CN114004826A - Appearance defect detection method for medical injection molded parts based on vision - Google Patents

Abstract

The invention discloses a method for detecting the appearance defects of medical injection molded parts based on visual template matching and deep learning. Ring light image and concentrated light image; use the deep learning model to detect the defect features of lack of glue, pits and parting line in the ring light image of the medical injection part to be tested, and save the ring light judgment result of the ring light image; use template matching Methods The light-collecting image of the medical injection molded part to be tested is detected for the defects of sharp edges and burrs, and the light-collecting judgment result of the light-collecting image is output; the ring-light judgment result and the light-collecting judgment result are integrated. The test results of the medical injection molded parts are qualified products, otherwise they are defective products. The invention can quickly detect unqualified products under the premise of ensuring production tact, and reduce the processing time of a single sample by comprehensively judging and outputting two types of detection results.

Description

Appearance defect detection method for medical injection molded parts based on vision

technical field

The invention relates to the technical field of computer vision recognition, in particular to a vision-based method for detecting appearance defects of medical injection molding parts.

Background technique

This injection molded part is used for the inhalation mixing part of the ventilator. It is used to mix the air and oxygen input into the ventilator and ensure the airtightness of the ventilator. It is a very important part of the ventilator. At present, the pain point of injection molding inspection lies in the existence of tiny defects, and the traditional manual inspection method has poor stability and efficiency.

In the prior art, Eggert et al. proposed a candidate frame extraction method based on feature map upsampling in ACM's Improving small object proposals for company logo detection. They used high-resolution feature maps to extract candidate frames for small objects. , on the one hand, the candidate frame is closer to the target frame, and on the other hand, the information of the small target is enhanced by enlarging the feature map.

In the IEEE's An analysis of scale invariance in object detectionsnip, Singh et al. proposed that by dividing the targets of different sizes in the dataset into different subsets, at the same time designing multiple detectors in the network to detect different sizes, and then letting different sizes The targets of each pass through the network to detect the corresponding size of the detector, thereby avoiding the problem of class imbalance in the data set, and making the detection effect of small targets better.

In IEEE/CVF's Scale match for tiny person detection, Yu et al. proposed to integrate the information around the target into the features of the current target, so as to obtain a better feature representation for small targets and effectively improve the effect of small target detection. .

The detection accuracy of the existing detection technology for the silicone parts cannot meet the detection requirements of the product, and the detection of the silicone parts cannot be completed within the time range of the production tact.

SUMMARY OF THE INVENTION

In order to solve the above technical problems, the embodiments of the present invention provide a visual-based method for detecting appearance defects of medical injection molded parts,

A visual defect detection method for medical injection molded parts, comprising the following steps:

Step S1: respectively photographing and imaging the medical injection molded part to be tested under two light sources of ring light and concentrated light, to obtain the ring light image and the concentrated light image of the medical injection molded part to be tested;

Step S2: use the deep learning model to detect the defect features of lack of glue, pits and parting lines on the ring light image of the medical injection molded part to be tested, and save the ring light judgment result of the ring light image;

Step S3: use the template matching method to detect the defect features of the sharp edges and burrs on the light collection image of the medical injection molded part to be tested, and output the light collection judgment result of the light collection image;

Step S4: Integrate the judgment result of the ring light and the judgment result of the light collection. When all of them are qualified, output the detection result of the medical injection molded part to be tested as a qualified product, otherwise, it is a defective product.

Preferably, in the above-mentioned visual-based method for detecting appearance defects of medical injection molded parts, the pre-training process of the deep learning model includes the steps:

Step S21: collecting the ring light image of the medical injection molded part under the ring light light source;

Step S22: marking the ring light images with pitting, lack of glue and parting line defects, and generating a defect detection data set of lack of glue, pockmarking and parting line;

Step S23: Input the defect detection data set generated in the step S22 into the defect detection network as a training data set;

Step S24: training to obtain defect type, center point, and bounding box information;

Step S25: Output the deep learning model.

