CN115661071A - Detection and evaluation method of surface defects in composite material processing based on deep learning - Google Patents

Abstract

The invention belongs to the technical field related to precision machining detection, and discloses a method for detecting and evaluating surface defects of composite material processing based on deep learning, including: (1) taking defect images and performing data marking on the images; (2) dividing training (3) Input the training set into the deep learning model for training; (4) Input the test set into the trained model to obtain the category, location, and area of the defect in the image Area and model evaluation index; (5) Form the mapping of evaluation index to evaluation score. The invention also discloses a corresponding system. Through the present invention, high-accuracy and high-efficiency particle-reinforced composite material processing surface defect detection can be realized, the quantitative evaluation of detected defects can be realized, and the selection of optimal process parameters can be effectively guided, so it is especially suitable for particle-reinforced composite materials. Application occasions for detection and evaluation of surface defects in material processing.

Description

Detection and evaluation method of surface defects in composite material processing based on deep learning

technical field

The invention belongs to the technical field related to precision machining detection, and more specifically relates to a method for detecting and evaluating surface defects of composite material processing based on deep learning.

Background technique

With the application of spacecraft in space and near space, the structural volume, weight and manufacturing accuracy of the core optical components of various systems are getting more and more attention. The only way to high-quality comprehensive performance. A single material cannot always meet certain needs in the industrial field, so particle-reinforced composite materials that can meet these needs are widely used.

Composite materials have small thermal expansion coefficient, high strength and high rigidity, and excellent properties such as wear resistance, corrosion resistance, and high temperature resistance. They are an ideal new lightweight material. However, due to the large difference in physical and mechanical properties between the metal matrix and the reinforced phase of the composite material, it is easy to form various forms of defects during the processing process, such as furrows, cracks, particle breakage, protrusions, and interface peeling, etc., which will cause serious damage. affect material properties.

In the prior art, the methods for surface defect detection of particle-reinforced material processing mainly include traditional artificial naked eye inspection, ultrasonic inspection, X-ray inspection, high-frequency pulsed eddy current inspection, and the like. Among them, manual inspection is biased towards subjective judgment, and small defects are found at the same time, so the efficiency and accuracy are very low; ultrasonic inspection is currently a relatively common and widely used inspection technology for composite materials, but its defect display is not intuitive enough, and it is difficult to identify and quantify defects , requires couplant and is mainly suitable for internal defect detection; eddy current testing requires the material itself to be conductive, and requires professional analysis and judgment. Most of these methods require manual or semi-manual methods for defect judgment, resulting in low efficiency.

Correspondingly, there is an urgent need for further research and improvement in this field in order to better meet the needs of high-precision and high-efficiency detection and quantitative evaluation of composite material processing surfaces.

Contents of the invention

In view of the above defects or requirements of the prior art, the purpose of the present invention is to provide a method for detecting and evaluating surface defects of composite materials processing based on deep learning, wherein by fully considering the relevant characteristics of surface defects of composite materials processing, the deep learning algorithm is selected and Targeted design training, testing, and evaluation operations can further improve the model's ability to detect surface defects in composite materials compared with existing technologies, and obtain high-accuracy, high-efficiency detection of surface defects in processing and more comprehensive and quantitative results. Evaluation results.

In order to achieve the above object, according to one aspect of the present invention, a method for detecting and evaluating surface defects in composite material processing based on deep learning is provided, characterized in that the method includes:

Step 1. Image acquisition and labeling

Take images of the processed surface of composite materials, collect the obtained images into an image set and perform defect annotation, thereby forming an image data set;

Step 2. Image dataset division and data enhancement

Divide the image data set formed in step 1 into a training set and a test set, which are used to train the model and test the model respectively, and simultaneously perform image data enhancement processing on the divided training set to expand the scale of the training set;

Step 3. Defect detection model training

Input the training set obtained in step 2 into the deep learning model for model training;

Step 4. Defect detection model testing

Input the test set obtained in step 2 into the trained deep learning model for testing, and obtain information such as defect category, defect location, defect depth, defect area, defect area ratio, length and width of the smallest circumscribed rectangle in the image, etc. , so as to obtain the corresponding defect evaluation index value;

Step 5. Defect Evaluation

Based on the defect evaluation index value obtained in step 4, combined with the preset defect evaluation score criterion, the mapping relationship between the evaluation index and the evaluation score is formed, thereby completing the entire detection and evaluation process.

