a step of image preprocessing, performing brightness normalization processing on the image information obtained by the micrographic equipment, sharpening processing based on Gaussian filtering, and obtaining the The image of the sharpened egg edge; use the mean shift algorithm to segment the target image to obtain the area judged as an egg; Carry out binarization processing, use the boundary tracking algorithm to acquire the target according to the boundary of the egg area, and obtain the segmented egg image; intercept the specified feature value of the egg image, and store it in the preset feature database; use the relative distance based The KNN (k=3) algorithm substitutes the obtained eigenvalues into the total database, and judges the egg category based on the KNN algorithm. The invention has an accuracy of more than 90% in identifying the eggs, achieving a relatively ideal result.

Description

A method for identifying parasite eggs based on multi-feature fusion

Technical field:

The invention belongs to the technical field of image recognition, and in particular relates to a method for identifying eggs, in particular to a method for identifying parasite eggs based on multi-feature fusion. the

Background technique:

Parasitic diseases are still one of the global public health problems. Microscopic examination of worm eggs is one of the key control techniques, and it is also a basic link in the analysis of parasite morphological characteristics and subsequent biological research. The identification of parasite eggs cannot be carried out with automated instruments like blood cell analysis. For a long time, it has only relied on human eyes to observe and distinguish under a microscope. However, identifying different eggs among numerous parasite samples is a cumbersome task and requires specialized training for technicians. At present, the method of manually smearing specimens and visually identifying them under a microscope is not only cumbersome to operate, and the recognition error varies with the experience and status of the inspectors, but also lacks objectivity and accuracy, and the images, data and results of the inspection specimens are not easy to store. Reproduction and retrieval cannot meet the needs of the development of modern medical information. Therefore, it is necessary to use computer technology to assist in the identification of parasite eggs. the

In 1995, Kong Xiangwei from the Department of Electronics of Dalian University of Technology carried out research on the microcomputer detection and recognition system of worm eggs under a microscope, and the correct recognition rate was close to 92.3%. In 1997, Zhao Ya'e also carried out research on the automatic recognition of 10 parasite egg images, extracted four features of the egg area's perimeter, area, circularity and density for recognition, and the recognition accuracy rate reached 92%. In 2002, Fu Chengbin of Sun Yat-Sen University and others developed a method to identify and classify 7 kinds of trematode adult specimens and extract corresponding 13 morphological features, with an identification accuracy rate of 89.04%. But the image needs to be preprocessed by using Photoshop, AutoCAD, etc. In 2004, Guo Xiaomin used wavelet classification to extract the wavelet variation coefficient feature of insect egg images, and selected probabilistic neural network to classify insect eggs. In 2005, Li Junfeng used the principle of tree layering combined with the principle of minimum distance classification, Bayes discriminant criterion and artificial neural network to build a classifier for identification, and the correct rate reached 94.91%. In 2005, Peng Shexin of Hunan University developed a parasite identification system, and the identification rate can reach 93.0%. In 2007, Luo Zeju, Song Lihong and others proposed a new image feature extraction method and used SVM to automatically recognize and classify nine kinds of parasite egg pictures such as schistosomiasis, and the recognition rate reached 93.9%. the

In foreign countries, in 1996, Sommer C. of Copenhagen Veterinary Laboratory in Denmark used to calculate the amplitude of its Fourier transform to classify, and the accurate recognition rate was 81.5%; Sommer C. extracted the size, Texture and shape features are used for classification and recognition, making the average correct recognition rate reach 91.2%. In 1999, Yang ySll31 of Seoul National University in Korea and others used 7 kinds of 52 human parasite egg images, and used artificial neural network recognition method to extract The four morphological features are classified, detected and recognized, and the recognition accuracy rate reaches 86%. Yang et al. increased the types of eggs and the number of pictures in 2001, and used the above method to classify, detect and identify, and the correct recognition rate was increased to 90.3%. In 2000, G. Theodoropoulos[141] of the National University of Science and Technology in Athens, Greece carried out digital image recognition processing on the images of five kinds of nematode larvae parasitizing in domestic animals, extracted 7 effective feature parameters for classification, and the correct recognition rate was 91.9%. In 2007, Jane S. Fraga and others at the University of Sao Paulo in Brazil used a Bayes classifier to realize the identification of poultry infection parasites, and the identification rate reached 85.75%. In the same year, Sudan's S. Raviraja used statistical methods to classify the classification of blood pictures infected with malaria pathogens. the

Although relevant researchers at home and abroad are trying to use computers to automatically identify parasite pathogens, there are still many difficulties in using computers to automatically identify parasite egg images, which are mainly reflected in the following aspects:

a) There are many types of parasites, making it difficult to find a method suitable for all eggs during image preprocessing, and the shapes and colors of parasite eggs are different, making it difficult to select the characteristics to distinguish various eggs;

b) Due to differences in image capture devices and shooting environments, even if it is the same kind of eggs, the captured images may have differences in the background and the color of the eggs themselves, which will also affect the recognition effect;

c) The parasite eggs themselves have different shapes at different stages, and some even have great differences. For example, roundworm eggs are obviously different in the stage of unfertilized and non-deproteinized membranes and the stage of fertilized and deproteinized membranes. the

Judging from the existing research data, it has been less than fifteen years since the research on the automatic recognition of digital images of parasite eggs has been carried out in China, and it is far from reaching the level of clinical application or automatic instrument recognition. It has a strong color of "pure research". in the following three areas:

a) There are few species that can be identified, and they are often limited to laboratory research on certain types, eggs or adults of a certain type, and systems with too narrow a scope of application are not of high clinical application value. the

b) There are many parts that require manual intervention in the processing of the recognition system. For example, some recognition systems need to use general-purpose software to measure characteristic parameters and enter them into the database before calling them for classification and recognition. It is not an integrated recognition system; some systems require Using the mouse to select the target or determine the starting point of the boundary tracking segmentation is far from the claimed "automatic" recognition. the

c) The extracted image features cannot accurately reflect the image characteristics, so that the feature value ranges of various recognition objects overlap more, and a complex classification algorithm has to be used to improve the recognition rate, which is different from the five-class blood cell analysis currently used in hospitals. Compared with the high-efficiency operation of the instrument, the fastest recognition processing time of 15s in foreign systems still needs to be improved. the

Therefore, it is necessary to develop an automatic egg image recognition system that can adapt to the identification and classification of common human parasite eggs, integrate various identification processing steps, and have a faster processing speed, and can adapt to clinical applications in future automated instruments. application. the

Invention content:

The present invention provides a method for identifying parasite eggs based on multi-feature fusion. The method for identifying parasite eggs based on multi-feature fusion needs to solve the problem of the identification rate of parasite eggs in the prior art. Low, few identification types, and long identification time technical problems. the

A method for identifying parasite eggs based on multi-feature fusion of the present invention comprises the following steps:

a) a step of preprocessing the image, in the step of preprocessing the image, performing brightness normalization processing on the image information acquired by the photomicrograph equipment, and performing grayscale processing on the normalized image, Generate a normalized grayscale image, and then perform sharpening processing on the entire image based on Gaussian filtering to obtain an image with sharpened edges of eggs;

b) A step of performing mean shift on an image with a sharpened edge of eggs to find eggs, in a step of performing mean shift on an image with sharpened edges of eggs to find eggs, using the mean shift algorithm Segment the image to obtain the color feature vector of the above image, plan and find the best target area based on the color feature vector, and obtain the area judged as an egg;

c) A step of target acquisition of the egg image based on the above-mentioned area identified as an egg, in the step of target acquisition, according to the established edge area information of the shape of the parasite egg to be identified, for each candidate Binarize the edge area of the area, use the boundary tracking algorithm to acquire the target according to the boundary of the egg area, and obtain the segmented egg image;

d) a step of intercepting a specified feature value from the segmented egg image and storing it in a preset feature database;

e) a step of classification recognition based on multiple algorithms, in a step of classification recognition based on multiple algorithms, using the KNN (k=3) algorithm based on relative distance, the acquired feature values are substituted into the total database, based on The KNN algorithm judges the type of eggs. the

Further, in a step of performing mean shift on an image with sharpened egg edges to find eggs, the mean shift algorithm is used to segment the target, and the mean shift algorithm is used to segment the target. In this method, the original image is first divided by X×Y to obtain X×Y intersection points, and these intersection points are merged, that is, the Euclidean distance between the color values corresponding to two points is less than a certain threshold, and the The threshold is the color average value of the 5% pixels with the highest brightness and the 5% pixels with the lowest brightness in the image, and then they are combined into one point, so that m points are obtained as the initial point set, and m represents a total of n X×Y points on the picture A collection of pixel points, each pixel point can be expressed as an independent variable _Xi {i=1...n}, the calculation method of the average displacement M of the sample point is:

${\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; h</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 2</span>}}\frac{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{<span\; class="notranslate">\; \&dtri;</span>}{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; E.</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span>}{{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; f</span>}}}_{\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span>}$