Preferably, in the above-mentioned visual-based method for detecting appearance defects of medical injection molded parts, the step S3 includes:

Step S31: Selecting a qualified medical injection molded part sample to take a photo and imaging under the light-collecting light source, obtain a template image of a qualified medical injection molded part, and mark the feature points in the template image;

Step S32: marking in order the feature points on the light collection image of the medical injection molded part to be tested that are consistent with the feature points in the template image;

Step S33: performing two affine transformations on the template image, performing a difference between the obtained affine transformed image and the light collection image to obtain two differential images, and then taking the AND of the two differential images;

Step S34: performing connected domain analysis on each area of the obtained image, and filtering out areas with smaller areas;

Step S35: Calculate the foreground area, contour similarity, and gradient respectively through the image area filtered by the area;

Step S36: Comprehensively judge whether the calculated value is greater than the threshold value, if yes, output this part as a defect feature, otherwise, judge the next area.

Preferably, in the above-mentioned visual-based method for detecting appearance defects of medical injection molded parts, in the step S31, before annotating the feature points in the template image, it further includes image preprocessing on the template image, and The image preprocessing of the template image includes:

ROI region extraction is performed on the obtained template image, so as to reduce the size of the image and speed up subsequent processing.

Preferably, in the above-mentioned visual-based method for detecting appearance defects of medical injection molded parts, in the step S32, before marking the feature points on the light collection image of the medical injection molded part to be tested, the method further comprises: Image preprocessing of the image, the image preprocessing of the light collection image includes:

Extracting the ROI area on the acquired light collection image of the medical injection molded part to be tested, and scaling the size of the light collection image to be consistent with the size of the template image.

Preferably, in the above-mentioned visual-based method for detecting appearance defects of medical injection molded parts, in step S31, the feature points in the annotated template image include:

Label the centroid points of the 8 feature points of the template image, sort them according to the x value of the centroid coordinates, and divide the 8 centroid points into two groups.

Preferably, in the above-mentioned visual-based method for detecting appearance defects of medical injection molded parts, in the step S32, the feature points in the light collection image of the medical injection molded part to be tested are marked in sequence and consistent with the feature points in the template image feature points, including:

Mark the centroid points of the 8 characteristic points of the light collection image of the medical injection molded part to be tested, and number the 8 centroid points of the light collection image in the same order as the 8 centroid points in the template image, and divide them into two parts. Group.

Preferably, in the above-mentioned visual-based method for detecting appearance defects of medical injection molded parts, in step S33, performing two affine transformations on the template image, including:

Using two sets of corresponding centroid points, the template image is subjected to two affine transformations to obtain an affine transformed image.

Preferably, in the above-mentioned visual-based method for detecting appearance defects of medical injection molded parts, in the step S34, the connected domain analysis is performed on each area of the obtained image, and the area with a smaller area is filtered out, include:

Perform a connected domain analysis on each area of the image after summation, find out the area with a connected domain area greater than 5, and filter out the area with an area less than 5.

Preferably, in the above-mentioned visual-based method for detecting appearance defects of medical injection molded parts, in the step S35, the calculation of the foreground area of the image area filtered by the area includes:

Perform a foreground area search to find the area with the largest foreground area, and compare the template image with the medical device to be tested.

The area of the light collection image of the injection molded part is intercepted;

According to the calculation and comparison of the Hu moment invariant between the template image and the collected light image of the medical injection part to be tested, the 7 invariant moment groups are expressed as follows:

I ₁ =y ₂₀ +y ₀₂ (1)

I ₃ =(y ₃₀ +3y) ² +(3y ₂₁ -y ₀₃ ) ² (3)

I ₄ =(y ₃₀ +y ₁₂ ) ² +(y ₂₁ +y ₀₃ ) ² (4)

I ₆ =(y ₂₀ -y ₀₂ )[(y ₃₀ +y ₁₂ ) ² -(y ₂₁ +y ₀₃ ) ² ]+4y ₁₁ (y ₃₀ +y ₁₂ )(y ₂₁ +y ₀₃ ) (6)

Among them, y _pq is the (p+q) moment, p and q are integers between 0 and 3, and I ₁ -I ₇ are seven invariant moments.

