CN116091506B - Machine vision defect quality inspection method based on YOLOV5 - Google Patents

Abstract

The invention discloses a machine vision defect quality inspection method based on YOLOV5, and relates to the technical field of visual inspection product quality. Step 1, respectively collect multiple sets of six-sided photos of standard qualified products and unqualified products with different defects; Step 2, respectively compose qualified Product six-sided standard diagram training library and unqualified product six-sided diagram training library; Step 3, build a quality inspection platform based on the YOLOV5 algorithm; Step 4, determine whether the product to be tested is qualified. The present invention adopts the same conditions to collect multiple sample images through the machine vision system as the training material of the quality inspection platform based on the deep learning YOLOV5 algorithm, and the final quality inspection platform will greatly improve the accuracy and efficiency of quality inspection after being trained with a large number of materials , using the model of four reference feature points to simplify the comparison, and introducing the concept of distance difference threshold in the comparison of reference feature points, which provides feedback data for the quality control of the product to be inspected, which is conducive to the improvement of product quality.

Description

A quality inspection method for machine vision defects based on YOLOV5

technical field

The invention relates to the technical field of visual inspection of product quality, in particular to a YOLOV5-based machine vision defect quality inspection method.

Background technique

Visual inspection is to use machines instead of human eyes for measurement and judgment. Visual inspection refers to converting the ingested target into an image signal through a machine vision product, that is, an image capture device, and transmitting it to a dedicated image processing system, and converting it into a digital signal according to pixel distribution, brightness, color and other information; Various calculations are performed on the signal to extract the characteristics of the target, and then the on-site equipment actions are controlled according to the results of the discrimination. Is a valuable mechanism for production, assembly or packaging. It is invaluable in its ability to detect defects and prevent defective products from being shipped to consumers.

At present, more and more visual systems are used in automated production to collect product appearance pictures, and then cooperate with visual algorithms to realize product defect quality inspection. The surface is the product quality inspection of the main use surface, and the applicability of products with important features on all six sides of the product is not high, and manual or other quality inspection procedures are still required to complete the defect quality inspection of other surfaces.

Contents of the invention

In order to solve the above-mentioned technical problem that the quality inspection of products in automation using visual technology cannot satisfy the defect quality inspection of six sides at one time, the present invention provides a machine vision defect quality inspection method based on YOLOV5. Adopt the following technical solutions:

A method for quality inspection of machine vision defects based on YOLOV5, comprising the following specific steps:

Step 1, collect multiple sets of six-sided photos of standard qualified products, and collect multiple sets of six-sided photos of unqualified products with different defects;

Step 2. Run the image stitching algorithm on the six-sided photos of standard qualified products and unqualified products respectively to regenerate the six-sided images of standard qualified products and the six-sided images of unqualified products. The six-sided images of standard qualified products include all six sides of standard qualified products Appearance characteristics, the six-sided picture of unqualified products includes all the appearance features of the six sides of unqualified products, multiple six-sided pictures of standard qualified products form the six-sided standard picture training library of qualified products, and multiple six-sided pictures of unqualified products form unqualified products Hexagram training library;

Step 3, build a quality inspection platform based on the YOLOV5 algorithm, and use the six-sided standard image training library for qualified products and the six-sided image training library for unqualified products generated in step 2 to train the quality inspection platform;

Step 4, determine whether the product to be inspected is qualified, use the machine vision system to obtain the six-sided inspection image of the product to be inspected, run the image stitching algorithm to stitch the six-sided inspection image into a product inspection image, and input the product inspection image to The quality inspection platform, the quality inspection platform uses the YOLOV5 algorithm to collect the characteristics of the product to be inspected, and compares the similarity. When the similarity IOU value between the product to be inspected and the six-sided image of the standard qualified product is greater than the set similarity threshold, the product is output qualified determination;

When the similarity IOU value between the six-sided image of the product to be inspected and the six-sided image of the standard qualified product is less than or equal to the set similarity threshold, the judgment of unqualified product is output.