As a further preference, in step 1, the number of images obtained is preferably not less than 800, and the corresponding number of each defect is preferably not less than 200; the defect annotation preferably includes: defect type, coordinate information of defect area frame, Defect instance boundary point, defect instance area, etc.

As further preferably, in step 2, the division ratio of the training set and the test set is preferably 8:2, and methods such as rotation, scaling, shearing, Mosaic, CutMix, etc. can be used to complete the image data of the training set Enhanced processing.

As a further preference, for the deep learning model, its preferred settings are as follows: its backbone uses the ResNet network for feature extraction, combined with the FPN network to output feature maps of different sizes; generates proposals through the RPN network, Fast-RCNN network Perform category prediction and position fine-tuning on the proposals generated by RPN; the Mask branch generates Masks of all categories, and extracts the Mask corresponding to the predicted category.

As further preferably, in step 3, the process of the model training is preferably designed as follows:

Carry out K-means clustering on the defect annotation frame of the training set to obtain the appropriate anchor size; extract feature maps of different levels through the backbone and FPN network, obtain proposals through the RPN network, and then map them back to the corresponding level feature map Get the proposal feature map, and its corresponding relationship is:

Among them, k ₀ is the number of layers mapped by w h = S ² , and w and h are the width and height of the proposal respectively;

In addition, the feature maps of different levels of proposals are converted into the same size through RoIAlign, and then through two fully connected layers, and finally through two parallel fully connected layers to realize the category prediction of the feature map and the prediction of the proposal offset. The Mask branch is in The input target during training is the proposals provided by RPN.

As further preferably, in step 3, the loss of the model training preferably includes RPN network loss, Fast-RCNN loss, Mask loss, wherein

The related RPN loss function is designed as:

当为正样本时为1，负样本时为0，N_reg为anchor位置点个数，t_i为预测第i个anchor对应回归参数，

为第i个anchor对应GTBox的回归参数；Among them, N _cls is the number of candidate frames selected to calculate the loss for a picture, p _i is the probability that the i-th anchor is predicted to be a positive sample,

When it is a positive sample, it is 1, when it is a negative sample, it is 0, N _reg is the number of anchor position points, t _i is the regression parameter corresponding to the predicted i-th anchor,

It is the regression parameter corresponding to GTBox for the i-th anchor;

The related Fast-RCNN loss function is designed as:

L(p,u,t ^u ,v)=L _cls (p,u)+λ[u≥1]L _loc (t ^u ,v)

Among them, t ^u is the regression parameter for predicting the corresponding category u, and v corresponds to the bounding box regression parameter of the real target;

The related Mask loss function is designed as:

L(m,n)=L _BCE (m,n)

Among them, m is the Mask corresponding to the predicted category, and n is the GT Mask.

As further preferably, in step 4, the process of the model testing is preferably designed as follows:

Multiple feature maps are obtained through backbone and FPN, the corresponding proposals of each feature map are generated through the RPN network, and these proposals are mapped to the corresponding feature maps to obtain the proposal feature maps; then, through RoIAlign, two fully connected layers, two Connect the fully connected layer in parallel to obtain a prediction category and related offset corresponding to a proposal, map the offset proposals output by the Fast-RCNN network back to the feature map, change the size through RoIAlign, and input it to the Mask branch, and select the corresponding target prediction category The Mask is mapped back to the original image.

As a further preference, in step 4, the defect area, the minimum circumscribed rectangle of the defect area and the defect type can preferably be obtained through the target detection algorithm based on deep learning, wherein the defect area can preferably be obtained by converting the pixel area and scale in the figure , the minimum circumscribed rectangle of the defect area can preferably be obtained by the rotation jamming algorithm.

As a further preference, in step 5, the defect evaluation score criteria can preferably be selected from mechanical properties, physical properties, chemical properties, service life, etc., wherein the mechanical properties further include yield strength, shear modulus, area shrinkage, etc., The physical properties further include resistivity, thermal conductivity, refractive index, etc., and the chemical properties further include corrosion resistance, oxidation resistance, etc., and the proportion of these parameters is adjusted through weighting coefficients.