Preferably, in the above-mentioned visual-based method for detecting appearance defects of medical injection molded parts, in step S35, the calculation of the contour similarity is performed on the image area filtered by the area, and the calculation expression of the contour similarity is: :

分别是图像A，B的Hu矩。in,

are the Hu moments of images A and B, respectively.

Preferably, in the above-mentioned visual-based method for detecting appearance defects of medical injection molded parts, in the step S35, the gradient calculation is performed on the image area filtered by the area, and the expression for the gradient calculation is:

为点(x,y)处的梯度值，θ(x,y)为梯度方向。Among them, G(x, y) is the gray value of the point whose coordinates are (x, y), G _x (x, y) is the gradient in the x direction, G _y (x, y) is the gradient in the y direction,

is the gradient value at point (x, y), and θ(x, y) is the gradient direction.

Preferably, in the above-mentioned visual-based method for detecting appearance defects of medical injection molded parts, in step S36, the comprehensive judgment is made as to whether the calculated value is greater than the threshold value, and if so, this part is output as the defect feature, Otherwise, judge the next area, including:

Step S361: Judging according to the foreground area value calculated in the step S35, if the foreground area is greater than the set threshold, the area is judged as a defect, otherwise the contour similarity is judged;

Step S362: if the contour similarity is less than the set threshold, remove the area;

Step S363: According to the sum of the contour similarity and the gradient as the judgment condition, if it is greater than the threshold, the position is judged as a defect area;

Step S364: After obtaining all the positions that may be defective areas, calculate the distance from the centroid of the candidate frame area to the centroid of the injection molded part, and judge the position of the candidate area based on this. If there is a defective area, the result of the light collection judgment for outputting the light collection image is unqualified; otherwise, the light collection judgment result for outputting the light collection image is qualified.

The beneficial effects of the present invention are as follows: the visual defect detection method for medical injection molded parts provided by the present invention can quickly and accurately detect the medical injection molded parts under the condition of ensuring the rapid production of the assembly line. The deep learning and template matching algorithm model adopted solves the problem that traditional methods cannot perform multiple defect detection on all samples in a short time. The processing time for a single sample is only 600ms, which fully meets the tact requirements of pipeline production and improves the The degree of automation of the production line.

Description of drawings

In order to illustrate the technical solutions of the embodiments of the present invention more clearly, the following briefly introduces the drawings that are used in the description of the embodiments of the present invention. Obviously, the drawings in the following description are only some embodiments of the present invention. , for those of ordinary skill in the art, other drawings can also be obtained from these drawings without creative labor.

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

Fig. 2 is the front view of the photographing structure of the medical injection molded part of the present invention;

3 is a top view of the medical injection molded part of the present invention on a glass carrier;

Fig. 4 is the light collection image of the medical injection molded part of the qualified product of the present invention;

FIG. 5 is a light collection image of a defective medical injection molded part with burrs according to the present invention;

FIG. 6 is a light collection image of a defective medical injection molded part with a sharp edge according to the present invention;

Fig. 7 is the ring light image of the medical injection molded part of the qualified product of the present invention;

8 is a ring light image of a defective medical injection molded part with poor mold clamping according to the present invention;

FIG. 9 is a ring light image of a defective medical injection molded part with pitting according to the present invention.

Detailed ways

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 are part of the embodiments of the present invention, but not all of the embodiments. 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.

The defect types of medical injection molded parts can be abstractly divided into morphology and texture. The optical imaging scheme is used to shoot in two ways: ring light and concentrated light, respectively, for these two defect types.

For morphological defects, it is mainly manifested as burrs and slashes. These defects will change the shape of the injection molded parts to a certain extent. Using condensing light for imaging and shooting can achieve a good representation of the morphological characteristics of the injection molded parts, while diluting other features. The effect is helpful to reduce the matching error in the algorithm processing, improve the detection accuracy, and solve the problem of missed detection of small defects.