By adopting the above-mentioned technical solution, the traditional product defect quality inspection based on visual inspection usually only collects top view or three-dimensional view for quality inspection, which cannot be applied to products to be inspected with important features on all six sides. Here, the inspection products are collected separately. After splicing the six-sided pictures to be inspected, the six-sided defect quality inspection is completed at one time, and the six-sided quality inspection is completed in one shot, which greatly improves the efficiency of visual quality inspection and can be applied to products to be inspected with important features on all six sides;

In order to meet the requirements of quality inspection, first of all, the same conditions are used for standard qualified products and unqualified products to collect multiple samples through the machine vision system, as the training material of the quality inspection platform based on deep learning YOLOV5 algorithm, standard qualified products and unqualified products The splicing operation is also required after the material is collected, and the final quality inspection platform will greatly improve the accuracy and efficiency of quality inspection after a large number of material training;

The quality inspection platform is essentially a visual server, which is equipped with various visual algorithms. The quality inspection platform uses the YOLOV5 algorithm to collect the characteristics of the product to-be-inspected image, and compares the similarity. When the IOU value of the similarity degree is greater than the set similarity threshold, the judgment of product qualification is output. This judgment method is to compare the six-sided diagram of standard qualified products first, that is, the priority qualification presumption, which is more suitable for product defect quality with a pass rate greater than 90%. Inspection can greatly improve the efficiency of defect quality inspection.

Optionally, in step 2, use the YOLOV5 target detection algorithm to detect defect features of multiple sets of six-sided photos of unqualified products, and form a defect feature comparison library.

Optionally, the yolov5 deep learning network of the YOLOV5 target detection algorithm model uses the data of the defect feature comparison library for iterative training, scales the width and height of the pictures in the defect feature comparison library, and fills the gray edges to a size of 224×224. Input, training image samples Three effective feature layers are extracted through the Focus network structure, namely (52, 52, 256), (26, 26, 512) and (13, 13, 1024), and the fourth feature FPN layer is constructed based on the three effective feature layers.

By adopting the above technical solution, after the defect features in the defect feature comparison library are captured by the YOLOV5 target detection algorithm, the feature layer based on the defect feature will be directly obtained, and finally divided into three effective feature layers according to the size of the defect feature , where the feature layer is (52,52,256) training image samples corresponding to small targets, the feature layer is (26,26,512) corresponding to the target in the detection granularity grid, and the feature layer is (13,13,1024) corresponding to the detection granularity grid For large targets, the fourth feature FPN layer can be obtained after performing convolution processing on the three initial feature layers.

Optionally, the features of the three effective feature layers (52, 52, 256), (26, 26, 512) and (13, 13, 1024) are extracted by 3×3 convolution to fuse different features, and the features obtained in the fusion Prediction is then performed on the graph to obtain the fourth feature FPN layer.

By adopting the above technical solution, the Focus network structure converts the information on the w-h plane to the channel dimension, and then extracts different features through 3×3 convolution. In this way, the information loss caused by downsampling can be reduced, and the features represented by the fourth feature FPN layer are more accurate.

Optionally, at least four reference feature points on the fourth feature FPN layer are sampled, the reference feature points refer to a feature area less than 3 pixels, and the degree of light and shade change in the feature area is greater than 50%, the four reference feature points Points are extracted respectively and a defect rapid comparison layer based on four reference feature points is constructed, and the distance detection algorithm is used to measure the mutual distance between the four reference feature points, and the distance data is used as an additional comparison item of the defect rapid comparison layer .

Optionally, a lightness detection algorithm is used to perform lightness change detection on the fourth feature FPN layer.

By adopting the above-mentioned technical scheme, because most of the defect features in unqualified products will be more, we use the mode of four reference feature points to simplify the comparison, and the four reference feature points plus the position distance relationship data can accurately Locating the defect feature is similar to the principle of fingerprint feature point recognition, and the biggest feature of the defect feature is the change of light and shade. Therefore, using the light and dark detection algorithm to detect the change of light and shade can quickly obtain four reference feature points.

