CN111275697B - Battery silk-screen quality detection method based on ORB feature matching and LK optical flow method - Google Patents

Abstract

The invention discloses a battery silk-screen quality detection method based on ORB feature matching and an LK optical flow method, which comprises the steps of collecting battery silk-screen image data, preprocessing the image data and extracting a battery silk-screen area; modeling is carried out in a rectangular partitioning mode, the template comprises an illustration part template and a character part template, and the characteristics of the template silk screen and the silk screen to be detected are extracted and matched based on an ORB algorithm to realize the positioning of silk screen contents; detecting by using an image difference method based on morphology, and if false alarm occurs, performing secondary detection by using a distortion correction detection method based on an L-K optical flow method; and if false alarm does not occur, outputting a result image and detection data, and executing sorting operation. The method has the advantages of good real-time performance and high detection rate, and the adaptability and the detection rate of non-accurate printing are greatly improved by improving the traditional differential image method and introducing the optical flow method into the field of printing defect detection.

Description

A battery screen printing quality detection method based on ORB feature matching and LK optical flow method

technical field

The invention belongs to the technical field of machine vision automatic surface detection, in particular to a battery silk screen quality detection method based on ORB feature matching and LK optical flow method.

Background technique

Screen printing/barcode detection of batteries is an important step in the battery assembly process. Different types of batteries have differences in character type (Chinese characters, Korean, English and numbers, etc.), character format, illustration content, and barcode format. At present, battery screen printing/barcode is affected by fixtures, equipment, personnel and other related factors, resulting in printing defects. The types of screen printing/barcode defects are mainly divided into: barcode defect/distortion/skew/blur/ghost/dirty/color difference, silkscreen defect/skew/ Fuzzy/ghosting/dirty/color difference, silk screen and barcode information do not match, barcode size and barcode/screen printing position do not meet specifications, etc.

The traditional appearance inspection of battery screen printing mainly relies on the naked eye, magnifying glass and CCD camera for manual inspection. Due to the inevitable influence of human emotions, surrounding environmental noise, work concentration and other factors, it is difficult to ensure the accuracy and real-time performance of the inspection results. The manual detection method is simple to operate, but the detection speed is slow, it is easy to introduce misjudgment, and the detection lacks objective standards.

In order to realize the automatic detection of battery screen printing quality, scholars at home and abroad have conducted a lot of research, and many classic methods have emerged, such as global template matching method, pixel-based hierarchical detection method, neural network algorithm, wavelet transform detection method, Gabor transform algorithms, feature extraction, etc. But there are two main problems with the above method:

1) The algorithm is too complex, and the detection takes a long time, which is not suitable for factories to be used in production line detection;

2) The characteristics of the application objects are relatively simple, and the versatility of the method and the detection ability of complex objects need to be strengthened.

There are many types of battery screen printing detected by the present invention, and the applicable methods for each pattern or text are also different, which requires higher processing precision. Using the previous classical methods can not achieve good results.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide a battery screen printing quality detection method based on ORB (Oriented FAST and Rotated BRIEF) feature matching and L-K (Lukas-Kanade) optical flow method for the above-mentioned deficiencies in the prior art. The appearance quality of silk screen printing on the factory assembly line is tested in real time, with high accuracy, which to a certain extent improves the automation level of battery screen printing quality detection in the domestic battery manufacturing industry.

The present invention adopts following technical scheme:

A battery screen printing quality detection method based on ORB feature matching and LK optical flow method, comprising the following steps:

S1. Collect battery silkscreen image data, and perform preprocessing to extract battery silkscreen area;

S2. Use the rectangular block method to model, including the illustration part template and the text part template, based on the ORB algorithm to extract and match the characteristics of the template silk screen and the silk screen to be tested, so as to realize the positioning of the silk screen content;

S3. Perform the detection based on the image aberration method based on morphology. If there is a false alarm, use the distortion correction detection method based on the L-K optical flow method for secondary detection; if there is no false alarm, output the result image and detection data, and perform sorting operate.

Specifically, step S1 is specifically:

S101, creating a crossed rectangular window according to the size of the original image, and cropping the original image, the purpose of which is to obtain the boundary information of the screen printing area of the cell;

S102, use the Ostu algorithm to calculate the binarization threshold, initially separate the silk screen area from other backgrounds, and appropriately offset the threshold value in combination with the grayscale features of the image to accurately separate the cell silk screen area and other backgrounds, and remove them by morphological opening operation Edge burr to eliminate small edge segmentation errors;

S103, using a relevant algorithm to perform minimum circumscribed rectangle fitting on the region extracted in step S102, and trimming the original image according to the fitting result to obtain an accurate battery silk screen area;

S104, performing grayscale correction on the image to be tested, and adjusting it to the same grayscale level as the template image.