Select an initial point in the center of the picture, calculate the average displacement M _{h , U} (x) in the window Sh (x) centered on this point, if the value is not less than a certain threshold, put the window _Sh (x ) to translate M _{h, U} (x), and then repeatedly calculate the average displacement in a new window to obtain a new central value, until M _{h, U} (x) is less than a certain threshold, stop the translation, and obtain a maximum local density position ; Repeat the above steps to obtain m points corresponding to the maximum local density position, and merge these points to obtain the center points of n clusters, that is, the main color of the original image. For each pixel in the original image, According to the Euclidean distance to determine which cluster to belong to, use a one-dimensional histogram to represent the main color information, the abscissa indicates each main color, and the ordinate indicates the proportion of the number of pixels contained in each main color, so that the color characteristics of the image can be obtained vector:

Q={(P _i ,W _i )i=1,…,n}, where P _i =(L ^* _i ,a ^* _i ,b ^* _i ),W _i ∈(0,1], in the above formula, W is the ratio, and P _i is the color value, which is different from the traditional RGB component notation. Here, the color value is represented by the LSH component notation, which is recorded as L _i , a _i , and _bi respectively. Based on the color feature vector Q, the traditional EMD The algorithm can plan the optimal target area, and the general form of the EMD function formula is

$\mathrm{<span\; class="notranslate">\; EMD</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; min</span>}\frac{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}}{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}}$

Among them, the area with the highest EMD similarity with the expected center point is the target area;

Further, in a step of acquiring the target of the egg image based on the area identified as an egg, based on the target area calculated by the above EMD function, that is, according to the established edge area information of the shape of the parasite egg to be identified, Binarize each candidate edge area, and use the boundary tracking algorithm to acquire the target. At the beginning of the algorithm, each pixel is searched in order from top to bottom. Set the sequence array as K, and first search for the first pixel from the upper left. target pixel, set it as k0, then the pixel k0 is the boundary pixel in the upper left corner of the area, that is, the starting point of the search, set the search direction counterclockwise, search in the eight-neighborhood direction, set k0 as the tracking flag, and set k0 is inserted as the first element of the sequence array, search for the next target pixel counterclockwise, and set it to k, if not found, then k is the isolated pixel area; if k is equal to the search start boundary pixel k0, then Continue to judge in order whether there are untracked boundary pixels in other adjacent directions, if not, return to the starting point, the algorithm ends, the boundary pixels in the sequence K form a closed area, enclosing the target area. the

Further, in a step of intercepting the specified eigenvalues of the segmented worm egg images and storing them in the preset feature database, the eigenvalues of the worm egg images are first obtained:

1) Find the smallest square area circumscribed by the edge area, and count the number of pixels to obtain the length (length) and width (width). The length refers to the length of the rectangle circumscribed by the target object, and the width refers to the length of the rectangle circumscribed by the target object;

2) Count the pixels of the target area and the surrounding area of the target to obtain the area and perimeter. The ratio of the area to the smallest circumscribed square is the ovality. The ellipticity refers to the ratio of the area of the target to the area of the circumscribed ellipse ;

3) Area refers to the area of the target object, and perimeter refers to the perimeter of the target object;

4) Obtain RGB components based on the color composition information of the target area; convert the image into grayscale to obtain the statistical histogram of grayscale values, and its mean value is the grayscale value; convert the target into HSV space to obtain HSL components; average Grayscale (grey) refers to the average color of the target object after grayscale; the average red component refers to the computer's expression of color using RGB combination, and the average red component (red) refers to the average value of the R part; The average green component (green) means that the computer uses RGB combination to express the color, and the average green component refers to the average value of the G part; the average blue component (blue) refers to the computer that uses RGB combination to express the color. The average blue component refers to the average value of part B; the average chroma (color) refers to the average value of part H after converting the RGB color model to the HSL color model; the average saturation (saturation) refers to the RGB color model After the model is converted to the HSL color model, the average value of the S part; the average brightness (bright) refers to the average value of the L part after the RGB color model is converted to the HSL color model; the statistical histogram of the gray value refers to the gray value The distribution of 0 to 255 is a vector obtained by staged statistics; the grayscale standard deviation (greyscale) refers to the difference of each local color of the target object; the color weight (weighted) refers to the color expression of the computer using the RGB combination method. The ratio of the average chromaticity to the position coordinates automatically generated by the pixel position, this value is only used for error correction and does not participate in the calculation;

5) After obtaining the feature values, input the preset text format database, and load the subsequent classification recognition algorithm in the form of a table. the

Further, in a step of classification recognition based on multiple algorithms, the KNN algorithm based on relative distance is adopted, and the steps of the KNN algorithm are as follows:

First of all, in order to avoid the calculation of sample distance due to different attribute value ranges, each sample in the eigenvalue database should calculate the maximum value Max[i] and minimum value Min[i] for the i-th dimension attribute value X[i], Then use the formula X[i]=(X[i]-Mini[i])/(Max[i]-Min[i]) to carry out the normalization operation. After the normalization of each attribute of the sample, its value range is [0 ,1], and then construct a data set D={X1,…,XL} according to the eigenvalue database, where X _i ∈ R ⁿ , i=1…L; let the samples have ClassNum classes in total; let C _i represent the i-th class The set of all samples of , and C _i ∩C _j =Ф(i, j=1,...,ClassNum), the sample set can also be expressed as: D=C ₁ ∪C ₂ ∪…∪C _r ;

Suppose the distance between two insect egg samples is Dist, the data set D has m attributes, and its data set composition is R(A ₁ , A ₂ ,...,A _m ), X and Y are two attributes in data set D respectively. samples, then the distance measure formula between X and Y is:

$\mathrm{<span\; class="notranslate">\; Dist</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\sqrt{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; .</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; .</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

The average K-nearest neighbor distance of the i-th class in the test sample is:

$\mathrm{<span\; class="notranslate">\; Avgdis</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; k</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; Dist</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; k</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; ClassNum</span>}$

K _i is the number of samples in C _i , Y is the nearest neighbor of Xj, the relative distance between the test sample X and the training sample Y is: D=Dist(X,Y)/Avgdis(i),Y∈C _i ;

When N=3, as long as the distance between each sample in the data set and the measured sample is calculated, and the three nearest neighbors of the selected test sample are compared, its category can be identified. The classification result is reflected by the score, and the feature of the input image is set as (f1 ,f2,...,fn), the feature of a sample in the database is (x11,x12,...,x1n), then score=s(f1,x11)*s(f2,x12)*.... .*s(fn,x1n);

Here the s function is constructed as follows: Let maxV=max(f1,x11); minV=min(f1,x11), then diff=(maxV-minV)/maxV; s=X*(pow(e,-diff)- 1/e)+B, where e is a natural number, X=(A-B)*e/(e-1), that is, A can be taken to the maximum value, if diff=0, then s is the maximum value A; if diff= 1, then s is the minimum value B. the

Furthermore, before using the mean shift algorithm to segment the target, the picture needs to pre-calculate the color histogram. the

Further, if there are breakpoints in the image due to problems such as over-normalization, resulting in the loss of boundary points and the inability to construct a region frame or a region pixel set, the region growing algorithm is used to make up for it. The specific steps of the region growing algorithm are: first For each region that needs to be segmented, find a seed pixel as the starting point of growth, and then merge the pixels with the same or similar properties as the seed pixel in the neighborhood around the seed pixel into this region, and use these new pixels as a new seed Pixels continue the above process until no more pixels satisfying the condition can be included. the

Further, the parasite is Clonorchis sinensis, Taenia solium, whipworm, pinworm, roundworm, hookworm, Schizocephalon lacunae, Schistosoma japonicum, Fasciola brucei, Lung fluke, or Streptomyces mansoni tapeworm. the

Further, the image to be recognized has rotation invariance, no difference in illumination, and no difference in individuals. the

Specifically, the feature database refers to feature values of the egg part of the image, such as length (length), width (width), ellipticity (ovality), area (area), perimeter (perimeter), average gray scale ( gray), average red component (red), average green component (green), average blue component (blue), average chroma (color), average saturation (saturation), average brightness (bright), gray standard deviation ( greyscale), color weights (weighted), see the attached table for the database format. the

Specifically, the total database refers to a total database composed of parasite characteristic databases with known parameters. the