Optionally, the machine vision defect quality inspection method also includes step 41, outputting product defect features, performing defect feature traversal before comparing the similarity between the product to-be-inspected image and the standard qualified product six-sided image, and calling the defect feature comparison library. The defect rapid comparison layer traverses the image of the product to be inspected. If the features of the four reference feature points are successfully matched, the matching similarity IOU value is greater than 0.9, and the mutual distance difference between the four reference feature points is less than the distance difference threshold, then The quality inspection platform directly determines that the product is unqualified, and at the same time frames the defect location and displays it based on the frame selection mark.

，距离差异阈值设0.05～0.1。Optionally, the distance difference value refers to the absolute value calculation of the difference between the distance value between any two reference feature points in the product to-be-inspected map and the distance value between the corresponding two reference feature points in the defect rapid comparison layer, Then, the absolute value is compared with the distance between two corresponding reference feature points in the defect rapid comparison layer, and the distance value between any two reference feature points in the product to-be-inspected image is set as S, and the corresponding value in the defect rapid comparison layer is The distance between two reference feature points is S1, and the distance difference threshold is D, then

, and the distance difference threshold is set to 0.05-0.1.

By adopting the above technical solution, the quality inspection platform needs to mark the specific defect location and defect type while making the unqualified judgment of the product to be inspected. Only such product defect quality inspection will give positive feedback to the production of the product, which is more Conducive to the improvement of product qualification rate;

Specifically, in order to mark the specific defect location and defect type, the defect feature similarity must be compared based on the defect quick comparison layer. If the similarity IOU value is greater than 0.9, the defect type is determined, and then the position data The comparison of the distance difference threshold is introduced here. The distance difference threshold is actually a key for defect location location. If the matching is successful, the defect location location is completed. Finally, the quality inspection platform outputs the unqualified product to be inspected. Judgment, and at the same time, mark the unqualified area and mark the defect type, so as to provide feedback data for the quality control of the product to be inspected, which is conducive to the improvement of product quality.

A machine vision system used to obtain six-sided images of products to be inspected, including a transparent platform for supporting products to be inspected, a bottom camera, a gantry and a top camera module, the transparent platform is set on the product to be inspected At the exit of the production line, the bottom camera is set under the transparent platform to take a bottom view of the product to be inspected, the gantry is set above the transparent platform, and the top camera module includes a front camera, a rear Camera, left camera, right camera and overlooking camera, the front camera, rear camera, left camera, right camera and overlooking camera are respectively mounted on the gantry, and are respectively used to photograph the front of the product to be inspected. The front view, the rear view, the left view, the right view and the top view, the bottom camera, the front camera, the rear camera, the left camera, the right camera and the top view camera are respectively connected to the quality inspection platform.

By adopting the above-mentioned technical scheme, the traditional machine vision system can only obtain the picture above the product, and the above-mentioned structure can complete the acquisition of the front view, rear view, left view, right view and top view of the inspected product. The inspection provides six pictures to be inspected in the same state at one time. Because it is shot at the same time, the light and shadow conditions are consistent, which is more conducive to the subsequent quality inspection platform to judge the quality inspection results.

In summary, the present invention includes at least one of the following beneficial technical effects:

The present invention can provide a machine vision defect quality inspection method based on YOLOV5, adopting the same conditions for standard qualified products and unqualified products to collect multiple sample images through the machine vision system, as a quality inspection platform based on deep learning YOLOV5 algorithm training Material, the final quality inspection platform will greatly improve the accuracy and efficiency of quality inspection after a large number of material training, using the model of four benchmark feature points to simplify the comparison, four benchmark feature points plus position distance relationship data can be accurate To accurately locate the defect feature, use the light and dark detection algorithm to realize the light and dark change detection, and quickly obtain four reference feature points. In the comparison of the reference feature points, the concept of distance difference threshold is introduced, and the unqualified area is selected at the same time. Mark and mark the type of defect, and provide feedback data for the quality control of the product to be inspected, which is conducive to the improvement of product quality.

Description of drawings

Fig. 1 is a schematic flow chart of a machine vision defect quality inspection method based on YOLOV5 of the present invention;

Fig. 2 is a schematic diagram of the electrical device connection principle of a machine vision system of the present invention;

Fig. 3 is a schematic diagram of the effect of a defect rapid comparison layer of a machine vision defect quality inspection method based on YOLOV5 in the present invention.