Further, step S104 is specifically:

S1041. Grayscale the image, convert the original three-channel image into a single-channel grayscale image, and use the psychological grayscale formula:

gray=0.299*red+0.587*green+0.114*blue

S1042, divide the foreground region R ₁ and the background region R ₂ by using a threshold segmentation method; calculate the gray mean value of the foreground region and the background region respectively, perform the above operations on the image to be tested and the template image respectively, and calculate the gray mean value as follows:

Among them, R _i is the region extracted after threshold segmentation, p is a pixel in the region, g(p) is the gray value at point p, and F is the total number of pixels in R _i ;

S1043. Perform grayscale transformation on the image to be tested to obtain the mean value M1 of the foreground grayscale of the template, the mean value _{M2 of} the background grayscale of the template, the mean value T1 of the foreground grayscale of the image to be tested, and the mean value T2 _of the background _grayscale of the image to be tested, and perform grayscale Transform, calculate the gray scale scaling coefficient Mult and the gray scale translation coefficient Add respectively, and use these two coefficients to map and correct the original gray scale of the image to be tested, as follows:

Mult=(M ₁ -M ₂ )/(T ₁ -T ₂ )

Add=M ₁ -Mult*T ₁

g'=g*Mult+Add

Among them, Mult is the gray scale scaling coefficient, Add is the gray scale translation coefficient, and g' is the new gray value after mapping.

Specifically, step S2 is specifically:

S201, oFAST feature point extraction;

S202, rBRIEF feature descriptor structure;

S203, feature point matching.

Further, in step S201, the gray value of point P is compared with the gray value of 16 pixel points in the neighborhood with P as the center and the radius of 3. If the pixel value of point P is the same as that of n pixels in the neighborhood on the circle The subtraction of the pixel values of consecutive pixel points is greater than the threshold t, and the point P is a feature point.

Further, in step S202, N point pairs are selected around the key point P, the comparison results of the N point pairs are combined as descriptors, and all point pairs are compared to generate a binary string of length n.

Further, in step S203, the Hamming distance between the descriptors is used as the similarity measure, and the Hamming distance between the two binary strings of equal length is the number of different characters in the corresponding positions of the two strings. By setting the similarity According to the corresponding relationship between the ORB feature points in the template image and the image to be tested, the template and the image to be tested are registered, and each rectangular sub-template in the template image is matched with the image to be tested. Matching is done on each part of the measured image.

Specifically, in step S3, the illustration part detection is specifically:

S3011, obtaining the corresponding point relationship according to the matching result, calculating the rigid body transformation matrix, aligning the sub-template image with the image to be tested, and using the threshold segmentation method to segment the detection content from the background;

S3012, perform morphological corrosion and expansion treatment on the divided screen printing content, and obtain multi-print defects and missing printing defects through the positive and negative difference between the product to be tested and the template. The multi-print defects are: T _(corrosion) - M _(expansion) ; missing printing defects are: M _(corrosion) -T _(expansion) ;

S3013, perform connectivity analysis and threshold screening on the difference image after the difference in step S3012, and the area exceeding the threshold is the defect area;

S3014 , taking the center of the defect as the center of the circle, and marking the defect by using the smallest circumcircle of the defect.

Specifically, in step S3, the text part defect detection process is as follows:

S3015, affine transformation registration;

S3016, perform skeleton extraction processing on the text, after the extraction, the text lines are unified into a single-pixel width, and only the topology structure is retained;

S3017, performing a rice-shaped translation process on the extracted text, namely up, down, left, right, upper left, upper right, lower left, and lower right in 8 directions, and then performing image difference to obtain 8 difference images;

S3018, performing an intersection operation on the 8 difference images obtained in step S3017, and the final reserved area is regarded as a real defect area;

S3019. Connectivity analysis threshold screening and defect marking.

Specifically, in step S3, the template image and the image to be tested are regarded as continuous frame images in the video, and the irregular and small distortions that appear in the image to be tested are regarded as local micro-motion, and the detection process is as follows:

S3021. The optical flow field between the two images can be calculated by using the LK optical flow method. The optical flow vector V _i,j at any position (i, j) of the optical flow field is composed as follows:

V _i,j =(x,y)

Among them, the x value represents the row offset of the image to be tested at point (i, j), and the y value represents the column offset;

S3022, decompose the optical flow vector V _i,j to obtain its components x and y to form two data sets {x} and {y};

S3023, perform threshold segmentation on the image to be tested, obtain the screen printing foreground area T, and count the coordinates (i, j) of all points in the area T;

S3024, calculate the corresponding point set, the image to be tested is compared with the template image, and the pixel at the coordinate (i, j) is moved to the coordinate (i+x, j+y) due to distortion;

S3025, using (i, j) and (i+x, j+y) as the corresponding point sets, calculate the projection transformation matrix A, and perform projection transformation on the silk screen to be measured with this matrix, so as to realize the correction of small deformation;

S3026 , performing shadow difference processing on the corrected image to be tested and the template image again, performing defect separation and extraction processing, and making final marks.