As a further preference, in step five, the mapping relationship between the evaluation index and the evaluation score is preferably established through an artificial neural network, wherein the defect category is represented by weighting the corresponding parameters of different types of defects.

According to another aspect of the present invention, there is also provided a corresponding deep learning-based composite material processing surface defect detection and evaluation, characterized in that the system includes:

An image acquisition and labeling module, the image acquisition and labeling module is used to take images of the composite material processing surface, collect the obtained images into an image set and perform defect labeling, thereby forming an image data set;

Image data set division and data enhancement module, the image data set division and data enhancement module is used to divide the formed image data set into training set and test set, which are respectively used for training model and test model, and at the same time, the divided training set The image data set is enhanced to expand the size of the training set;

A defect detection model training module, the defect detection model training module is used to input the obtained training set into the deep learning model for model training;

Defect detection model test module, the defect detection model test module is used to input the obtained test set into the trained deep learning model for testing, and obtain the defect category, defect position, defect depth, defect area and defect area in the image The proportion, the length and width of the minimum circumscribed rectangle of the defect area, etc., so as to obtain the corresponding defect evaluation index value;

The defect evaluation module is used to form the mapping relationship between the evaluation index and the evaluation score based on the obtained defect evaluation index value and combined with the preset defect evaluation score criterion, thus completing the entire detection and evaluation process .

Generally speaking, compared with the prior art, the above technical solution conceived by the present invention has the following

Beneficial effect:

(1) Compared with traditional manual inspection with naked eyes, ultrasonic inspection and X-ray inspection, the present invention can effectively improve the efficiency and accuracy of defect identification, and can avoid subjective interference caused by manual or semi-manual methods in traditional methods;

(2) The present invention can effectively improve the generalization ability of the defect detection model through a large number of defect samples and data enhancement methods;

(3) The present invention can effectively improve the ability to identify defects of different sizes and reduce the rate of missed detection of defects through the combination of feature extraction network and FPN network;

(4) The present invention further proposes a defect evaluation system, which can make up for the problem that the defect evaluation standard comes from the subjective judgment of the inspectors, and there is no set of good quantitative evaluation methods. This method can be used to quantify defects to guide processing, which can effectively promote The processing of granular composite materials is developing towards intelligent manufacturing.

Description of drawings

Fig. 1 is the overall flowchart of the composite material processing surface defect detection and evaluation method based on deep learning according to the present invention;

Fig. 2 is a schematic diagram for exemplary display of a defect detection model according to a preferred embodiment of the present invention;

Fig. 3 is a schematic diagram of a feature extraction network for exemplary display combining FPN according to a preferred embodiment of the present invention;

Fig. 4 is a schematic flow chart for an exemplary display defect evaluation stage according to a preferred embodiment of the present invention;

Fig. 5 is a schematic diagram of the actual output of the defect detection model according to a preferred embodiment of the present invention.

Detailed ways

In order to make the purpose, 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 in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments It is a part of embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Fig. 1 is an overall flowchart of the method for detecting and evaluating surface defects of composite material processing based on deep learning according to the present invention. The present invention will be explained in more detail below with reference to FIG. 1 .

First, there are image acquisition and labeling steps.

In this step, images of the processed surface of the composite material are taken, and the obtained images are assembled into an image set and defect annotation is performed, thereby forming an image data set.

More specifically, for example, images of the processed surface of particle-reinforced composite materials can be taken by an electron microscope, and the acquired images can be collected into an image set, and defects can be marked on the image set with appropriate data labeling software to form an image data set.

For example, 1,000 aluminum-based silicon carbide defect pictures can be obtained through SEM, and at the same time, the corresponding number of each defect is not less than 200. Create a COCO dataset through labelme, label the defect type, defect area frame information (upper left corner coordinates, height and width), defect instance polygon boundary points, and defect area.

Next, is the image data set division and data enhancement steps.

In this step, the previously formed image data set is divided into a training set and a test set, which are used to train the model and test the model respectively, and image data enhancement processing is performed on the divided training set to expand the scale of the training set.

More specifically, it is preferable to enhance the picture data to 3000 by methods such as rotation, scaling, flipping, CutMix, etc., and divide the pictures into a training set and a test set according to a ratio of 8:2, that is, 2400 training set pictures and 600 test sets set pictures.