Explanation of reference numerals: 1. Bottom camera; 2. Front camera; 3. Rear camera; 4. Left camera; 5. Right camera; 6. Looking down camera; 100. Quality inspection platform.

Implementation

The present invention will be described in further detail below in conjunction with the accompanying drawings.

The embodiment of the present invention discloses a machine vision defect quality inspection method based on YOLOV5.

Referring to Figure 1-Figure 3, a machine vision defect quality inspection method based on YOLOV5 includes the following specific steps:

Step 1, collect multiple sets of six-sided photos of standard qualified products, and collect multiple sets of six-sided photos of unqualified products with different defects;

Step 2. Run the image stitching algorithm on the six-sided photos of standard qualified products and unqualified products respectively to regenerate the six-sided images of standard qualified products and the six-sided images of unqualified products. The six-sided images of standard qualified products include all six sides of standard qualified products Appearance characteristics, the six-sided picture of unqualified products includes all the appearance features of the six sides of unqualified products, multiple six-sided pictures of standard qualified products form the six-sided standard picture training library of qualified products, and multiple six-sided pictures of unqualified products form unqualified products Hexagram training library;

Step 3, build a quality inspection platform 100 based on the YOLOV5 algorithm, and use the six-sided standard map training library for qualified products and the six-sided diagram training library for unqualified products generated in step 2 to train the quality inspection platform 100;

Step 4, determine whether the product to be inspected is qualified, use the machine vision system to obtain the six-sided inspection image of the product to be inspected, run the image stitching algorithm to stitch the six-sided inspection image into a product inspection image, and input the product inspection image to The quality inspection platform 100, the quality inspection platform 100 adopts the YOLOV5 algorithm to collect the characteristics of the product to-be-inspected image, and through similarity comparison, when the similarity IOU value between the product to-be-inspected image and the six-sided image of the standard qualified product is greater than the set similarity threshold, Output product qualification judgment;

When the similarity IOU value between the six-sided image of the product to be inspected and the six-sided image of the standard qualified product is less than or equal to the set similarity threshold, the judgment of unqualified product is output.

The traditional quality inspection of product defects based on visual inspection usually only collects the top view or three-dimensional view for quality inspection, which cannot be applied to products to be inspected with important features on all six sides. After splicing, the defect quality inspection of six sides is completed at one time, and the quality inspection of six sides is completed in one shot, which greatly improves the efficiency of visual quality inspection and can be applied to products to be inspected with important features on all six sides;

In order to meet the requirements of quality inspection, first of all, the same conditions are used for standard qualified products and unqualified products to collect multiple samples through the machine vision system, as the training material of the quality inspection platform based on deep learning YOLOV5 algorithm, standard qualified products and unqualified products The splicing operation is also required after the material is collected, and the final quality inspection platform will greatly improve the accuracy and efficiency of quality inspection after a large number of material training;

The quality inspection platform 100 is essentially a visual server, which is equipped with various visual algorithms. The quality inspection platform 100 adopts the YOLOV5 algorithm to collect the characteristics of the image of the product to be inspected. When the similarity IOU value of the surface diagram is greater than the set similarity threshold, the judgment of product qualification is output. This judgment method is to compare the six-sided diagram of the standard qualified product first, that is, the priority presumption of qualification, which is more suitable for products with a qualification rate greater than 90%. Defect quality inspection can greatly improve the efficiency of defect quality inspection.

In step 2, the YOLOV5 target detection algorithm is used to detect the defect features of multiple sets of six-sided photos of unqualified products, and a defect feature comparison library is formed.

The yolov5 deep learning network of the YOLOV5 target detection algorithm model uses the data of the defect feature comparison library for iterative training, scales the width and height of the pictures in the defect feature comparison library and fills the gray edges to a size of 224×224 input, and the training image samples pass through the Focus network structure Three effective feature layers are extracted (52, 52, 256), (26, 26, 512) and (13, 13, 1024), and the fourth feature FPN layer is constructed based on the three effective feature layers.