Compared with the prior art, the present invention at least has the following beneficial effects:

The invention is a battery silk screen quality detection method based on ORB feature matching and LK optical flow method, which has good real-time performance and higher detection rate than similar methods, and can correct irregular distortion by introducing optical flow method, which greatly enhances the accuracy of The adaptability of screen printing quality inspection.

Further, after the image is preprocessed, the background redundancy is eliminated, the image quality is improved, and the subsequent detection can be carried out.

Further, after grayscale correction, the silkscreen to be tested and the template silkscreen can be adjusted to basically the same gray level, which greatly reduces the influence of illumination changes on detection.

Further, after block matching, the positioning accuracy of each sub-template can be improved, preventing false positives caused by the position and angle errors between the contents of the screen printing blocks, and the matching based on ORB features has high accuracy and real-time performance. .

Further, image feature points can be quickly extracted, which is efficient and time-saving, and provides prerequisites for constructing rBRIEF descriptors.

Furthermore, a feature descriptor in the form of a binary string is constructed, which has a simple form, greatly saves storage space, and reduces the time-consuming of matching. At the same time, the descriptor is invariant to rotation, which improves the adaptability of the algorithm.

Further, by calculating the Hamming distance to measure the similarity of the binary descriptors, it is simple and efficient, and the corresponding points between the template and the image to be tested are screened out to complete the matching and positioning.

Further, through image aberration and corrosion expansion processing in the illustration detection, contour false defects can be effectively eliminated, and the real defects of the illustration can be accurately separated.

Further, in the text detection, by extracting the text skeleton, the false positives caused by the uneven thickness of the text lines can be effectively avoided, and the false positive interference caused by the poor text position consistency can be further reduced by the "mi" font translation strategy.

Further, the template and the silk screen to be tested are regarded as continuous frame images, and the slight distortion in the silk screen can be accurately corrected through L-K optical flow calculation and projection transformation, eliminating the false alarm interference of such distortion.

To sum up, the present invention has good real-time performance and high detection rate. By improving the traditional aberration method and introducing the optical flow method into the field of printing defect detection, the adaptability and detection rate of inaccurate printing are greatly improved.

The technical solutions of the present invention will be further described in detail below through the accompanying drawings and embodiments.

Description of drawings

Fig. 1 is the flow chart of the present invention;

Figure 2 is a preprocessing image, wherein (a) is the original image, and (b) is the preprocessing image;

Fig. 3 is a preprocessing flow chart;

Fig. 4 is a flow chart of cross segmentation, wherein (a) is cross windowing, (b) is Ostu binarization, (c) is Ostu threshold shift, and (d) is morphological opening operation;

5 is a grayscale correction diagram, wherein (a) is a template image, (b) is an image to be tested before correction, and (c) is an image to be tested after correction;

6 is a block diagram;

Fig. 7 is a flowchart of template matching based on ORB feature;

Figure 8 is a schematic diagram of oFAST detection;

Fig. 9 is oFAST detection result graph;

Figure 10 is a schematic diagram of rBRIEF point pair;

Figure 11 is the ORB algorithm matching result diagram;

12 is a flowchart of defect separation and extraction;

Figure 13 is a flowchart of illustration detection;

Figure 14 is a flow chart of text detection;

15 is a schematic diagram of a translation strategy;

Figure 16 is a schematic diagram of twisted silk screen printing, wherein (a) is the comparison of the coincidence degree, and (b) is the difference shadow detection result;

17 is a schematic diagram of an optical flow field;

Fig. 18 is the L-K optical flow method detection flow chart;

FIG. 19 is a schematic diagram of screen printing calibration.

Detailed ways

The invention provides a battery silk screen quality detection method based on ORB feature matching and LK optical flow method. First, a preprocessing stage is performed: the original image collected by the camera contains a lot of background interference and noise, and the inclination degrees are different. Therefore, it is necessary to perform cropping correction, affine transformation and grayscale correction at this stage; then the template matching stage: based on the ORB algorithm, the characteristics of the template silkscreen and the silkscreen to be tested are extracted and matched to realize the positioning of the silkscreen content; the final defect Separation and extraction stage: In this stage, the difference shadow method based on morphology is used for preliminary detection, and the distortion correction detection method based on L-K optical flow method is used for the secondary detection of screen printing with irregular distortion.