For texture defects, it is mainly manifested as pits and parting lines. Using ring light to shoot can show good details in imaging, optimize the ability of deep learning models to extract feature information of injection molded parts, and improve the accuracy of model judgment. It can accurately mark the location of the defect, which is convenient for inspectors to observe the defect and better guide the improvement of production.

The combination of two lighting methods, ring light and concentrated light, is used to restore the morphological and texture characteristics of medical injection molded parts to the greatest extent in imaging. This not only improves the detection speed, but also ensures the accurate detection of defects, fully meets the production and detection requirements of medical injection molding parts, and greatly improves the automation degree of production.

Referring to FIG. 1, the present invention proposes a visual-based method for detecting appearance defects of medical injection molded parts, comprising the following steps:

Step S1: The glass carrier 1 moves along the set direction. When the medical injection molded part 2 to be tested moves to the preset position, first the lateral low-angle ring light source 3 (ring light source) above the medical injection molded part 2 to be tested is turned on Then, turn off the lateral low-angle ring light source 3, turn on the bottom parallel light source 4 (collecting light source), and take the collected light shooting of the medical injection part 2 to be tested. Thereby, the imaging effect diagrams of the medical injection molded part 2 to be tested under two light sources are obtained, which are the ring light image and the concentrated light image respectively;

Step S2: The surface defects of the medical injection molded part 2 to be tested can be clearly displayed under the ring light, and there is a large texture difference between the defective area and the normal area. The optical image is used to detect the defect characteristics of lack of glue, pitting and parting line. The position of the defect will be displayed in a rectangular frame in the picture, and the ring light judgment result of the ring light image will be saved;

Step S3: Use the template matching method to detect the defect features of the sharp edges and burrs on the light-collecting image of the medical injection molded part 2 to be tested. Find out the centroids of the 8 holes of the medical injection molded part 2 to be tested, and sort them from small to large according to the abscissa. Then, the corrected detection image is differentiated from the produced standard template image, and there will be a certain area of white area after the difference where there are burrs and drapes. Perform a step-by-step analysis on these white areas, and calculate the area, gradient, similarity and other parameters of these areas. If all the regional parameters do not exceed the set conditions, the medical injection molded part 2 to be tested is judged to be qualified, otherwise it is judged to be unqualified. For the position with defects, a rectangular box will be used to select and display in the final displayed figure. output the judgment result of the light collection image;

Step S4: Integrate the judgment result of the ring light and the judgment result of the light collection, and when all are qualified, output the test result of the medical injection molded part 2 to be tested as a qualified product, otherwise, it is a defective product.

As shown in FIG. 2 and FIG. 3 , FIG. 2 is a front view of the photographing structure of the medical injection molded part 2 to be tested according to the present invention, wherein the medical injection molded part 2 (product) to be tested is distributed on the high translucent glass 1, and the high translucent glass The glass 1 is arranged on a high-transparency glass carrier 5 and can move in a linear direction. A camera 6 is arranged above the highly transparent glass 1, and the camera 6 takes pictures of the medical injection molded parts 2 (products) to be tested passing below. Among them, a bottom parallel light source 4 (light collection light source) is arranged directly under the high light transmission glass 1 to provide a light collection light source for the medical injection molded part 2 (product) to be tested. Above the high light-transmitting glass 1, on the upper peripheral side of the medical injection molded part 2 (product) to be tested, there is a lateral low-angle annular light source 3 (ring light source) surrounded in a circumferential direction. The light source 3 can provide a ring light source for the medical injection molded part 2 (product) to be tested.

Specifically, the glass carrier 5 moves along the set direction. When the medical injection molded part 2 (product) to be tested moves to the preset position, first, the lateral low-angle ring light source 3 above the medical injection molded part 2 (product) to be tested Open it, and take a ring light photograph of the passing medical injection molded part 2 (product) to be tested, which is used for the detection of pitting, parting line and lack of glue. Then, turn off the lateral low-angle ring light source 3, turn on the bottom parallel light source 4 under the glass carrier 5, and collect light for the passing medical injection molded part 2 (product) to be tested for detection of burrs and sharp edges.