The present invention also provides a method for identifying parasite eggs based on multi-feature fusion, which is characterized in that it comprises the following steps:

a) A step of image preprocessing, in which the image information obtained by the photomicrograph equipment is subjected to brightness normalization processing, and the normalized image is grayscaled to generate a normalized image. First convert the grayscale image, and then perform sharpening processing based on Gaussian filtering on the entire image to obtain an image with sharpened edges of eggs;

b) A step of finding eggs using artificially assisted recognition. In a step of finding eggs using artificially assisted recognition, the enhanced Grab Cut method is used to segment the edge-sharpened images of eggs, and the user provides a limited box for manual support. , get the color feature vector of the egg image, plan and find the best target area based on the color feature vector, and get the area judged as the egg;

c) A step of performing target acquisition on the egg image based on the above-mentioned area identified as an egg, in the step of target acquisition, according to the established edge area information of the shape of the parasite egg to be identified, for each candidate edge The area is binarized, and the boundary tracking algorithm is used to obtain the target according to the boundary of the egg area, and the segmented egg image is obtained;

d) a step of intercepting a specified feature value from the segmented egg image and storing it in a preset feature database;

e) A step of classification and recognition based on multiple algorithms, in a step of classification and recognition based on multiple algorithms, using the KNN (k=3) algorithm based on relative distance, and substituting the acquired eigenvalues into the 11 kinds of parasitic The total database composed of the insect characteristic database is based on the KNN algorithm to judge the type of insect eggs. the

Further, in a step of using artificially assisted recognition to find eggs, the enhanced Grab Cut method is used to segment the image with edge sharpening of eggs, and the user provides a limited box for manual support, and the parts outside the box are not processed, and the user The trimap T is initialized by setting the background area TB, the foreground area TF is set to be empty, and the unknown area TU is set as the complement of the background area TB. For all pixels in the background area, their Alpha values are set to 0, that is, a = 0; for the pixels in the unknown area, set their Alpha value to 1, that is, a=1, and use the two sets of a=0 and a=1 to initialize the Gaussian mixture model of creating the foreground and background, for the unknown area Set the parameters of the Gaussian mixture model for each pixel n in:

kn=arg min Dn(a _n , k _n , θ, Z _n ),

Gaussian mixture model parameters are obtained from the data of each pixel in the image

θ = arg min U(a, k _n , θ, Z _n ),

Use the minimized energy formula to get the initial split:

min _k E(a, k _n , θ, Z _n ),

Repeat 3 times for boundary optimization. the

Further, in the step of acquiring the target of the egg image based on the above-mentioned area identified as an egg, according to the established edge area information of the shape of the parasite egg to be identified, binarization is performed on each candidate edge area, The boundary tracking algorithm is used for target acquisition. At the beginning of the algorithm, each pixel is searched in order from top to bottom, and the sequence array is set to K. First, the first target pixel is searched from the upper left, which is set to k0, and the pixel k0 is The boundary pixel in the upper left corner of the area is the starting point of the search. Set the search direction to counterclockwise, search in the eight-neighborhood direction, set k0 as the tracking flag, and insert k0 as the first element of the sequence array, press Search the next target pixel counterclockwise and set it to k. If it cannot be found, then k is an isolated pixel area; if k is equal to the search start boundary pixel k0, continue to judge whether there are untracked in other adjacent directions in order If there is no border pixel, it has returned to the starting point, and the algorithm ends, and the border pixels in the sequence K form a closed area, enclosing the target area. the

Further, in a step of storing the specified eigenvalues of the obtained target interception into the feature database, the eigenvalues of the worm egg image are obtained first:

1) Find the smallest square area circumscribed by the edge area, and count the number of pixels to obtain the length (length) and width (width). The length refers to the length of the rectangle circumscribed by the target object, and the width refers to the length of the rectangle circumscribed by the target object;

2) Count the pixels of the target area and the surrounding area of the target to obtain the area and perimeter. The ratio of the area to the smallest circumscribed square is the ovality. The ellipticity refers to the ratio of the area of the target to the area of the circumscribed ellipse ;

3) Area (area) refers to the area of the target object, and perimeter (perimeter) refers to the circumference of the target object;

4) Obtain RGB components based on the color composition information of the target area; convert the image into grayscale to obtain the statistical histogram of grayscale values, and its mean value is the grayscale value; convert the target into HSV space to obtain HSL components; average Grayscale (grey) refers to the average color of the target object after grayscale; the average red component refers to the computer's expression of color using RGB combination, and the average red component (red) refers to the average value of the R part; The average green component (green) means that the computer uses RGB combination to express the color, and the average green component refers to the average value of the G part; the average blue component (blue) refers to the computer that uses RGB combination to express the color. The average blue component refers to the average value of part B; the average chroma (color) refers to the average value of part H after converting the RGB color model to the HSL color model; the average saturation (saturation) refers to the RGB color model After the model is converted to the HSL color model, the average value of the S part; the average brightness (bright) refers to the average value of the L part after the RGB color model is converted to the HSL color model; the statistical histogram of the gray value refers to the gray value The distribution of 0 to 255 is a vector obtained by staged statistics; the grayscale standard deviation (greyscale) refers to the difference of each local color of the target object; the color weight (weighted) refers to the color expression of the computer using the RGB combination method. The ratio of the average chromaticity automatically generated by the pixel position to the position coordinates, this value is only used for error correction and does not participate in the calculation;

5) After obtaining the feature values, input the preset text format database, and load the subsequent classification recognition algorithm in the form of a table. the

Further, in a step of classification recognition based on multiple algorithms, a KNN algorithm based on relative distance is adopted, and the steps of the KNN algorithm are as follows:

First of all, in order to avoid the calculation of sample distance due to different attribute value ranges, each sample in the eigenvalue database should calculate the maximum value Max[i] and minimum value Min[i] for the i-th dimension attribute value X[i], Then use the formula X[i]=(X[i]-Mini[i])/(Max[i]-Min[i]) to carry out the normalization operation. After the normalization of each attribute of the sample, its value range is [0 ,1], and then construct a data set D={X1,…,XL} according to the eigenvalue database, where X _i ∈ R ⁿ , i=1…L; let the samples have ClassNum classes in total; let C _i represent the i-th class The set of all samples of , and C _i ∩C _j =Ф(i, j=1,...,ClassNum), the sample set can also be expressed as: D=C ₁ ∪C ₂ ∪…∪C _r ;

Suppose the distance between two insect egg samples is Dist, the data set D has m attributes, and its data set composition is R(A ₁ , A ₂ ,...,A _m ), X and Y are two attributes in data set D respectively. samples, then the distance measure formula between X and Y is:

$\mathrm{<span\; class="notranslate">\; Dist</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\sqrt{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; .</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; .</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

The average K-nearest neighbor distance of the i-th class in the test sample is:

$\mathrm{<span\; class="notranslate">\; Avgdis</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; k</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; Dist</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; k</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; ClassNum</span>}$

K _i is the number of samples in C _i , Y is the nearest neighbor of Xj, the relative distance between the test sample X and the training sample Y is: D=Dist(X,Y)/Avgdis(i),Y∈C _i ;

When N=3, as long as the distance between each sample in the data set and the measured sample is calculated, and the three nearest neighbors of the selected test sample are compared, its category can be identified. The classification result is reflected by the score, and the feature of the input image is set as (f1 ,f2,...,fn), the feature of a sample in the database is (x11,x12,...,x1n), then score=s(f1,x11)*s(f2,x12)*.... .*s(fn,x1n);

Here the s function is constructed as follows: Let maxV=max(f1,x11); minV=min(f1,x11), then diff=(maxV-minV)/maxV; s=X*(pow(e,-diff)- 1/e)+B, where e is a natural number, X=(A-B)*e/(e-1), that is, A can be taken to the maximum value, if diff=0, then s is the maximum value A; if diff= 1, then s is the minimum value B. the

Further, if there are breakpoints in the image due to problems such as over-normalization, resulting in the loss of boundary points, the region growing algorithm is used to make up for it. The specific steps of the region growing algorithm are: first find each region that needs to be segmented. A seed pixel is used as the starting point of growth, and then the pixels with the same or similar properties as the seed pixel in the neighborhood around the seed pixel are merged into this area, and these new pixels are regarded as new seed pixels to continue the above process until No more pixels satisfying the condition can be included. the

Further, the parasite is Clonorchis sinensis, Taenia solium, whipworm, pinworm, roundworm, hookworm, Schizocephalon lacunae, Schistosoma japonicum, Fasciola brucei, Lung fluke, or Streptomyces mansoni tapeworm. the

Further, the image to be recognized has rotation invariance, no difference to illumination, and no difference to individuals. the

The present invention provides a fully automatic technology for segmenting insect eggs from images to be identified to obtain feature values. Due to the irregular morphological structure, spatial orientation, and many impurities in the specimens of the parasite eggs, it is difficult to fully automatically identify them. The present invention also provides artificial assistance means based on the Grab Cut digital map-matching method. Since the difficulty of the present invention is to automatically segment the target object from the egg picture, the system is required to have rotation invariance to the input image (that is, it is not sensitive to the shooting angle), and it has strong adaptability to illumination differences and individual differences. Egg types are scalable. the