After the defect features in the defect feature comparison library are captured by the YOLOV5 target detection algorithm, the feature layer based on the defect feature will be directly obtained, and finally divided into three effective feature layers according to the size of the defect feature, where the feature layer is ( 52, 52, 256) training image samples correspond to small targets, the feature layer is (26, 26, 512) detection granularity grid corresponding to the medium target, and the feature layer is (13, 13, 1024) detection granularity grid corresponds to large targets, for the 3 The fourth feature FPN layer can be obtained after the initial feature layer is subjected to convolution processing.

The features of the three effective feature layers (52, 52, 256), (26, 26, 512) and (13, 13, 1024) are extracted and fused by 3×3 convolution, and then performed on the fused feature map The prediction gets the fourth feature FPN layer.

The Focus network structure converts the information on the w-h plane to the channel dimension, and then extracts different features through 3×3 convolution. In this way, the information loss caused by downsampling can be reduced, and the features represented by the fourth feature FPN layer are more accurate.

Sampling at least four reference feature points on the fourth feature FPN layer. The reference feature point refers to a feature area less than 3 pixels, and the degree of light and shade change in the feature area is greater than 50%. The four reference feature points are extracted respectively and constructed based on The defect rapid comparison layer of the four reference feature points, and the distance detection algorithm is used to measure the mutual distance between the four reference feature points, and the distance data is used as an additional comparison item of the defect rapid comparison layer.

A light and dark detection algorithm is used to detect changes in light and dark on the fourth feature FPN layer.

Because there are many defect features in most unqualified products, we use the model of four reference feature points to simplify the comparison. The four reference feature points plus the position and distance relationship data can accurately locate the defect feature. Similar to the principle of fingerprint feature point recognition, the biggest feature of defect features is the change of light and shade. Therefore, using the light and shade detection algorithm to detect the change of light and shade can quickly obtain four reference feature points.

The machine vision defect quality inspection method also includes step 41, outputting product defect features, performing defect feature traversal before comparing the similarity between the product to-be-inspected image and the standard qualified product six-sided image, and calling the defect quick comparison of the defect feature comparison library The layer traverses the image of the product to be inspected. If the features of the four reference feature points are successfully matched, the matching similarity IOU value is greater than 0.9, and the mutual distance difference between the four reference feature points is less than the distance difference threshold, the quality inspection platform 100 Directly determine that the product is unqualified, and at the same time frame the defect location and display it based on the frame selection mark.

，距离差异阈值设0.05～0.1。The distance difference value refers to the absolute value calculation of the difference between the distance value between any two reference feature points in the product to-be-inspected map and the distance value between the corresponding two reference feature points in the defect quick comparison layer, and then the absolute value and the defect In the rapid comparison layer, the ratio operation is performed on the distance value corresponding to two reference feature points, and the distance between any two reference feature points in the product to-be-inspected image is set as S, and the distance between the two reference feature points in the defect rapid comparison layer is distance value S1, and the distance difference threshold is D, then

, and the distance difference threshold is set to 0.05-0.1.

The quality inspection platform 100 needs to mark the specific defect location and defect type while making the unqualified judgment of the product to be inspected. Only such product defect quality inspection will give positive feedback to the production of the product, which is more conducive to the product qualification rate promotion;

Specifically, in order to mark the specific defect location and defect type, the defect feature similarity must be compared based on the defect quick comparison layer. If the similarity IOU value is greater than 0.9, the defect type is determined, and then the position data For comparison, the concept of distance difference threshold is introduced here. The distance difference threshold is actually a key to locate the defect location. If the matching is successful, the location of the defect location is completed. Finally, the quality inspection platform 100 outputs unqualified products to be inspected. At the same time, the unqualified areas are framed and marked, and the type of defect is marked, so as to provide feedback data for the quality control of the product to be inspected, which is conducive to the improvement of product quality.