Please refer to FIG. 1 , a battery silk screen quality detection method based on ORB feature matching and LK optical flow method of the present invention includes the following steps:

S1, image preprocessing

The original image collected by the camera contains a lot of background interference and noise, and the degree of inclination is different, so it is necessary to perform cropping correction, affine transformation, and grayscale correction at this stage; as shown in Figure 2, ( a) is the original image, (b) is the preprocessed image.

Please refer to Figure 3. The original image field of view includes interference areas such as the background stage, fixtures, and metal electrode sheets, which affect subsequent detection. Therefore, preprocessing is required to extract the cell silk screen area. The extraction steps are:

S101, creating a crossed rectangular window according to the size of the original image, and cropping the original image, the purpose of which is to obtain the boundary information of the screen printing area of the cell;

Referring to FIG. 4 , through the above processing, the information of the upper, lower, left and right borders of the screen printing area of the battery is successfully obtained, and a large amount of background redundancy irrelevant to detection is removed.

S102, use the Ostu algorithm to calculate the binarization threshold, initially separate the silk screen area from other backgrounds, and appropriately offset the threshold value in combination with the grayscale features of the image to accurately separate the cell silk screen area and other backgrounds, and finally perform a morphological opening operation Remove edge burrs and eliminate small edge segmentation errors. The processing effect is shown in Figure 4;

S103 , using a correlation algorithm to perform minimum circumscribed rectangle fitting on the extracted area, and crop the original image according to the fitting result to obtain an accurate battery silk screen area;

S104, perform grayscale correction on the image to be tested, and adjust it to the same grayscale level as the template image, so as to solve the influence of unstable illumination and improve the detection accuracy, which is very important for subsequent detection.

The grayscale correction method adopted by the present invention is specifically:

S1041. First, the image is grayscaled, and the original three-channel image is converted into a single-channel grayscale image. The human eye is most sensitive to green and the least sensitive to blue, so the psychological grayscale formula is adopted:

gray=0.299*red+0.587*green+0.114*blue (1)

S1042, divide the foreground area R ₁ and the background area R ₂ by using a threshold segmentation method; calculate the gray mean value of the foreground area and the background area, respectively, perform the above operations on the image to be tested and the template image, and the gray mean value calculation formula is shown in formula 2 :

Among them, R _i is the region (foreground or background region) extracted after threshold segmentation, p is a pixel in the region, g(p) is the gray value at point p, and F is the total number of pixels in R _i ;

S1043. Perform grayscale transformation on the image to be tested, and through the above calculation, four key values are obtained: the grayscale mean value M ₁ of the template foreground (screen printing content), the template background gray mean value M ₂ , the foreground (screen printing content) gray value of the image to be tested The gray scale mean T ₁ and the background gray mean T ₂ of the image to be tested are based on these four data, and the gray scale transformation is performed to calculate the gray scale scaling coefficient Mult and the gray scale translation coefficient Add respectively, and use these two coefficients to be tested. The original grayscale of the image is mapped and corrected, and the formulas are as follows:

Mult=(M ₁ -M ₂ )/(T ₁ -T ₂ ) (3)

Add=M ₁ -Mult*T ₁ (4)

g'=g*Mult+Add (5)

Among them, Mult is the gray scale scaling coefficient, Add is the gray scale translation coefficient, g' is the new gray value after mapping, and the effect of the original image after the above processing is shown in Figure 5, where (a) is the template image , (b) is the image to be tested (before correction), (c) is the image to be tested (after correction); it can be seen from Figure 5 that the contrast, brightness and other characteristics of the image to be tested after grayscale transformation are very different from the template image. close, which provides good conditions for the implementation of subsequent template matching algorithms and defect separation and extraction.

S2, template matching stage

Based on the ORB algorithm, the characteristics of the template silkscreen and the silkscreen to be tested are extracted and matched to realize the positioning of the silkscreen content;

Please refer to Fig. 6. After the preprocessing, it is necessary to identify and locate each part of the silkscreen content on the image to be tested. The running time of ORB feature detection is much better than that of the SIFT algorithm and the SURF algorithm, and can be applied to real-time feature detection. Therefore, the present invention The template matching algorithm based on ORB features is adopted. At the same time, due to the large size of the screen printing image of the battery, the poor consistency of the screen printing content, and the rotation and position errors between different screen printing contents, the present invention adopts a rectangular block method for modeling.