Next, is the defect detection model training step.

In this step, the obtained training set is input into the deep learning model for model training.

More specifically, the present invention can adopt the Mask-RCNN algorithm, for example, as shown in FIG. 2 , the model in this embodiment includes a feature extraction network (ResNet50+FPN), an RPN network, a Fast-RCNN network, and a Mask branch. As shown in Figure 3, using a feature extraction network with FPN can effectively detect defects of different sizes. In the training process, it is first necessary to carry out K-means clustering on the defect labeling boxes of the training set to obtain a suitable anchor size. Before training the entire model, pre-train the backbone, and solve the problem of small amount of data and slow network convergence through transfer learning. question. During the training process, the output of the RPN network is used by the Fast-RCNN network and the Mask branch, and the proposals are mapped back to the corresponding feature map according to the size. The mapping formula is:

Among them, k ₀ is the number of layers mapped by w·h=S ² , and w and h are the width and height of the proposal respectively.

The two branches use two different ROIAlign to convert the size of the proposal feature map, and the relevant weights need to be trained separately.

According to a preferred embodiment of the present invention, the total training loss includes RPN network loss, Fast-RCNN loss, and Mask loss.

RPN loss function:

当为正样本时为1，负样本时为0，N_reg为anchor位置点个数，t_i为预测第i个anchor对应回归参数，

为第i个anchor对应GTBox的回归参数。Among them, N _cls is the number of candidate frames selected to calculate the loss for a picture, p _i is the probability that the i-th anchor is predicted to be a positive sample,

When it is a positive sample, it is 1, and when it is a negative sample, it is 0. N _reg is the number of anchor position points, and t _i is the regression parameter corresponding to the predicted i-th anchor.

It is the regression parameter corresponding to GTBox for the i-th anchor.

Fast-RCNN loss function:

L(p,u,t ^u ,v)=L _clS (p,u)+λ[u≥1]L _loc (t ^u ,v)

Among them, t ^u is the regression parameter for predicting the corresponding category u, and v corresponds to the bounding box regression parameter of the real object.

Mask loss function:

L(m,n)=L _BCE (m,n)

Among them, m is the Mask corresponding to the predicted category, and n is the GT Mask.

In addition, the model training environment can use the pytorch framework, the hardware uses GeForce RTX 3080, and the SGD algorithm with momentum is used for gradient update. The momentum parameter is set to 0.9, the learning rate is 0.0004, and the weight decay is 0.0001. A total of 26 epochs are iterated, and the 16th , 22 epochs carry out the learning rate decay of the gradient 0.1 ratio.

Next, is the defect detection model testing step.

In this step, input the test set obtained above into the trained deep learning model for testing, and obtain the defect category, defect position, defect depth, defect area, defect area ratio, and the smallest circumscribed rectangle of the defect area in the image. information such as length and width, so as to obtain the corresponding defect evaluation index value.

More specifically, the weights obtained from training can be used to input the test set pictures into the defect detection model. The test process is shown in Figure 3. First, the final prediction frame and prediction category are obtained through the Fast-RCNN network, and then used as the input of the Mask branch , to get the corresponding Mask of this category, and then map back to the original image to get the final required defect category, location, area and other parameters. The model evaluation index of target detection and image segmentation uses mAP, mAP is the average precision of all categories, mAP The larger the value, the better the performance of the model. If the mAP value of the final test set meets the requirements, the trained model can be used for actual defect picture detection.

Finally, there is the defect evaluation step.

In this step, based on the obtained defect evaluation index value and combined with the preset defect evaluation score criterion, a mapping relationship between the evaluation index and the evaluation score is formed, thereby completing the entire detection and evaluation process.

More specifically, the defect evaluation in the present invention includes selecting defect evaluation indicators and evaluation score criteria to form a mapping of evaluation indicators to evaluation scores. The entire defect evaluation process is shown in Figure 4, where the defect evaluation indicators include defect area, defect area ratio, length and width of the smallest circumscribed rectangle of the defect area, defect type, defect depth, etc. The type of defect is directly obtained from the deep learning model; the defect area, the proportion of the defect area, the length and width of the smallest circumscribed rectangle of the defect area, etc., are obtained by the post-processing algorithm on the output of the deep learning model; the depth of the defect cannot be obtained directly from the two-dimensional image Indicators obtained indirectly need to be obtained through other measurement methods. For example, the defect area can be obtained by converting the pixel area and scale in the figure, the minimum circumscribed rectangle of the defect area can be obtained by the rotation jamming algorithm, and the defect depth can be obtained by SEM measurement.