A machine vision system used to obtain the six-sided image of the product to be inspected, including a transparent platform for supporting the product to be inspected, a bottom camera 1, a gantry and a top camera module, and the transparent platform is set on the production line of the product to be inspected At the exit, the bottom camera 1 is set under the transparent platform to take the bottom view of the product to be inspected, the gantry is set above the transparent platform, and the top camera module includes the front camera 2, the rear camera 3, the left side Camera 4, right camera 5 and overlooking camera 6, front camera 2, rear camera 3, left camera 4, right camera 5 and overlooking camera 6 are respectively mounted on the gantry, and are used to photograph the products to be inspected respectively The front view, rear view, left side view, right side view and top view, the bottom camera 1, the front camera 2, the rear camera 3, the left camera 4, the right camera 5 and the top view camera 6 are respectively connected with the quality inspection The platform 100 is communicatively connected.

The traditional machine vision system can only obtain the picture above the product. The above structure can complete the acquisition of the front view, rear view, left view, right view and top view of the inspected product, and provide the same state for the subsequent quality inspection at one time. The following picture of the six sides to be inspected is taken at the same time, so the light and shadow conditions are consistent, which is more conducive to the follow-up quality inspection platform 100 to make a judgment on the quality inspection result.

The implementation principle of a machine vision defect quality inspection method based on YOLOV5 in the embodiment of the present invention is as follows:

In a specific product quality inspection scenario of an industrial production line, the product to be inspected after processing at a certain moment is transported from the industrial assembly line to the designated position on the transparent platform, the bottom camera 1, the front camera 2, the rear camera 3, the left camera 4, The right camera 5 and the top-view camera 6 simultaneously take pictures to be inspected and transmit them to the quality inspection platform 100 .

The quality inspection platform 100 runs an image stitching algorithm to splice the six-sided images to be inspected into a product image to be inspected, and inputs the product image to be inspected to the quality inspection platform 100,

First perform defect feature traversal, call the defect quick comparison layer of the defect feature comparison library to traverse the product to-be-inspected image, and for the defect features of surface cracks, the features of the four reference feature points are successfully matched, and the matching similarity IOU values are all greater than 0.9. And the mutual distance differences between the four reference feature points are all less than the distance difference threshold value of 0.05, then the quality inspection platform 100 directly determines that the product is unqualified, and at the same time performs frame selection processing on the defect position, and the display of the quality inspection platform 100 is based on the frame selection mark. The display shows that the type of defect is a crack. After the staff finds the above-mentioned unqualified products, they analyze the defects, and then adjust the process, which plays a feedback role in improving the product qualification rate.

All the above are preferred embodiments of the present invention, not to limit the protection scope of the present invention, so: all equivalent changes made according to the structure, shape and principle of the present invention should be covered by the protection scope of the present invention Inside.

Claims (7)

1. A quality inspection method for machine vision defects based on YOLOV5 is characterized in that: the method comprises the following specific steps:

step 1, collecting a plurality of six-sided photos of standard qualified products, and collecting a plurality of six-sided photos of unqualified products with different defects;

step 2, respectively running an image stitching algorithm on six photos of a standard qualified product and a unqualified product to regenerate a six-sided image of the standard qualified product and a six-sided image of the unqualified product, wherein the six-sided image of the standard qualified product comprises all appearance characteristics of six faces of the standard qualified product, the six-sided image of the unqualified product comprises all appearance characteristics of six faces of the unqualified product, a six-sided standard image training library of the qualified product is formed by a plurality of six-sided images of the standard qualified product, and a six-sided image training library of the unqualified product is formed by a plurality of six-sided images of the unqualified product;

step 3, a quality inspection platform (100) based on a YOLOV5 algorithm is built, and the quality inspection platform (100) is trained by adopting the six-face standard diagram training library of the qualified products and the six-face diagram training library of the unqualified products generated in the step 2;

step 4, judging whether the product to be detected is qualified or not, acquiring six-sided images to be detected of the product to be detected by adopting a machine vision system, splicing the six-sided images to be detected into a product image to be detected by running an image splicing algorithm, inputting the product image to be detected into a quality inspection platform (100), acquiring the characteristics of the product image to be detected by the quality inspection platform (100) by adopting a YOLOV5 algorithm, and outputting the judgment of the product qualification when the similarity IOU value of the product image to be detected and the six-sided image of the standard qualified product is larger than a set similarity threshold value through similarity comparison;