Including the illustration part template and the text part template, each sub-template is established in sequence, and each sub-template will be covered by a rectangular mask with a gray value of 0 after the current modeling is completed, and then continue modeling, so each sub-template in the figure will be covered. There is overlap between the rectangular boxes; then the ORB template matching algorithm is used for matching positioning (Oriented FAST and RotatedBRIEF ORB), which is mainly divided into the following three steps:

(1) oFAST feature point detection;

(2) rBRIEF feature description;

(3) Matching of feature points.

Please refer to Figure 7, the detection process and technical details are as follows:

S201, oFAST feature point extraction;

Please refer to Figure 8. oFAST corner detection is a fast corner feature detection algorithm. Specifically, it judges the grayscale of 16 pixels of the point of interest and its surrounding area. Usually, a circular area is selected for this area. The judgment method is as follows :

Compare the gray value of point P with the gray value of 16 pixels in the neighborhood with P as the center and a radius of 3. If the pixel value of point P and the pixel value of n consecutive pixels in the neighborhood on the circle The subtraction is greater than the threshold t, then the point P is considered to be a feature point.

The oFAST algorithm has good scale invariance and rotation invariance. The feature points of a sub-image extracted by the oFAST algorithm are shown in Figure 9.

S202, rBRIEF feature descriptor structure;

After obtaining the feature points, we need to describe the attributes of these feature points in a certain way. This description is called a feature descriptor. The ORB algorithm uses the rBRIEF algorithm to calculate the descriptor. The operation is to select N point pairs in a certain pattern around the key point P, and combine the comparison results of these N point pairs as a descriptor. All point pairs are compared, and a binary string of length n is generated. Generally, N takes 128, 256 or 512.

Please refer to Figure 10, take the key point P as the center of the circle, and take d as the radius to make a circle O, and select N point pairs in a certain pattern in the circle O. Assuming that the currently selected 4 point pairs are shown in Figure 10, at this time, compare the relative sizes of the gray values of A _i and B _i in each point pair, and obtain a binary description substring of length 4, such as 1011. The coordinate system established by ORB when calculating the rBRIEF descriptor takes the key point as the origin, and establishes a 2-dimensional coordinate system with the line connecting the key point and the centroid of the point region as the X-axis, so that the descriptor has rotation invariance.

S203, feature point matching;

The biggest feature of ORB algorithm is the fast calculation speed. The present invention uses the Hamming distance between descriptors as a similarity measure when matching, and the Hamming distance between two binary strings of equal length is the number of different characters in the corresponding positions of the two strings, and also represents the two binary strings. The similarity of the strings is selected by setting the similarity threshold to filter the matching point pairs. The matching results obtained according to the Hamming distance are shown in Figure 11.

After the above matching results are obtained, the template and the image to be tested are registered according to the corresponding relationship between the ORB feature points in the template image and the image to be tested. So far, the template matching stage is completed, and each rectangular sub-template in the template image is matched with each part in the image to be tested.

S3. Defect separation and extraction based on optical flow method and morphological image aberration method

First, all silk screen prints are detected by the morphological-based difference shadow method. If the detection results are normal, the sorting operation is performed; if there are a large number of false positives in the detection results, the distortion correction detection method based on the L-K optical flow method is used for the second time. detection to eliminate false positives.

Please refer to Figure 12. The flowchart and technical details of the defect separation and extraction stage are as follows:

S301. An image aberration defect analysis algorithm based on morphology;

The screen printing content is divided into two types: illustration and text, and sub-algorithms are designed according to their respective characteristics to achieve the best detection effect.

Please refer to Figure 13. The ink in the illustration area is concentrated and the lines are simple. According to this feature, the following detection process is designed:

S3011, the corresponding point relationship can be obtained according to the matching result, and the rigid body transformation matrix is calculated accordingly, the sub-template image is aligned with the image to be tested, and the detection content is segmented from the background by a threshold segmentation method;

S3012 , performing morphological corrosion and expansion processing on the segmented silk-screen content, in order to prevent false contour defects caused by misalignment of pattern boundaries. Multi-print defects and missing-print defects can be obtained by the difference between the positive and negative of the product to be tested and the template, as shown below:

Multiprint Defects:

T _(corrosion) -M _(expansion)

Missing printing defects:

M _(corrosion) -T _(expansion)

S3013, performing connectivity analysis and threshold screening on the difference image after the difference, and the area exceeding the threshold is the defect area;

S3014, defect location and marking: take the center of the defect as the center of the circle, and use the minimum circumcircle of the defect to mark the defect.

The above is an introduction to the detection method of the illustration part, while the lines of the text part are fine and complex, and the defects are more subtle.