Among them, Figure 4 shows a qualified medical injection molded part under the light source, Figure 5 is a concentrated image of a defective medical injection molded part with burrs, and Figure 6 is a condensed image of a defective medical injection molded part with a sharp edge. , Figure 7 shows the ring light image of the qualified medical injection molded part under the ring light image, Figure 8 is the ring light image of the defective medical injection molded part with poor mold clamping, and Figure 9 is the ring light of the defective medical injection molded part with pitting. light image.

Further, in a preferred implementation of the present invention, the pre-training process of the deep learning model includes the steps:

Step S21: collecting the ring light image of the medical injection molded part under the ring light light source, and selecting the image with defects;

Step S22: Use labelimg software to mark the ring light image with pitting, lack of glue, and parting line defects, and use a rectangular frame to frame the position of the defect. The software can automatically generate an .xml file to save the labeling information, so that Defect detection datasets for lack of glue, pitting, and parting lines can be generated;

Step S23: Input the defect detection data set generated in the step S22 as a training data set into the defect detection network, and the network will optimize the network parameters according to the marked position and the marked defect type;

Step S24: After training the model, the data set defect type, bounding box and confidence information will be obtained;

Step S25: Use the trained weight file, defect name, and network structure information as the output file, and output the deep learning model.

Further, in a preferred implementation of the present invention, the step S3 includes:

Step S31: Selecting a qualified medical injection molded part sample and taking a photo and imaging under the light-collecting light source, obtaining a template image of a qualified medical injection molding part, marking the feature points in the template image, and comparing it with the image of the sample to be inspected during detection;

Step S32: Mark the feature points on the light collection image of the medical injection molded part to be tested that are consistent with the feature points in the template image in sequence, calculate the centroid positions of the eight holes in the image, and perform the centroids of the eight holes in the order of coordinates. number, the centroid points are divided into two groups, and they are matched with the corresponding numbered centroids in the template image.

Step S33: Using the two sets of centroid matching, perform two affine transformations on the template image, and make a difference between the obtained affine transformed image and the collected light image to obtain two differential images, and then take the sum of the two differential images to obtain: There are different areas between the image to be inspected and the standard template image;

Step S34: perform a connected domain analysis on each area of the obtained image, calculate the area of each connected domain, set a screening threshold according to the size requirements of the defect, and filter out areas with smaller areas;

Step S35: Calculate the foreground area, contour similarity, and gradient respectively through the image area filtered by the area;

Step S36: a comprehensive design judger, judges whether the calculated value is greater than the set threshold, if so, this part is output as a defect feature, otherwise, the judgment of the next area is carried out until all connected domains are analyzed, and if there is no Defect area, the product to be inspected is judged as qualified, otherwise it is judged as unqualified product.

Further, in a preferred implementation of the present invention, in the step S31, before the feature points in the template image are marked, image preprocessing on the template image is also included, and the image preprocessing on the template image includes: :

Find the position of the injection molded part on the obtained template image, and then take the minimum circumscribed rectangle of the injection molded part as the basis, and expand each of the four sides by 50 pixels, as the ROI area of the image, and extract the ROI area to reduce the size of the image. , to speed up subsequent processing.

Further, in a preferred implementation of the present invention, in the step S32, before marking the feature points on the light collection image of the medical injection molded part to be tested, image preprocessing on the light collection image is also included, and the light collection image is preprocessed. The image preprocessing of the collected light image includes:

Find the position of the injection molded part for the acquired light collection image to be tested, and then take the minimum circumscribed rectangle of the injection molded part as the basis, and each of the four sides expands 50 pixels outward as the ROI area of the image, and extracts the ROI area. The size of the collection image is scaled to match the size of the template image.