Evaluation criteria can be selected from mechanical properties (yield strength, shear modulus, area shrinkage, etc.), physical properties (resistivity, thermal conductivity, refractive index, etc.), chemical properties (corrosion resistance, oxidation resistance, etc.) , service life, etc., the parameters can be selected according to the material use background. When selecting multiple parameters, considering the different units and numerical magnitudes of different parameters, it is necessary to perform normalization processing, and it is also necessary to adjust the proportion of multiple parameters through weighting coefficients. For example, aluminum-based silicon carbide materials are processed to make space mirrors. The performance parameters such as reflection performance, thermal conductivity, and yield strength meet the requirements of use. The selected parameters are normalized to the interval [0,1], and the sum of the weighting coefficients is 1, that is, the score formula can be expressed as:

score=w ₁ a+w ₂ b+w ₃ c

Among them, a, b, and c are the normalized values of reflection performance, thermal conductivity, and yield strength, w ₁ , w ₂ , and w ₃ are the weighting coefficients of the corresponding parameters, and w ₁ +w ₂ +w ₃ = 1.

According to another preferred embodiment of the present invention, the mapping relationship between the evaluation index and the evaluation score is established through an artificial neural network, that is, the mapping weight between multi-layer neurons and a nonlinear activation function through a large amount of data self-learning. The type of defect is reflected by weighting the corresponding parameters of different types of defects, such as the common particle breakage and scratch defects of aluminum-based silicon carbide. When performing neural network mapping, the relevant parameters of the two defects are individually used as input neurons. The weights representing defect categories are also self-learned via backpropagation.

In summary, the present invention can realize high-accuracy and high-efficiency surface defect detection of particle-reinforced composite materials, realize quantitative evaluation of detected defects, and can effectively guide the selection of optimal process parameters, so it is especially suitable for particle-reinforced composite materials. Application occasions for detection and evaluation of surface defects of composite materials. .

It is easy for those skilled in the art to understand that the above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention, All should be included within the protection scope of the present invention.

Claims (10)

1. A composite material processing surface defect detection and evaluation method based on deep learning is characterized by comprising the following steps:

step one, image acquisition and labeling

Shooting images of the processed surface of the composite material, summarizing the obtained images into an image set and executing defect labeling, thereby forming an image data set;

step two, dividing the image data set and enhancing the data

Dividing the image data set formed in the step one into a training set and a testing set, respectively training the model and the testing model, and simultaneously carrying out image data enhancement processing on the divided training set to enlarge the scale of the training set;

step three, training a defect detection model

Inputting the training set obtained in the step two into a deep learning model for model training;

step four, testing a defect detection model

Inputting the test set obtained in the second step into a trained deep learning model for testing to obtain information such as defect category, defect position, defect depth, defect area ratio, length and width of a minimum circumscribed rectangle of a defect area and the like in the image, thereby obtaining a corresponding defect evaluation index value;

step five, defect evaluation

And forming a mapping relation between the evaluation indexes and the evaluation scores based on the defect evaluation index values obtained in the step four and by combining a preset defect evaluation score criterion, thereby completing the whole detection and evaluation process.

2. The method for detecting and evaluating defects on composite material processing surfaces according to claim 1, wherein in step one, the number of images obtained is preferably not less than 800, and the corresponding number of each defect is preferably not less than 200; the defect labeling preferably comprises: defect type, coordinate information of defect region box, defect instance boundary point, defect instance area, etc.

3. The method for detecting and evaluating defects on composite material processing surfaces according to claim 1 or 2, wherein in step two, the training set and the test set are preferably divided into 8:2, and rotation, scaling, cutting, mosaic, cutMix and other methods can be adopted to complete the image data enhancement processing of the training set.

4. The method for detecting and evaluating defects on a composite material processing surface according to any one of claims 1 to 3, wherein the deep learning model is preferably set as follows: the backbone of the method uses a ResNet network to extract features and combines with an FPN network to output feature maps with different sizes; generating propulses through an RPN (resilient packet network), and performing category prediction and position fine adjustment on the propulses generated by the RPN through a Fast-RCNN (Fast-RCNN) network; and generating masks of all categories by Mask branches, and extracting masks corresponding to the prediction categories.