Please refer to Figure 14, the process of the text part defect detection method is as follows:

S3015, affine transformation registration: same as illustration detection;

S3016, skeleton extraction: the reason for using skeleton extraction is that the text lines are relatively fine and complex, and the thickness of the lines is uneven. If the expansion corrosion treatment is used, the topological structure of the text will be destroyed, and the parameters are difficult to adjust, so the present invention performs skeleton extraction processing on the text. After extraction, the text lines are unified to a single-pixel width, and only their topology is retained without being affected by the thickness;

S3017, image translation difference: perform "m" font translation processing on the extracted text, namely up, down, left, right, upper left, upper right, lower left, and lower right 8 directions, and then perform image difference to obtain 8 poor image. The reason for the translation error is that the position consistency of the silk screen content is poor, and there will be a small amount of position deviation. A single direct error will bring a large number of false alarms, as shown in Figure 15;

S3018, difference shadow intersection operation: perform the intersection operation on the 8 difference images obtained in the previous step, and the final reserved area is regarded as the real defect area;

S3019. Connectivity analysis threshold screening and defect marking: same as illustration detection;

S302. Distortion correction detection algorithm based on L-K optical flow method

After testing, most of the defects can be detected normally, but for a very small part of screen printing products, due to the printing process, the printed content will be slightly distorted and deformed, and it cannot cause abnormal human vision. This category belongs to the category of good printing products, but such products are in the 3.1 The detection method, that is, the differential shadow method, will be judged as a defective product, which is a difficult point in this detection. Figure 16 shows a schematic diagram of a small amount of distortion. It can be seen from the observation that in the picture (a), the template silk screen and the under test The screen printing has a low degree of coincidence and is irregularly distorted, but the screen printing to be tested belongs to the category of good products. After direct detection, a large number of false positives have been generated, as shown in the figure (b) on the right. Therefore, the present invention proposes a distortion correction detection algorithm based on the L-K optical flow method, and creatively introduces the optical flow method into the field of surface defect detection, which can accurately detect products with slight distortion and prevent false alarms.

The testing process and technical details are as follows:

Optical flow is the instantaneous speed of pixel motion of space moving objects on the observation imaging plane. The optical flow method uses the changes of pixels in the image sequence in the time domain and the correlation between adjacent frames to find the previous frame and the current frame. A method for calculating the motion information of objects between adjacent frames. Usually, the instantaneous rate of change of gray level on a specific coordinate point of the two-dimensional image plane is defined as the optical flow vector. When the time interval is small (such as between two consecutive frames of the video), it is also equivalent to the displacement of the target point. The schematic diagram of the optical flow field calculated from the silk screen image is shown in Figure 17.

Because it is more convenient to apply to a set of points in the input image, the present invention selects the L-K optical flow method, which is based on the following three assumptions:

(1) The brightness between adjacent frames is constant;

(2) The frame-taking time of adjacent video frames is continuous, or the movement of objects between adjacent frames is relatively "small";

(3) Maintaining spatial consistency; that is, the pixels of the same sub-image have the same motion.

The basic idea is:

Let I(x,y,t) be the brightness at point P(x,y) on the image at time t, I(x+dx,y+dy,t+dt) is the point (x, t+dt) on the image at time t+dt The brightness at the corresponding point P' of y), according to the assumption of brightness consistency, we get:

I(x,y,t)=I(x+dx,y+dy,t+dt) (9)

According to the definition of optical flow (u, v) = (dx/dt, dy/dt), Taylor expansion can be used to obtain:

I _x u+I _y v+I _t =0 (10)

The above formula is the basic optical flow constraint equation. One equation cannot solve two unknowns. At this time, another constraint needs to be introduced to solve the two velocity vectors. According to assumption (3), n equations can be obtained, and then combined with the least squares Multiplication solution can get L-K optical flow:

即为所求，也就是计算得到的L-K光流。in,

That is, what is required, that is, the calculated LK optical flow.

Referring to Figure 18, the template image and the image to be tested are regarded as continuous frame images in the video, and the irregular and small distortions in the image to be tested are regarded as local micro-motion. The detection process is as follows:

S3021. The optical flow field between the two images can be calculated by using the LK optical flow method. As shown in Figure 17, the optical flow vector V _i,j at any position (i, j) of the optical flow field is composed of:

Vi _,j = (x,y) (12)

Among them, the x value represents the row offset of the image to be tested at point (i, j), and the y value represents the column offset;

S3022, decompose the optical flow vector V _i,j to obtain its components x and y to form two data sets {x} and {y};

S3023, perform threshold segmentation on the image to be tested, obtain the screen printing foreground area T, and count the coordinates (i, j) of all points in the area T;