outputting the unqualified judgment of the product when the similarity IOU value of the six-surface to-be-detected drawing of the product and the six-surface drawing of the standard qualified product is smaller than or equal to a set similarity threshold value;

step 4 further includes step 41, outputting product defect characteristics, namely performing defect characteristic traversal before comparing the similarity of the product to be detected with the six-sided drawing pieces of the standard qualified product, invoking a defect quick comparison layer of a defect characteristic comparison library to traverse the product to be detected, and if the matching is successful, matching the characteristics of four reference characteristic points, wherein the similarity IOU value is greater than 0.9, and the mutual distance difference between the four reference characteristic points is less than a distance difference threshold, directly judging that the product is unqualified by a quality inspection platform (100), and meanwhile, performing frame selection processing on the defect positions and displaying based on frame selection marks;

the distance difference value refers to the absolute value calculation of the difference between the distance value between any two reference feature points of the product to be detected and the distance value between the corresponding two reference feature points in the defect rapid comparison layer, then the ratio operation is carried out on the absolute value and the distance value between the corresponding two reference feature points in the defect rapid comparison layer, the distance value between any two reference feature points of the product to be detected is set as S1, the distance difference threshold value is set as D, and then

The distance difference threshold is set to 0.05-0.1.

2. A YOLOV 5-based machine vision defect quality inspection method according to claim 1, characterized in that: in the step 2, a Yolov5 target detection algorithm is adopted to detect the defect characteristics of six-sided photos of a plurality of sets of unqualified products, and a defect characteristic comparison library is formed.

3. A YOLOV 5-based machine vision defect quality inspection method according to claim 2, characterized in that: the YOLOV5 deep learning network of the YOLOV5 target detection algorithm model uses defect feature comparison library data to carry out iterative training, the width and height equal proportion of pictures in the defect feature comparison library are scaled and filled with gray edges to 224×224 size input, a training image sample is subjected to Focus network structure to extract three effective feature layers (52,52,256), (26,26,512) and (13,13,1024) respectively, and a fourth feature FPN layer is constructed based on the three effective feature layers.

4. A YOLOV 5-based machine vision defect quality inspection method according to claim 3, characterized in that: and extracting different features from the features of the three effective feature layers (52,52,256), (26,26,512) and (13,13,1024) in a 3×3 convolution mode, fusing, and predicting on the fused feature map to obtain a fourth feature FPN layer.

5. The YOLOV 5-based machine vision defect quality inspection method of claim 4, wherein: sampling at least four reference feature points on a fourth feature FPN layer, wherein the reference feature points are feature areas smaller than 3 pixel points, the brightness variation degree in the feature areas is larger than 50%, respectively extracting the four reference feature points and constructing a defect rapid comparison layer based on the four reference feature points, measuring the mutual distance between the four reference feature points by adopting a distance detection algorithm, and taking distance data as an additional comparison item of the defect rapid comparison layer.

6. The YOLOV 5-based machine vision defect quality inspection method of claim 5, wherein: and (3) detecting the brightness change of the fourth characteristic FPN layer by adopting a brightness detection algorithm, marking all areas with the brightness change degree of more than 50%, and determining four characteristic points with the maximum brightness change degree as reference characteristic points.

7. A machine vision system, characterized by: the method for inspecting quality of machine vision defects based on YOLOV5 used for obtaining six-sided images to be inspected of products to be inspected comprises a transparent platform for supporting the products to be inspected, a bottom camera (1), a portal frame and a top camera module, wherein the transparent platform is arranged at an outlet of a production line of the products to be inspected, the bottom camera (1) is arranged below the transparent platform and is used for shooting bottom views of the products to be inspected, the portal frame is arranged above the transparent platform, the top camera module comprises a front camera (2), a rear camera (3), a left side camera (4), a right side camera (5) and a top camera (6), and the front camera (2), the rear camera (3), the left side camera (4), the right side camera (5) and the top camera (6) are respectively mounted on the portal frame and are respectively used for shooting front views, rear views, left side views, right side views and top views of the products to be inspected, and the bottom camera (1), the front camera (2), the rear camera (3), the left side camera (4), the right side camera (5) and the top camera (6) are respectively connected with the quality inspection platform (100).

Images