Human parasites can be divided into three categories: single-celled protozoons, multicellular worms (helminth) and arthropods (arthorpods). In my country, there are mainly more than ten kinds of worms, which are mainly divided into nematodes. , I fluke (trematode) and tapeworm (eestode) and acanthocphala (aeanthocphala) four categories. The present invention uses 11 kinds of common human body such as Clonorchis sinensis, tapeworm, whipworm, pinworm, roundworm, hookworm, Schizophrenia ladenum, Schistosoma japonicum, Fasciola brucei, Paragonimia, and Diagonal mansoni. The eggs of parasites are the intended identification objects. the

The invention provides a KNN algorithm-based insect egg classifier, which utilizes feature values to classify. With the continuous development of computer vision and various advanced medical imaging equipment, a single image feature is difficult to fully and accurately express the content of medical images, and multi-feature fusion has become an inevitable way to extract effective features of medical images. On the basis of retaining the original information as much as possible, it overcomes the large and unstable characteristics of the original data, and the extracted fusion features can be effectively used for image recognition. The identification of human parasites depends on the shape, size, content, color and special structure (such as egg shell thickness, specific shape) and other characteristics of eggs. The present invention converts the above-mentioned visual features into 15 classification features of perimeter, area, circularity, color and texture that can be quantitatively extracted by correlation algorithms. the

Compared with the prior art, the present invention has obvious technical effects. (1) The system is not sensitive to the rotation of the picture (that is, it is not only sensitive to the shooting angle), but also has a certain tolerance to the difference in illumination and the individual difference of eggs; (2) In the stage of mathematical feature extraction, the system closely combines the The biological characteristics of the eggs are highly targeted and highly differentiated, so that the identification is more accurate; (3) The system provides two options: automatic segmentation and semi-automatic segmentation, and the semi-automatic segmentation algorithm is more accurate for insects with complex backgrounds. Egg images have good adaptability; fully automatic segmentation has a high degree of automation, which provides convenience for subsequent automatic testing, statistics, analysis and other work. (4) The recognition accuracy of 11 kinds of insect eggs by this system is more than 90%, achieving a relatively ideal result. the

Description of drawings:

Figure 1 is the process of egg image recognition based on automatic segmentation. the

Figure 2 is the picture after edge sharpening. the

Figure 3 is based on the mean shift algorithm to find the target area and calculate the area box. the

Figure 4 is based on the target area and boundary tracking algorithm to extract the target. the

Figure 5 is the feature value database structure and collection use case. the

Fig. 6 is the flow chart of insect egg image recognition based on artificial assistance. the

Figure 7 is a picture after edge sharpening. the

Figure 8 is an artificial frame-assisted segmentation based on the Grab Cut method. the

Figure 9 is based on artificial assistance and boundary tracking algorithm to extract the target. the

Figure 10 is the feature value database structure and collection use case. the

Detailed ways:

The following examples use photomicrographs of a blood smear of Schistosoma japonicum eggs. the

Example 1

A fully automatic identification method based on multi-feature fusion of parasite eggs, including the following process (see Figure 1 for the flow chart):

a) A step of image preprocessing, which is to normalize the brightness of the image information obtained by the photomicrograph equipment, and to grayscale the normalized image to generate a normalized grayscale image, and then to normalize the whole The sharpening process based on Gaussian filtering is performed on the picture to obtain the sharpened image of the egg edge;

b) A step of performing mean shift on the image with sharpened egg edges to find the eggs, using the mean shift algorithm to segment the target image, obtaining the color feature vector of the image, planning and finding the best color vector based on the color vector The target area is the area judged to be an egg;

c) A step of performing target acquisition on the egg image based on the region identified as the egg. Based on the above shape segmentation information, that is, according to the established edge area information of the shape of parasite eggs to be identified, binary processing is performed on each candidate edge area, and the boundary tracking algorithm is used to obtain the target according to the boundary of the egg area. Obtain the segmented worm egg image;

d) a step of intercepting a specified feature value from the segmented egg image and storing it in a preset feature database;

e) a step of classification recognition based on multiple algorithms, in the step of classification recognition based on multiple algorithms according to the present invention, the KNN (k=3) algorithm based on relative distance is adopted to convert the acquired eigenvalues Substitute into the total database containing 11 kinds of parasite characteristic databases, and judge the type of eggs based on the KNN algorithm;

Further, in the step of preprocessing the image, the image information acquired by the photomicrograph equipment is subjected to brightness normalization processing, and the normalized image is grayscaled to generate a normalized grayscale image , and then perform sharpening processing based on Gaussian filtering on the entire picture to obtain an image with sharpened edges of eggs (see Figure 2 for the use case diagram);

Further, a step of performing mean shift on the image with edge sharpening of worm egg, in a step of performing mean shift on the image with edge sharpening of worm egg, using mean shift algorithm to segment the target, in In the process of using the mean shift algorithm to segment the target, the original image is first divided into X×Y to obtain X×Y intersection points, and these intersection points are merged, that is, the color values corresponding to two points The Euclidean distance between them is less than a certain threshold (in order to ensure the software running speed, this threshold is defined as the color average value of the 5% pixels with the highest image brightness and the 5% pixels with the lowest brightness in the present invention), then they are combined into one Points, so that m points are obtained as the initial point set, m represents the set of n pixels X×Y on the picture, and each pixel point can be expressed as an independent variable X _i {i=1...n}, the average value of the sample points The calculation method of displacement M is:

${\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; h</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 2</span>}}\frac{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{<span\; class="notranslate">\; \&dtri;</span>}{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; E.</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span>}{{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; f</span>}}}_{\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span>}$

Select an initial point in the center of the picture, calculate the average displacement M _{h , U} (x) in the window Sh (x) centered on this point, if the value is not less than a certain threshold, put the window _Sh (x ) to translate M _{h, U} (x), and then repeatedly calculate the average displacement in a new window to obtain a new center value, until M _{h, U} (x) is less than a certain threshold, stop the translation, and obtain a maximum local density position ; Repeat the above steps to obtain m points corresponding to the maximum local density position, and merge these points to obtain the center points of n clusters, that is, the main color of the original image. For each pixel in the original image, According to the Euclidean distance to determine which cluster to belong to, use a one-dimensional histogram to represent the main color information, the abscissa indicates each main color, and the ordinate indicates the proportion of the number of pixels contained in each main color, so that the color characteristics of the image can be obtained vector:

P={(P _i ,W _i )i=1,...,n}, where P _i =(L ^* _i ,a ^* _i ,b ^* _i ),W _i ∈(0,1].

Plan the optimal target area based on the color vectors p and q, the formula is

$\mathrm{<span\; class="notranslate">\; EMD</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; min</span>}\frac{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}}{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}}$

Among them, the area with the highest EMD similarity with the expected central point is the target area (see Figure 3 for the use case diagram). the

Traditionally, image segmentation mainly includes four categories: parallel boundary segmentation, serial boundary segmentation, parallel region segmentation and serial region segmentation. Previous studies have pointed out that parallel boundary segmentation is not suitable for images with uniform grayscale changes, while serial region segmentation algorithms (mainly region growing method and relaxation iteration method) are very popular, but it is generally considered that the calculation amount is large, which is not in line with the present invention the limiting conditions. Closed value segmentation needs to set a threshold first, then compare the pixels in the image with the threshold, divide the pixels into target and background and segment them. However, in the detection photos of parasite eggs, due to the pollution of impurities, the naturally formed gray level difference between the target and the background cannot meet the segmentation requirements. the

The mean shift algorithm can find the local maximum point of the density function for each pixel along the shortest path in complex probability distributions such as parasite egg pictures. After testing, using the statistical robustness of the mean shift algorithm and the characteristics of fast convergence along the density gradient direction and the matching of the color histogram algorithm to the target shape can solve the problem of non-rigid target shape change and high tracking complexity . the

The mean shift in the present invention refers to: a non-parameter-based kernel density estimation theory, which is a method for iteratively calculating the extreme point of the probability density function by using the gradient method. The algorithm has the characteristics of no parameters and fast pattern matching, and is an effective target tracking algorithm. the

In the present invention, region growing refers to: combining pixels with similar properties to form a region. The specific steps are: first find a seed pixel for each region that needs to be segmented as the starting point of growth, and then merge pixels with the same or similar properties as the seed pixel in the neighborhood around the seed pixel into this region. The above process continues with these new pixels as new seed pixels until no more pixels satisfying the condition can be included. the

Further, based on the step of acquiring the target of the egg image based on the above-mentioned area identified as an egg, based on the above-mentioned shape segmentation information, that is, according to the established edge area information of the shape of the parasite egg to be identified, for each candidate edge area Carry out binarization processing, and use the boundary tracking algorithm to acquire the target. At the beginning of the algorithm, each pixel is searched in order from top to bottom, and the sequence array is set to K. First, the first target pixel is searched from the upper left, which is set to k0, then the pixel k0 is the boundary pixel in the upper left corner of the area, that is, the starting point of the search, set the search direction counterclockwise, search in the direction of the eight neighbors, set k0 as the tracking flag, and use k0 as the first sequence array Insert an element, search for the next target pixel counterclockwise, and set it to k, if not found, then k is an isolated pixel area; if k is equal to the search start boundary pixel k0, continue to judge other adjacent directions in order Whether there are any boundary pixels that have not been tracked, if not, it has returned to the starting point, and the algorithm ends, and the boundary pixels in the sequence K form a closed area, enclosing the target area (see Figure 4 for the use case diagram) ;