Further, in a preferred implementation of the present invention, in the step S31, the feature points in the standard template image include:

Extract the centroid points of the 8 holes of the standard template image, sort them according to the x value of the centroid coordinates, and divide the 8 centroid points into two groups.

Further, in a preferred implementation of the present invention, in the step S32, the feature points on the light collection image of the medical injection molded part to be tested that are consistent with the feature points in the template image are marked in sequence, including:

Mark the centroid points of the 8 characteristic points of the light collection image of the medical injection molded part to be tested, and number the 8 centroid points of the light collection image in the same order as the 8 centroid points in the template image, and divide them into two parts. Group.

Further, in a preferred implementation of the present invention, in the step S33, the affine transformation is performed on the template image twice, including:

Use two sets of corresponding centroid points to correspond the template image with the same number to the image to be inspected, perform two affine transformations on the template image to obtain an affine transformed image, and then perform two differences between the template image and the image to be inspected. , and then take the AND to get the area where the two images are different.

Further, in a preferred implementation of the present invention, in the step S34, the connected domain analysis is performed on each area of the image after summing, and the area with a smaller area is filtered out, including:

Perform a connected domain analysis on each area of the image after summation, calculate the area size of each connected domain, find out the area with a connected domain area greater than 5, and filter out the area with an area less than 5.

Further, in a preferred implementation of the present invention, in the step S35, the calculation of the foreground area of the image area filtered by the area includes:

Perform a foreground area search to find the area with the largest foreground area, then intercept the template image and the area of the light collection image of the medical injection molded part to be tested, extract the contour information of this area, and calculate the similarity of the contours of the two images according to the following formula Information such as degree and gradient are used as the basis for judgment.

According to the calculation and comparison of the Hu moment invariant between the template image and the collected light image of the medical injection part to be tested, the 7 invariant moment groups are expressed as follows:

I ₁ =y ₂₀ +y ₀₂ (1)

I ₃ =(y ₃₀ +3y) ² +(3y ₂₁ -y ₀₃ ) ² (3)

I ₄ =(y ₃₀ +y ₁₂ ) ² +(y ₂₁ +y ₀₃ ) ² (4)

I ₆ =(y ₂₀ -y ₀₂ )[(y ₃₀ +y ₁₂ ) ² -(y ₂₁ +y ₀₃ ) ² ]+4y ₁₁ (y ₃₀ +y ₁₂ )(y ₂₁ +y ₀₃ ) (6)

Among them, y _pq is the (p+q) moment, p and q are integers between 0 and 3, and I ₁ -I ₇ are seven invariant moments.

Preferably, in the above-mentioned visual-based method for detecting appearance defects of medical injection molded parts, in step S35, the calculation of the contour similarity is performed on the image area filtered by the area, and the calculation expression of the contour similarity is: :

分别是图像A，B的Hu矩。in,

are the Hu moments of images A and B, respectively.

Further, in a preferred implementation of the present invention, in the step S35, the gradient calculation is performed on the image area filtered by the area, and the expression of the gradient calculation is:

为点(x,y)处的梯度值，θ(x,y)为梯度方向。Among them, G(x, y) is the gray value of the point whose coordinates are (x, y), G _x (x, y) is the gradient in the x direction, G _y (x, y) is the gradient in the y direction,

is the gradient value at point (x, y), and θ(x, y) is the gradient direction.

Further, in the preferred implementation of the present invention, in the step S36, the comprehensive judgment is made whether the calculated value is greater than the threshold value, if so, this part is output as the defect feature, otherwise the judgment of the next area is performed, include:

Step S361: Judging according to the foreground area value calculated in the step S35, if the foreground area is greater than the set threshold, the area is judged as a defect, otherwise the contour similarity is judged;

Step S362: if the contour similarity is less than the set threshold, remove the area;

Step S363: According to the sum of the contour similarity and the gradient as the judgment condition, if it is greater than the threshold, the position is judged as a defect area;

Step S364: After obtaining all the positions that may be defective areas, calculate the distance from the centroid of the candidate frame area to the centroid of the injection molded part, and judge the position of the candidate area based on this. If there is a defective area, the result of the light collection judgment for outputting the light collection image is unqualified; otherwise, the light collection judgment result for outputting the light collection image is qualified.