S3024, calculate the corresponding point set, the image to be tested is compared with the template image, and the pixel at the coordinate (i, j) is moved to the coordinate (i+x, j+y) due to distortion;

S3025, using (i, j) and (i+x, j+y) as the corresponding point sets, calculate the projection transformation matrix A, and perform projection transformation on the silk screen to be measured with this matrix, so as to realize the correction of the slight deformation, and the corrected image As shown in Figure 19. As can be seen from Figure 19, the degree of coincidence between the silk screen to be tested and the template silk screen has been relatively high, and the distortion phenomenon has been basically eliminated, which is convenient for subsequent detection;

S3026 , perform aberration processing on the corrected image to be tested and the template image again, perform defect separation and extraction processing similar to that proposed in step S301 , and make a final mark.

After the processing of the above method, the normal detection of micro-distorted silk screen printing can be realized without false alarms, which greatly improves the practicability of the method, can meet the real-time detection of enterprise pipelines, and can ensure high detection accuracy. Rate.

In order to make the purposes, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments These are some embodiments of the present invention, but not all embodiments. The components of the embodiments of the invention generally described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations. Thus, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but is merely representative of selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

By adopting the battery screen printing quality detection method based on ORB feature matching and LK optical flow method of the present invention for online testing, 981 screen printings can be correctly detected by online testing of 1000 screen printings to be inspected, and the detection rate is 98.1%, of which , 11 pieces of good screen printing were judged as defective products, and 8 pieces of bad screen printing were judged as good products.

However, the detection rate of the traditional global template matching method for the same sample is only 49%, which is completely unsuitable for the detection of such inaccurate prints.

To sum up, the detection method of the present invention has high detection rate and good real-time performance compared with the traditional detection method, and has strong adaptability to inaccurate printed matter such as distortion, position angle error, etc., and can accurately separate real defects. without generating false positives.

The above content is only to illustrate the technical idea of the present invention, and cannot limit the protection scope of the present invention. Any modification made on the basis of the technical solution proposed in accordance with the technical idea of the present invention falls within the scope of the claims of the present invention. within the scope of protection.

Claims (7)

1. a battery silk screen quality detection method based on ORB feature matching and LK optical flow method, is characterized in that, comprises the following steps: S1. Collect battery silkscreen image data, and perform preprocessing to extract battery silkscreen area; S2. Use the rectangular block method to model, including the illustration part template and the text part template, based on the ORB algorithm to extract and match the characteristics of the template silk screen and the silk screen to be tested, so as to realize the positioning of the silk screen content; S3. Perform the detection based on the image aberration method based on morphology. If there is a false alarm, use the distortion correction detection method based on the L-K optical flow method for secondary detection; if there is no false alarm, output the result image and detection data, and perform sorting Operation, the illustration part detection is as follows: S3011, obtaining the corresponding point relationship according to the matching result, calculating the rigid body transformation matrix, aligning the sub-template image with the image to be tested, and using the threshold segmentation method to segment the detection content from the background; S3012. Perform morphological corrosion and expansion treatment on the screen-printed content of the battery. The multi-print defects and missing-print defects can be obtained by the positive and negative differences between the test object and the template. The multi-print defects are: T _(corrosion) -M _(expansion) ; missing printing defects are: M _(corrosion) -T _(expansion) , T _(corrosion) is the content area of the screen printing of the battery to be tested after corrosion, M _(expansion) is the content area of the template screen printing after expansion, M _(corrosion) is the content area of the stencil screen printing after corrosion, and T _(expansion) is the content area of the screen printing of the battery to be tested after expansion; S3013, perform connectivity analysis and threshold screening on the difference image after the difference in step S3012, and the area exceeding the threshold is the defect area; S3014, take the center of the defect as the center of the circle, and mark the defect with the smallest circumcircle of the defect; The text part defect detection process is as follows: S3015, affine transformation registration; S3016, perform skeleton extraction processing on the text, after the extraction, the text lines are unified into a single-pixel width, and only the topology structure is retained; S3017, performing a rice-shaped translation process on the extracted text, namely up, down, left, right, upper left, upper right, lower left, and lower right in 8 directions, and then performing image difference to obtain 8 difference images; S3018, performing an intersection operation on the 8 difference images obtained in step S3017, and the final reserved area is regarded as a real defect area; S3019. Connectivity analysis threshold screening and defect marking; The template image and the image to be tested are regarded as continuous frame images in the video, and the irregular and small distortions in the image to be tested are regarded as local micro-motion. The detection process is as follows: S3021, using the LK optical flow method to calculate the optical flow field between the two images, the optical flow vector V _i,j at any position (i, j) of the optical flow field is composed as follows: V _i,j =(x,y) Among them, the x value represents the row offset of the image to be tested at point (i, j), and the y value represents the column offset; S3022, decompose the optical flow vector V _i,j to obtain its components x and y to form two data sets {x} and {y}; S3023, perform threshold segmentation on the image to be tested, obtain the battery screen printing foreground area T, and count the coordinates (i, j) of all points in the area T; S3024, calculate the corresponding point set, the image to be tested is compared with the template image, and the pixel at the coordinate (i, j) is moved to the coordinate (i+x, j+y) due to distortion; S3025, using (i, j) and (i+x, j+y) as the corresponding point sets, calculate the projection transformation matrix A, and perform projection transformation on the screen printing of the battery under test with this matrix, so as to realize the correction of small deformation; S3026 , performing shadow difference processing on the corrected image to be tested and the template image again, performing defect separation and extraction processing, and making final marks. 2. the battery silkscreen quality detection method based on ORB feature matching and LK optical flow method according to claim 1, is characterized in that, step S1 is specifically: S101, creating a crossed rectangular window according to the original image size, and cropping the original image, the purpose of which is to obtain boundary information of the battery silk screen area; S102, use the Ostu algorithm to calculate the binarization threshold, initially separate the battery silk screen area from other backgrounds, and appropriately offset the threshold value in combination with the image grayscale features to accurately separate the battery silk screen area and other backgrounds, and remove them by morphological opening operation Edge burr to eliminate small edge segmentation errors; S103, performing minimum circumscribed rectangle fitting on the region extracted in step S102, and cropping the original image according to the fitting result to obtain an accurate battery silk screen area; S104, performing grayscale correction on the image to be tested, and adjusting it to the same grayscale level as the template image. 3. the battery silk screen quality detection method based on ORB feature matching and LK optical flow method according to claim 2, is characterized in that, step S104 is specifically: S1041. Grayscale the image, convert the original three-channel image into a single-channel grayscale image, and use the psychological grayscale formula: gray=0.299*red+0.587*green+0.114*blue S1042, divide the foreground region R ₁ and the background region R ₂ by using a threshold segmentation method; calculate the gray mean value of the foreground region and the background region respectively, perform the above operations on the image to be tested and the template image respectively, and calculate the gray mean value as follows:

Among them, R _i is the region extracted after threshold segmentation, p is a pixel in the region, g(p) is the gray value at point p, and F is the total number of pixels in R _i ; S1043. Perform grayscale transformation on the image to be tested to obtain the mean value M1 of the foreground grayscale of the template, the mean value _{M2 of} the background grayscale of the template, the mean value T1 of the foreground grayscale of the image to be tested, and the mean value T2 _of the background _grayscale of the image to be tested, and perform grayscale Transform, calculate the gray scale scaling coefficient Mult and the gray scale translation coefficient Add respectively, and use these two coefficients to map and correct the original gray scale of the image to be tested, as follows: Mult=(M ₁ -M ₂ )/(T ₁ -T ₂ ) Add=M ₁ -Mult*T ₁ g'=g*Mult+Add Among them, Mult is the gray scale scaling coefficient, Add is the gray scale translation coefficient, and g' is the new gray value after mapping. 4. the battery silk screen quality detection method based on ORB feature matching and LK optical flow method according to claim 1, is characterized in that, step S2 is specifically: S201, oFAST feature point extraction; S202, rBRIEF feature descriptor structure; S203, feature point matching. 5. the battery silk screen quality detection method based on ORB feature matching and LK optical flow method according to claim 4, it is characterized in that, in step S201, the gray value of point P and the neighbor whose radius is 3 with P as the center. The grayscale values of 16 pixels in the domain are compared. If the pixel value of point P and the pixel value of n consecutive pixels in the neighborhood on the circle are less than the threshold t, then point P is a feature point. 6. the battery silk screen quality detection method based on ORB feature matching and LK optical flow method according to claim 4, is characterized in that, in step S202, select N point pairs around key point P, and N point pairs The comparison results of are combined as descriptors, and all point pairs are compared to generate a binary string of length n. 7. the battery silk screen quality detection method based on ORB feature matching and LK optical flow method according to claim 4, is characterized in that, in step S203, adopts Hamming distance between descriptors as similarity measure, two etc. The Hamming distance between long binary strings is the number of different characters in the corresponding positions of the two strings, and the matching point pairs are screened by setting the similarity threshold; after obtaining the matching result, according to the ORB features in the template image and the image to be tested The corresponding relationship between the points, the template and the image to be tested are registered, and each rectangular sub-template in the template image is matched with each part of the image to be tested.

Images