Further, in a step of intercepting the specified feature value for the acquisition target and storing it in the feature database, in the step of intercepting the specified feature value for the acquisition target and storing it in the feature database, first obtain the feature value of the egg image ( See attached table):

1) Find the smallest square area circumscribed by the edge area, and count the number of pixels to obtain the length (length) and width (width). The length refers to the length of the rectangle circumscribed by the target object, and the width refers to the length of the rectangle circumscribed by the target object;

2) Count the pixels of the target area and the surrounding area of the target to obtain the area and perimeter. The ratio of the area to the smallest circumscribed square is the ovality. The ellipticity refers to the ratio of the area of the target to the area of the circumscribed ellipse ;

3) Area (area) refers to the area of the target object, and perimeter (perimeter) refers to the circumference of the target object;

4) Obtain RGB components based on the color composition information of the target area; convert the image into grayscale to obtain the statistical histogram of grayscale values, and its mean value is the grayscale value; convert the target into HSV space to obtain HSL components; average Grayscale (grey) refers to the average color of the target object after grayscale; the average red component refers to the computer's expression of color using RGB combination, and the average red component (red) refers to the average value of the R part; The average green component (green) means that the computer uses RGB combination to express the color, and the average green component refers to the average value of the G part; the average blue component (blue) refers to the computer that uses RGB combination to express the color. The average blue component refers to the average value of part B; the average chroma (color) refers to the average value of part H after converting the RGB color model to the HSL color model; the average saturation (saturation) refers to the RGB color model After the model is converted to the HSL color model, the average value of the S part; the average brightness (bright) refers to the average value of the L part after the RGB color model is converted to the HSL color model; the statistical histogram of the gray value refers to the gray value The distribution of 0 to 255 is a vector obtained by staged statistics; the grayscale standard deviation (greyscale) refers to the difference of each local color of the target object; the color weight (weighted) refers to the color expression of the computer using the RGB combination method. The ratio of the average chromaticity automatically generated by the pixel position to the position coordinates, this value is only used for error correction and does not participate in the calculation;

5) After obtaining the characteristic values, input the preset text format database, and load the subsequent classification recognition algorithm in the form of a table. The database structure is shown in the table below. the

Further, in a step of classification and recognition based on multiple algorithms, the KNN algorithm based on relative distance is adopted, and the specific steps of the KNN algorithm are as follows:

First of all, in order to avoid the calculation of sample distance due to different attribute value ranges, each sample in the eigenvalue database should calculate the maximum value Max[i] and minimum value Min[i] for the i-th dimension attribute value X[i], Then use the formula X[i]=(X[i] -Mini[i])/(Max[i]–Min[i]) to perform the normalization operation. After the normalization of each attribute of the sample, its value range is [0 ,1], and then construct a data set D={X1,…,XL} according to the eigenvalue database, where X _i ∈ R ⁿ , i=1…L; let the samples have ClassNum classes in total; let C _i represent the i-th class The set of all samples of , and C _i ∩C _j =Ф(i, j=1,...,ClassNum), the sample set can also be expressed as: D=C ₁ ∪C ₂ ∪…∪C _r ;

Suppose the distance between two insect egg samples is Dist, the data set D has m attributes, and its data set composition is R(A ₁ , A ₂ ,...,A _m ), X and Y are two attributes in data set D respectively. samples, then the distance measure formula between X and Y is:

$\mathrm{<span\; class="notranslate">\; Dist</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\sqrt{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; .</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; .</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

The average K-nearest neighbor distance of the i-th class in the test sample is:

$\mathrm{<span\; class="notranslate">\; Avgdis</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; k</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; Dist</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; k</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; ClassNum</span>}$

K _i is the number of samples in C _i , Y is the nearest neighbor of Xj, the relative distance between the test sample X and the training sample Y is: D=Dist(X,Y)/Avgdis(i),Y∈C _i ;

When N=3, as long as the distance between each sample in the data set and the measured sample is calculated, and the three nearest neighbors of the selected test sample are compared, its category can be identified. The classification result is reflected by the score, and the feature of the input image is set as (f1 ,f2,...,fn), the feature of a sample in the database is (x11,x12,...,x1n), then score=s(f1,x11)*s(f2,x12)*.... .*s(fn,x1n);

Here the s function is constructed as follows: Let maxV=max(f1,x11); minV=min(f1,x11), then diff=(maxV-minV)/maxV; s=X*(pow(e,-diff)- 1/e)+B, where e is a natural number, X=(A-B)*e/(e-1), that is, A can be taken to the maximum value, if diff=0, then s is the maximum value A; if diff= 1, then s is the minimum value B. the

Furthermore, before using the mean shift algorithm to segment the target, the picture needs to pre-calculate the color histogram. Here, the color histogram calculation standard algorithm module that comes with VC++ is used, and the default parameters are used. the

Further, if there are breakpoints in the image due to problems such as over-normalization, resulting in the loss of boundary points and the inability to construct a region frame or a region pixel set, the region growing algorithm is used to make up for it. The specific steps of the region growing algorithm are: first For each region that needs to be segmented, find a seed pixel as the starting point of growth, and then merge the pixels with the same or similar properties as the seed pixel in the neighborhood around the seed pixel into this region, and use these new pixels as a new seed Pixels continue the above process until no more pixels satisfying the condition can be included. the

Example 2

A kind of artificial assisted identification method of the parasite egg based on multi-feature fusion, it is characterized in that comprising following flow process (flow chart is shown in Fig. 6):

a) A step of image preprocessing. Normalize the brightness of the image information acquired by the micrographic equipment, grayscale the normalized image, generate a normalized grayscale image, and then perform sharpening processing on the entire image based on Gaussian filtering , to obtain the image of the sharpened edge of the egg;

b) A step of using the enhanced Grab Cut method to artificially assist the identification of the edge-sharpened image of the eggs to find the eggs. In the step of target segmentation, the enhanced Grab Cut method is adopted, that is, the user provides a limited box for manual support, more accurately divides the foreground and background, obtains the color feature vector of the image, plans and finds the best target area based on the color vector, Obtain the area judged as insect eggs;

c) A step of performing target acquisition on the egg image based on the region identified as the egg. Based on the above shape segmentation information, that is, according to the established edge area information of the parasite egg shape to be identified, binary processing is performed on each candidate edge area, and the boundary tracking algorithm is used to obtain the target according to the boundary of the egg area. The segmented egg image;

d) a set of steps for intercepting the specified feature value of the segmented egg image and storing it in the preset feature database;

e) a step of classification recognition based on multiple algorithms, in the step of classification recognition based on multiple algorithms according to the present invention, the KNN (k=3) algorithm based on relative distance is adopted to convert the acquired eigenvalues Substitute into the general database, and judge the type of eggs based on the KNN algorithm;

Further, in the step of preprocessing the image, the image information acquired by the photomicrograph equipment is subjected to brightness normalization processing, and the normalized image is grayscaled to generate a normalized grayscale image , and then perform sharpening processing based on Gaussian filtering on the entire picture to obtain an image with sharpened edges of eggs (see Figure 7 for a use case diagram);

Further, a step of using the enhanced Grab Cut method to artificially assist the identification of the edge-sharpened image of the eggs to find the eggs. The user provides a limited box for manual support, and the parts outside the box are not processed, so it is faster and more accurate than the fully automatic calculation of the mean displacement algorithm. The user needs to initialize the tripartite map T by setting the background area TB, the foreground area TF is set to be empty, and the unknown area TU is set as the complement of the background area TB. For all pixels in the background area, their Alpha (transparency) values are set to 0, that is, a=0; for the pixels in the unknown area, set their Alpha value to 1, that is, a=1, and use the two sets of a=0 and a=1 to initialize the Gaussian mixture of the foreground and background Model, set Gaussian mixture model parameters for each pixel n in the unknown area:

kn=arg min Dn(a _n , k _n , θ, Z _n ),

Gaussian mixture model parameters are obtained from the data of each pixel in the image

θ = arg min U(a, k _n , θ, Z _n ),

Use the minimized energy formula to get the initial split:

min _k E(a, k _n , θ, Z _n ),

Repeat the execution 3 times to perform boundary optimization (see Figure 8 for the use case diagram);