The beneficial effects of the present invention are as follows: the visual defect detection method for medical injection molded parts provided by the present invention can quickly and accurately detect the medical injection molded parts under the condition of ensuring the rapid production of the assembly line. The deep learning and template matching algorithm model adopted solves the problem that traditional methods cannot perform multiple defect detection on all samples in a short time. The processing time for a single sample is only 600ms, which fully meets the tact requirements of pipeline production and improves the The degree of automation of the production line.

The above-mentioned embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it is still possible to implement the foregoing implementations. The technical solutions described in the examples are modified, or some technical features thereof are equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the within the protection scope of the present invention.

Claims (13)

1. a visual defect detection method based on medical injection molding, is characterized in that, comprises the steps: Step S1: respectively photographing and imaging the medical injection molded part to be tested under two light sources of ring light and concentrated light, to obtain the ring light image and the concentrated light image of the medical injection molded part to be tested; Step S2: use the deep learning model to detect the defect features of lack of glue, pits and parting lines on the ring light image of the medical injection molded part to be tested, and save the ring light judgment result of the ring light image; Step S3: use the template matching method to detect the defect features of the sharp edges and burrs on the light collection image of the medical injection molded part to be tested, and output the light collection judgment result of the light collection image; Step S4: Integrate the judgment result of the ring light and the judgment result of the light collection. When all of them are qualified, output the detection result of the medical injection molded part to be tested as a qualified product, otherwise, it is a defective product. 2. The visual defect detection method for medical injection molded parts according to claim 1, wherein the pre-training process of the deep learning model comprises the steps: Step S21: collecting the ring light image of the medical injection molded part under the ring light light source; Step S22: marking the ring light images with pitting, lack of glue and parting line defects, and generating a defect detection data set of lack of glue, pockmarking and parting line; Step S23: Input the defect detection data set generated in the step S22 into the defect detection network as a training data set; Step S24: training to obtain defect type, center point, and bounding box information; Step S25: Output the deep learning model. 3. The visual-based method for detecting appearance defects of medical injection molded parts according to claim 1, wherein the step S3 comprises: Step S31: Selecting a qualified medical injection molded part sample to take a photo and imaging under the light-collecting light source, obtain a template image of a qualified medical injection molded part, and mark the feature points in the template image; Step S32: marking in order the feature points on the light collection image of the medical injection molded part to be tested that are consistent with the feature points in the template image; Step S33: performing two affine transformations on the template image, performing a difference between the obtained affine transformed image and the light collection image to obtain two differential images, and then taking the AND of the two differential images; Step S34: performing connected domain analysis on each area of the obtained image, and filtering out areas with smaller areas; Step S35: Calculate the foreground area, contour similarity, and gradient respectively through the image area filtered by the area; Step S36: Comprehensively judge whether the calculated value is greater than the threshold value, if yes, output this part as a defect feature, otherwise, judge the next area. 4. The visual-based method for detecting appearance defects of medical injection molded parts according to claim 3, characterized in that, in the step S31, before marking the feature points in the template image, it further comprises: Image preprocessing, the image preprocessing for the template image includes: ROI region extraction is performed on the obtained template image, so as to reduce the size of the image and speed up subsequent processing. 5. The visual-based method for detecting appearance defects of medical injection molded parts according to claim 3, characterized in that, in the step S32, before marking the feature points on the light collection image of the medical injection molded parts to be tested, further It includes image preprocessing on the light collection image, and the image preprocessing on the light collection image includes: Extracting the ROI area on the acquired light collection image of the medical injection molded part to be tested, and scaling the size of the light collection image to be consistent with the size of the template image. 6. The visual-based method for detecting appearance defects of medical injection molded parts according to claim 3, wherein in the step S31, the feature points in the annotated template image include: Label the centroid points of the 8 feature points of the template image, sort them according to the x value of the centroid coordinates, and divide the 8 centroid points into two groups. 7 . The visual-based method for detecting appearance defects of medical injection molded parts according to claim 6 , wherein in the step S32 , the collection images of the medical injection molded parts to be tested are marked in sequence with the template. 8 . The feature points in the image are consistent with the feature points, including: Mark the centroid points of the 8 characteristic points of the light collection image of the medical injection molded part to be tested, and number the 8 centroid points of the light collection image in the same order as the 8 centroid points in the template image, and divide them into two parts. Group. 8 . The visual-based method for detecting appearance defects of medical injection molded parts according to claim 7 , wherein, in the step S33 , performing two affine transformations on the template image, comprising: 9 . Using two sets of corresponding centroid points, the template image is subjected to two affine transformations to obtain an affine transformed image. 9 . The visual-based method for detecting appearance defects of medical injection molded parts according to claim 3 , wherein, in the step S34 , the connected domain analysis is performed on each area of the image after summing, and filtering out Smaller areas, including: Perform a connected domain analysis on each area of the image after summation, find out the area with a connected domain area greater than 5, and filter out the area with an area less than 5. 10. The visual-based method for detecting appearance defects of medical injection molded parts according to claim 3, wherein, in the step S35, the calculation of the foreground area is performed on the image area filtered by the area, including: Perform a foreground area search to find the area with the largest foreground area, and intercept the template image and the area of the light collection image of the medical injection molded part to be tested; According to the calculation and comparison of the Hu moment invariant between the template image and the collected light image of the medical injection part to be tested, the 7 invariant moment groups are expressed as follows: I ₁ =y ₂₀ +y ₀₂ (1)