For some images with a lot of impurities, the boundary tracking algorithm cannot guarantee to generate a closed boundary, and the algorithm may also get out of control and deviate from the image boundary, especially for some thin boundaries, multiple boundaries, and gray changes near the boundary contour line are not obvious Images are even more so. So the present invention provides manual intervention as an aid. The sets of foreground and background extractions are initialized as the unknown region part and the background region part of the trimap, respectively. The user interaction part in the initialization will affect the final segmentation result. According to the given initial information, K components of the Gaussian mixture model are respectively created for the initial foreground matting area and the background matting area. the

Enhancing the Grab Cut method in the present invention refers to that based on the limited parasite worm egg picture, on the basis of the Graph Cuts method, the enhancement of the algorithm adaptability has been done in the following three aspects: the first, using the Gaussian Mixture Model (Gaussian Mixture Model, GMM) Instead of histograms to describe the distribution of foreground and background pixels, the processing of grayscale images rises to the processing of color images; second, the iterative method is used to obtain the parameters in the Gaussian mixture model instead of a minimum estimate to complete The calculation process of energy minimization; thirdly, through the incomplete marking method, the workload of the user in the interaction process is reduced, and the user only needs to use the rectangular frame to mark the background area. the

Further, the step of performing target acquisition on the insect egg image based on the above-mentioned region identified as the insect egg. Based on the above shape segmentation information, that is, according to the established edge area information of the parasite egg shape to be identified, each candidate edge area is binarized, and the boundary tracking algorithm is used for target acquisition. The algorithm starts from top to bottom. Search each pixel in the following order, set the sequence array as K, first search for the first target pixel point from the upper left, set it as k0, then pixel k0 is the boundary pixel in the upper left corner of the area, that is, the starting point of the search, Set the search direction counterclockwise, search in the eight-neighborhood direction, set k0 as the tracking flag, and insert k0 as the first element of the sequence array, search the next target pixel counterclockwise, and set it to k, if If not found, then k is an isolated pixel area; if k is equal to the search start boundary pixel k0, continue to judge in order whether there are untracked boundary pixels in other adjacent directions, if not, return to the starting point, At the end of the algorithm, the boundary pixels in the sequence K form a closed area, enclosing the target area (see Figure 9 for the use case diagram);

Further, a step of intercepting the specified characteristic value of the acquisition target and storing it in the characteristic database. In the step of intercepting the specified characteristic value of the acquisition target and storing it in the characteristic database, the name and acquisition method of the acquired characteristic value are first required:

1) Find the smallest square area circumscribed by the edge area, and count the number of pixels to obtain the length (length) and width (width). The length refers to the length of the rectangle circumscribed by the target object, and the width refers to the length of the rectangle circumscribed by the target object;

2) Count the pixels of the target area and the surrounding area of the target to obtain the area and perimeter. The ratio of the area to the smallest circumscribed square is the ovality. The ellipticity refers to the ratio of the area of the target to the area of the circumscribed ellipse ;

3) Area (area) refers to the area of the target object, and perimeter (perimeter) refers to the circumference of the target object;

4) Obtain RGB components based on the color composition information of the target area; convert the image into grayscale to obtain the statistical histogram of grayscale values, and its mean value is the grayscale value; convert the target into HSV space to obtain HSL components; average Grayscale (grey) refers to the average color of the target object after grayscale; the average red component refers to the computer's expression of color using RGB combination, and the average red component (red) refers to the average value of the R part; The average green component (green) means that the computer uses RGB combination to express the color, and the average green component refers to the average value of the G part; the average blue component (blue) refers to the computer that uses RGB combination to express the color. The average blue component refers to the average value of part B; the average chroma (color) refers to the average value of part H after converting the RGB color model to the HSL color model; the average saturation (saturation) refers to the RGB color model After the model is converted to the HSL color model, the average value of the S part; the average brightness (bright) refers to the average value of the L part after the RGB color model is converted to the HSL color model; the statistical histogram of the gray value refers to the gray value The distribution of 0 to 255 is a vector obtained by staged statistics; the grayscale standard deviation (greyscale) refers to the difference of each local color of the target object; the color weight (weighted) refers to the color expression of the computer using the RGB combination method. The ratio of the average chromaticity automatically generated by the pixel position to the position coordinates, this value is only used for error correction and does not participate in the calculation;

5) After obtaining the feature values, input the preset text format database, and load the subsequent classification recognition algorithm in the form of a table. The database example is shown in the table below. the

See Figure 10 for the running use case of feature value collection;

Further, in a step of classification recognition based on multiple algorithms, a KNN algorithm based on relative distance is adopted, and the steps of the KNN algorithm are as follows:

First of all, in order to avoid the calculation of sample distance due to different attribute value ranges, each sample in the eigenvalue database should calculate the maximum value Max[i] and minimum value Min[i] for the i-th dimension attribute value X[i], Then use the formula X[i]=(X[i]-Mini[i])/(Max[i]-Min[i]) to carry out the normalization operation. After the normalization of each attribute of the sample, its value range is [0 ,1], and then construct a data set D={X1,…,XL} according to the eigenvalue database, where X _i ∈ R ⁿ , i=1…L; let the samples have ClassNum classes in total; let C _i represent the i-th class The set of all samples of , and C _i ∩C _j =Ф(i, j=1,...,ClassNum), the sample set can also be expressed as: D=C ₁ ∪C ₂ ∪…∪C _r ;

Suppose the distance between two insect egg samples is Dist, the data set D has m attributes, and its data set composition is R(A ₁ , A ₂ ,...,A _m ), X and Y are two attributes in data set D respectively. samples, then the distance measure formula between X and Y is:

$\mathrm{<span\; class="notranslate">\; Dist</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\sqrt{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; .</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; .</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

The average K-nearest neighbor distance of the i-th class in the test sample is:

$\mathrm{<span\; class="notranslate">\; Avgdis</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; k</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; Dist</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; k</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; ClassNum</span>}$

K _i is the number of samples in C _i , Y is the nearest neighbor of Xj, the relative distance between the test sample X and the training sample Y is: D=Dist(X,Y)/Avgdis(i),Y∈C _i ;

When N=3, as long as the distance between each sample in the data set and the measured sample is calculated, and the three nearest neighbors of the selected test sample are compared, its category can be identified. The classification result is reflected by the score, and the feature of the input image is set as (f1 ,f2,...,fn), the feature of a sample in the database is (x11,x12,...,x1n), then score=s(f1,x11)*s(f2,x12)*.... .*s(fn,x1n);

Here the s function is constructed as: let maxV=max(f1,x11); minV=min(f1,x11), then diff=(maxV-minV)/maxV; s=X*(pow(e,-diff)- 1/e)+B, where e is a natural number, X=(A-B)*e/(e-1), that is, A can be maximized, if diff=0, then s is the maximum value A; if diff= 1, then s is the minimum value B. the

Further, if there are breakpoints in the image due to problems such as over-normalization, resulting in the loss of boundary points, the region growing algorithm is used to make up for it. The specific steps of the region growing algorithm are: first find each region that needs to be segmented. A seed pixel is used as the starting point of growth, and then the pixels with the same or similar properties as the seed pixel in the neighborhood around the seed pixel are merged into this area, and these new pixels are regarded as new seed pixels to continue the above process until No more pixels satisfying the condition can be included. the

Multiple features in the present invention are combined in a multiplicative manner. Since two eggs may be similar in other aspects, but have a very large difference in a certain characteristic (in actual microscopic examination, the possibility of the same kind is often denied based on one characteristic), multiplication can better express this great difference , to affect the value of the total score. the

Claims (17)

1. a kind of identification method based on the parasite ovum of multi-feature fusion, comprise a process that utilizes photomicrograph equipment to obtain the image of parasite ovum, it is characterized in that: described process also comprises the steps: a) a step of preprocessing the image, in the step of preprocessing the image, performing brightness normalization processing on the image information acquired by the photomicrograph equipment, and performing grayscale processing on the normalized image, Generate a normalized grayscale image, and then perform sharpening processing on the entire image based on Gaussian filtering to obtain an image with edge sharpening of eggs; b) A step of performing mean shift on an image with a sharpened edge of eggs to find eggs, in the step of performing mean shift on an image with sharpened edges of eggs to find eggs, using a mean shift algorithm to Segment the target image to obtain the color feature vector of the above image, plan and find the best target area based on the color feature vector, and obtain the area judged as an egg; c) A step of target acquisition of the egg image based on the above-mentioned area identified as an egg, in the step of target acquisition, according to the established edge area information of the shape of the parasite egg to be identified, for each candidate Binary processing is performed on the edge area of the egg, and the boundary tracking algorithm is used to obtain the target according to the boundary of the egg area, and the segmented egg image is obtained; d) a step of intercepting specified eigenvalues from the segmented egg images and storing them in a preset eigenvalue database; e) A step of classifying and identifying. In a step of classifying and identifying, the relative distance-based KNN (k=3) algorithm is used to substitute the obtained feature values into the total database, and the egg category is judged based on the KNN algorithm. 2. the identification method of the parasite egg based on multi-feature fusion as claimed in claim 1, is characterized in that: in a step that carries out mean value shift to search for egg to the image of egg edge sharpening, uses mean value shift In the process of using the mean shift algorithm to segment the target, the original image is first divided into X×Y to obtain X×Y intersection points, and these intersection points are merged. , that is, the Euclidean distance between the color values corresponding to two points is less than a certain threshold, and the threshold is the color average value of the 5% pixels with the highest brightness and the 5% pixels with the lowest brightness in the image, then they are combined into One point, so m points are obtained as the initial point set, m represents the set of n pixels in X×Y on the picture, each pixel can be expressed as an independent variable X _i {i=1...n}, the average of the sample points The calculation method of value displacement M is: ${\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; h</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 2</span>}}\frac{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{<span\; class="notranslate">\; \&dtri;</span>}{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; E.</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span>}{{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; f</span>}}}_{\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span>}$ Select an initial point in the center of the picture, calculate the average displacement M _{h , U} (x) in the window Sh (x) centered on this point, if the value is not less than a certain threshold, put the window _Sh (x ) translate M _{h, U} (x), and then repeatedly calculate the average displacement in a new window to obtain a new center value, until M _{h, U} (x) is less than a certain threshold, stop the translation, and obtain a maximum local density position ; Repeat the above steps to obtain m points corresponding to the maximum local density position, and merge these points to obtain the center points of n clusters, that is, the main color of the original image. For each pixel in the original image, According to the Euclidean distance to determine which cluster to belong to, use a one-dimensional histogram to represent the main color information, the abscissa indicates each main color, and the ordinate indicates the proportion of the number of pixels contained in each main color, so that the color characteristics of the image can be obtained vector: Q={(P _i ,W _i )i=1,...,n}, where P _i =(L ^* _i ,a ^* _i ,b ^* _i ),W _i ∈(0,1], In the above formula, W is the ratio, P _i is the color value, and the color value is represented by LSH component notation, which is recorded as L _i , a _i , and _bi respectively. The color feature vector Q uses the EMD algorithm to plan the best target area, and EMD The general form of the function formula is $\mathrm{<span\; class="notranslate">\; EMD</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; min</span>}\frac{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}}{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}}$ Among them, the area with the highest EMD similarity with the expected center point is the target area. 3. the identification method of the parasite egg based on multi-feature fusion as claimed in claim 1, is characterized in that: in a step of carrying out target acquisition to the egg image based on the region of above-mentioned discrimination as egg, according to the established To identify the edge area information of the parasite egg shape, binarize each candidate edge area, and use the boundary tracking algorithm to obtain the target. At the beginning of the algorithm, each pixel is searched in order from top to bottom, and the sequence is set The array is K, first search for the first target pixel point from the upper left, set it to k0, then the pixel k0 is the boundary pixel in the upper left corner of the area, that is, the starting point of the search, set the search direction counterclockwise, eight neighbors Domain direction search, k0 is set as the tracking flag, and k0 is inserted as the first element of the sequence array, the next target pixel is searched counterclockwise, and set to k, if not found, then k is the isolated pixel area ; If k is equal to the search start boundary pixel k0, continue to judge in order whether there are boundary pixels that have not been tracked in other adjacent directions, if not, return to the starting point, the algorithm ends, and the boundary pixel points in the sequence K Form a closed area to enclose the target area. 4. A kind of identification method based on the multi-feature fusion of parasite worm eggs as claimed in claim 1, is characterized in that: in the step of intercepting the specified feature value to the segmented worm egg image, storing it in the preset feature database , first obtain the eigenvalues of the egg image: 1) Find the smallest square area circumscribed by the edge area, count the number of pixels to obtain the length and width, the length refers to the length of the rectangle circumscribed by the target object, and the width refers to the length of the rectangle circumscribed by the target object; 2) Count the pixel points of the target area and the surrounding area of the target to obtain the area and perimeter. The ratio of its area to the smallest circumscribed square is the ellipticity, and the ellipticity refers to the ratio of the area of the target object to the area of the circumscribed ellipse; 3) Area refers to the area of the target object, and perimeter refers to the circumference of the target object; 4) Obtain RGB components based on the color composition information of the target area; convert the image into grayscale to obtain the statistical histogram of grayscale values, and its mean value is the grayscale value; convert the target into HSV space to obtain HSL components; average Grayscale refers to the average color of the target object after grayscale; the average red component refers to the computer's expression of color using RGB combination, the average red component refers to the average value of the R part; the average green component refers to the computer The expression of color adopts the way of RGB combination, the average green component refers to the average value of G part; value; the average chromaticity refers to the average value of the H part after the RGB color model is converted to the HSL color model; the average saturation refers to the average value of the S part after the RGB color model is converted to the HSL color model; the average brightness refers to The average value of the L part after converting the RGB color model to the HSL color model; the statistical histogram of the gray value refers to the vector obtained by staging the distribution of the gray value 0 to 255; the gray standard deviation refers to the target object The difference of each local color; the color weight refers to the ratio of the average chromaticity to the position coordinate automatically generated according to the position of the pixel when the computer adopts the RGB combination method for the expression of the color. This value is only used for error correction and does not participate in the calculation; 5) After obtaining the feature values, input the preset text format database, and load the subsequent classification recognition algorithm in the form of a table. 5. the identification method based on the parasite ovum of multi-feature fusion as claimed in claim 1, is characterized in that: in the step of classification identification, adopts the KNN algorithm based on relative distance, the step of described KNN algorithm is as follows: Each sample in the eigenvalue database should calculate the maximum value Max[i] and minimum value Min[i] for the i-th dimension attribute value X[i], and then use the formula X[i]=(X[i]-Mini [i])/(Max[i]–Min[i]) for normalization operation, the value range of each attribute of the sample is [0,1] after normalization, and then construct the data set D={ X1,...,XL}, where Xi _i ∈ R ⁿ , i=1...L; Let the samples have a total of ClassNum classes; Let C _i represent the set of all samples in the i-th class, and C _i ∩C _j =Ф( i, j=1,...,ClassNum), the sample set can also be expressed as: D=C ₁ ∪C ₂ ∪...∪C _r ; Suppose the distance between two insect egg samples is Dist, the data set D has m attributes, and its data set composition is R(A ₁ , A ₂ ,...,A _m ), X and Y are two attributes in data set D respectively. samples, then the distance measure formula between X and Y is: $\mathrm{<span\; class="notranslate">\; Dist</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\sqrt{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; .</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; .</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$ The average K-nearest neighbor distance of the i-th class in the test sample is: $\mathrm{<span\; class="notranslate">\; Avgdis</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; k</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; Dist</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; k</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; ClassNum</span>}$ K _i is the number of samples in C _i , Y is the nearest neighbor of Xj, the relative distance between the test sample X and the training sample Y is: D=Dist(X,Y)/Avgdis(i),Y∈C _i ; When N=3, as long as the distance between each sample in the data set and the measured sample is calculated, and the three nearest neighbors of the selected test sample are compared, its category can be identified. The classification result is reflected by the score, and the feature of the input image is set as (f1 ,f2,...,fn), the feature of a sample in the database is (x11,x12,...,x1n), then score=s(f1,x11)*s(f2,x12)*.... .*s(fn,x1n); Here the s function is constructed as follows: Let maxV=max(f1,x11); minV=min(f1,x11), then diff=(maxV-minV)/maxV; s=X*(pow(e,-diff)- 1/e)+B, where e is a natural number, X=(A-B)*e/(e-1), that is, A can be taken to the maximum value, if diff=0, then s is the maximum value A; if diff= 1, then s is the minimum value B. 6. The method for identifying parasite eggs based on multi-feature fusion as claimed in claim 1, characterized in that: before using the mean shift algorithm to segment the target, the picture needs to pre-calculate the color histogram. 