I ₃ =(y ₃₀ +3y) ² +(3y ₂₁ -y ₀₃ ) ² (3) I ₄ =(y ₃₀ +y ₁₂ ) ² +(y ₂₁ +y ₀₃ ) ² (4)

I ₆ =(y ₂₀ -y ₀₂ )[(y ₃₀ +y ₁₂ ) ² -(y ₂₁ +y ₀₃ ) ² ]+4y ₁₁ (y ₃₀ +y ₁₂ )(y ₂₁ +y ₀₃ ) (6)

Among them, y _pq is the (p+q) moment, p and q are integers between 0 and 3, and I ₁ -I ₇ are seven invariant moments. 11. The visual-based method for detecting appearance defects of medical injection molded parts according to claim 10, characterized in that, in the step S35, the contour similarity is calculated in the image area filtered by the area, and the contours are similar. The calculation expression of degree is:

in,

are the Hu moments of images A and B, respectively. 12. The visual-based method for detecting appearance defects of medical injection molded parts according to claim 10, wherein in the step S35, the gradient calculation is performed on the image area filtered by the area, and the expression of the gradient calculation for:

Among them, G(x, y) is the gray value of the point whose coordinates are (x, y), G _x (x, y) is the gradient in the x direction, G _y (x, y) is the gradient in the y direction,

is the gradient value at point (x, y), and θ(x, y) is the gradient direction. 13. The visual-based method for detecting appearance defects of medical injection molded parts according to claim 3, characterized in that, in the step S36, the comprehensive judgment is made to determine whether the calculated value is greater than the threshold value, and if so, this Part of it is output as defect features, otherwise, the next area is judged, including: Step S361: Judging according to the foreground area value calculated in the step S35, if the foreground area is greater than the set threshold, the area is judged as a defect, otherwise the contour similarity is judged; Step S362: if the contour similarity is less than the set threshold, remove the area; Step S363: According to the sum of the contour similarity and the gradient as the judgment condition, if it is greater than the threshold, the position is judged as a defect area; Step S364: After obtaining all the positions that may be defective areas, calculate the distance from the centroid of the candidate frame area to the centroid of the injection molded part, and judge the position of the candidate area based on this. If there is a defective area, the result of the light collection judgment for outputting the light collection image is unqualified; otherwise, the light collection judgment result for outputting the light collection image is qualified.

Images