7. The method for identifying parasite worm eggs based on multi-feature fusion as claimed in claim 1, characterized in that: such as problems such as excessive normalization cause breakpoints in the image, resulting in loss of boundary points and unable to construct a region frame or region The pixel set is supplemented by a region growing algorithm. The specific steps of the region growing algorithm are: first find a seed pixel for each region that needs to be segmented as the starting point of growth, and then combine the seed pixel with the seed pixel in the neighborhood around the seed pixel Pixels with the same or similar properties are merged into this region, and these new pixels are used as new seed pixels to continue the above process until no more pixels satisfying the condition can be included. 8. the identification method of the parasite egg based on multi-feature fusion as claimed in claim 1, is characterized in that: described parasite is clonorchis sinensis, tapeworm, whipworm, pinworm, roundworm, hookworm, Schistocephalus lacunae, Schistosoma japonicum, Fasciola brucei, Paragonimia, or D. mansoni. 9. The method for identifying parasite eggs based on multi-feature fusion as claimed in claim 1, characterized in that: the image to be identified has rotation invariance, no difference to light, and no difference to individuals. 10. A method for identifying parasite eggs based on multi-feature fusion, characterized in that it comprises the following steps: a) A step of image preprocessing, in which the image information obtained by the photomicrograph equipment is subjected to brightness normalization processing, and the normalized image is grayscaled to generate a normalized image. Firstly convert the grayscale image, and then perform sharpening processing based on Gaussian filtering on the entire image to obtain an image with sharpened edges of eggs; b) A step of finding eggs using artificially assisted recognition. In a step of finding eggs using artificially assisted recognition, the enhanced Grab Cut method is used to segment the edge-sharpened images of eggs, and the user provides a limited box for manual support. , get the color feature vector of the egg image, plan and find the best target area based on the color feature vector, and get the area judged as the egg; c) A step of performing target acquisition on the egg image based on the above-mentioned area identified as an egg, in the step of target acquisition, according to the established edge area information of the shape of the parasite egg to be identified, for each candidate edge The area is binarized, and the boundary tracking algorithm is used to obtain the target according to the boundary of the egg area, and the segmented egg image is obtained; d) a step of intercepting a specified feature value from the segmented egg image and storing it in a preset feature database; e) A step of classifying and identifying. In a step of classifying and identifying, the relative distance-based KNN (k=3) algorithm is used to substitute the obtained feature values into the total database, and the egg category is judged based on the KNN algorithm. 11. a kind of identification method based on the multi-feature fusion of parasite eggs as claimed in claim 10, is characterized in that: a) In a step of using artificial aided recognition to find eggs, the enhanced Grab Cut method is used to segment the image with edge sharpening of eggs, and the user provides a limited box for manual support. The part outside the box is not processed, and the user passes Set the background area TB to initialize the trimap T, set the foreground area TF to be empty, set the unknown area TU as the complement of the background area TB, and set their Alpha values to 0 for all pixels in the background area, that is, a=0 ;For the pixels in the unknown area, their Alpha values are set to 1, that is, a=1, and the two sets of a=0 and a=1 are used to initialize the Gaussian mixture model of creating the foreground and background, for the unknown area Set the parameters of the Gaussian mixture model for each pixel n: kn=arg min Dn(a _n , k _n , θ, Z _n ), Obtain the Gaussian mixture model parameters from the data of each pixel in the image θ = arg min U(a, k _n , θ, Z _n ), Use the minimized energy formula to get the initial split: min _k E(a, k _n , θ, Z _n ), Repeat 3 times for boundary optimization. 12. A method for identifying parasite eggs based on multi-feature fusion as claimed in claim 10, characterized in that: in the step of target acquisition of the eggs image based on the region identified as eggs, according to the established To identify the edge area information of the parasite egg shape, binarize each candidate edge area, and use the boundary tracking algorithm to obtain the target. At the beginning of the algorithm, each pixel is searched in order from top to bottom, and the sequence is set The array is K, first search for the first target pixel from the upper left, set it to k0, then the pixel k0 is the boundary pixel in the upper left corner of the area, which is the starting point of the search, set the search direction counterclockwise, eight neighbors Domain direction search, k0 is set as the tracking flag, and k0 is inserted as the first element of the sequence array, the next target pixel is searched counterclockwise, and set to k, if not found, then k is the isolated pixel area ; If k is equal to the search start boundary pixel k0, continue to judge in order whether there are boundary pixels that have not been tracked in other adjacent directions, if not, return to the starting point, the algorithm ends, and the boundary pixel points in the sequence K Form a closed area to enclose the target area. 13. A method for identifying parasite eggs based on multi-feature fusion as claimed in claim 10, characterized in that: in a step of intercepting the specified feature value of the acquisition target and storing it in the feature database, the egg image is first acquired The eigenvalues of 1) Find the smallest square area circumscribed by the edge area, count the number of pixels to obtain the length and width, the length refers to the length of the rectangle circumscribed by the target object, and the width refers to the length of the rectangle circumscribed by the target object; 2) Count the pixel points of the target area and the surrounding area of the target to obtain the area and perimeter. The ratio of its area to the smallest circumscribed square is the ellipticity, and the ellipticity refers to the ratio of the area of the target object to the area of the circumscribed ellipse; 3) Area refers to the area of the target object, and perimeter refers to the circumference of the target object; 4) Obtain RGB components based on the color composition information of the target area; convert the image into grayscale to obtain the statistical histogram of grayscale values, and its mean value is the grayscale value; convert the target into HSV space to obtain HSL components; average Grayscale refers to the average color of the target object after grayscale; the average red component refers to the computer's expression of color using RGB combination, the average red component refers to the average value of the R part; the average green component refers to the computer The expression of color adopts the way of RGB combination, the average green component refers to the average value of G part; value; the average chromaticity refers to the average value of the H part after the RGB color model is converted to the HSL color model; the average saturation refers to the average value of the S part after the RGB color model is converted to the HSL color model; the average brightness refers to The average value of the L part after converting the RGB color model to the HSL color model; the statistical histogram of the gray value refers to the vector obtained by staging the distribution of the gray value 0 to 255; the gray standard deviation refers to the target object The difference of each local color; the color weight refers to the ratio of the average chromaticity to the position coordinate automatically generated according to the position of the pixel when the computer adopts the RGB combination method for the expression of the color. This value is only used for error correction and does not participate in the calculation; 5) After obtaining the feature values, input the preset text format database, and load the subsequent classification recognition algorithm in the form of a table. 14. a kind of identification method based on the parasite egg of multi-feature fusion as claimed in claim 10, it is characterized in that: in the step of a classification identification, adopt the KNN algorithm based on relative distance, the step of KNN algorithm is as follows: First of all, in order to avoid the calculation of sample distance due to different attribute value ranges, each sample in the eigenvalue database should calculate the maximum value Max[i] and minimum value Min[i] for the i-th dimension attribute value X[i], Then use the formula X[i]=(X[i]-Mini[i])/(Max[i]-Min[i]) to carry out the normalization operation. After the normalization of each attribute of the sample, its value range is [0 ,1], and then construct a data set D={X1,…,XL} according to the eigenvalue database, where X _i ∈ R ⁿ , i=1…L; let the samples have ClassNum classes in total; let C _i represent the i-th class The set of all samples of , and Ci∩C _j = Ф(i, j = 1, ..., ClassNum), the sample set can also be expressed as: D = C ₁ ∪C ₂ ∪...∪C _r ; Suppose the distance between two insect egg samples is Dist, the data set D has m attributes, and its data set composition is R(A ₁ , A ₂ ,...,A _m ), X and Y are two attributes in data set D respectively. samples, then the distance measure formula between X and Y is: $\mathrm{<span\; class="notranslate">\; Dist</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\sqrt{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; .</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; .</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$ The average K-nearest neighbor distance of the i-th class in the test sample is: $\mathrm{<span\; class="notranslate">\; Avgdis</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; k</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; Dist</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Y</span>}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; k</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; ClassNum</span>}$ K _i is the number of samples in C _i , Y is the nearest neighbor of Xj, the relative distance between the test sample X and the training sample Y is: D=Dist(X,Y)/Avgdis(i),Y∈C _i ; When N=3, as long as the distance between each sample in the data set and the measured sample is calculated, and the three nearest neighbors of the selected test sample are compared, its category can be identified. The classification result is reflected by the score, and the feature of the input image is set as (f1 ,f2,...,fn), the feature of a sample in the database is (x11,x12,...,x1n), then score=s(f1,x11)*s(f2,x12)*.... .*s(fn,x1n); Here the s function is constructed as: let maxV=max(f1,x11); minV=min(f1,x11), then diff=(maxV-minV)/maxV; s=X*(pow(e,-diff)- 1/e)+B, where e is a natural number, X=(A-B)*e/(e-1), that is, A can be maximized, if diff=0, then s is the maximum value A; if diff= 1, then s is the minimum value B. 15. A method for identifying parasite eggs based on multi-feature fusion as claimed in claim 10, characterized in that: if a breakpoint occurs in the image due to problems such as over-normalization, resulting in the loss of boundary points, then use the region The specific steps of the region growing algorithm are: first find a seed pixel for each region that needs to be segmented as the starting point of growth, and then use the same or similar properties as the seed pixel in the neighborhood around the seed pixel Pixels are merged into this region, and these new pixels are used as new seed pixels to continue the above process until no more pixels satisfying the condition can be included. 16. A method for identifying parasite eggs based on multi-feature fusion as claimed in claim 10, wherein said parasites are Clonorchis sinensis, tapeworm, whipworm, pinworm, roundworm, Hookworm, Schistocephalus lacunae, Schistosoma japonicum, Fasciola brucei, Paragonimiasis, or D. mansoni. 17. A method for identifying parasite eggs based on multi-feature fusion as claimed in claim 10, characterized in that: the image to be identified has rotation invariance, no difference to light, and no difference to individuals.