5. The method for detecting and evaluating the defects on the processed surface of the composite material according to any one of the claims 1 to 4, wherein in the third step, the model training process is preferably designed as follows:

performing K-means clustering on the defect marking frame of the training set to obtain a proper anchor size; extracting feature graphs of different levels through a backbone network and an FPN network, obtaining explosals through an RPN network, and mapping the obtained explosals feature graphs to corresponding level feature graphs to obtain the explosals feature graphs, wherein the corresponding relation is as follows:

wherein k is ₀ Is w · h = S ² The mapped layer number, w and h are the width and the height of the propofol respectively;

in addition, the propusals feature graphs of different levels are converted into the same size through RoIAlign, class prediction and propusal offset prediction of the feature graphs are achieved through two full-connection layers and finally through two parallel full-connection layers, and input targets of Mask branches are propusals provided by RPNs during training.

6. The method according to claim 5, wherein the loss of model training in step three preferably comprises RPN network loss, fast-RCNN loss and Mask loss, wherein

The associated RPN loss function is designed to:

wherein N is _cls Selecting the number of candidate frames for calculating loss, p, for a picture _i The probability of being a positive sample is predicted for the ith anchor,

1 in the case of positive samples and 0 in the case of negative samples _reg Number of anchors, t _i To predict the regression parameters corresponding to the ith anchor,

the regression parameter of the GTBox corresponding to the ith anchor is taken as the regression parameter of the GTBox;

the relevant Fast-RCNN loss function is designed as:

L(p，u，t ^u ，v)＝L _cls (p，u)+λ[u≥1]L _loc (t ^u ，v)

wherein, t ^u V is a bounding box regression parameter corresponding to the real target for predicting the regression parameter corresponding to the category u;

the relevant Mask loss function is designed as:

L(m，n)＝L _BCE (m，n)

wherein m is Mask corresponding to the prediction category, and n is GT Mask.

7. The method for detecting and evaluating the defects on the processed surface of the composite material according to any one of claims 1 to 6, wherein in the fourth step, the model test process is preferably designed as follows:

obtaining a plurality of feature maps through a backbone and an FPN, generating corresponding explosals of each feature map through an RPN, and mapping the explosals onto the corresponding feature maps to obtain the explosals feature maps; and then, obtaining a prediction type and a related offset corresponding to the propsall through RoIAlign, two full-connection layers and two parallel full-connection layers, mapping the offset propsals output by the Fast-RCNN network back to a characteristic diagram, changing the size through RoIAlign, inputting the result to a Mask branch, and selecting a Mask corresponding to the target prediction type to map back to the original image.

8. The method for detecting and evaluating the defects on the processed surface of the composite material as claimed in any one of claims 1 to 7, wherein in the fifth step, the defect evaluation score criterion is preferably selected from the group consisting of mechanical properties, physical properties, chemical properties, service life and the like, wherein the mechanical properties further include yield strength, shear modulus, section shrinkage and the like, the physical properties further include electrical resistivity, thermal conductivity, refractive index and the like, the chemical properties further include corrosion resistance, oxidation resistance and the like, and the parameters are proportionally adjusted by weighting coefficients.

9. The method for detecting and evaluating the defects on the processed surface of the composite material according to any one of claims 1 to 8, wherein in the fifth step, the mapping relation between the evaluation index and the evaluation score is established preferably through an artificial neural network, wherein the defect types are represented by weighting corresponding parameters of different types of defects.

10. A composite material processing surface defect detection and evaluation system based on deep learning is characterized by comprising:

an image acquisition and labeling module for capturing images of the composite material machined surface, summarizing the acquired images into an image set and performing defect labeling, thereby forming an image data set;

the image data set dividing and data enhancing module is used for dividing the formed image data set into a training set and a testing set, respectively training a model and the testing model, and simultaneously carrying out image data enhancing processing on the divided training set so as to enlarge the scale of the training set;

the defect detection model training module is used for inputting the obtained training set into the deep learning model to carry out model training;

the defect detection model testing module is used for inputting the obtained testing set into a trained deep learning model for testing to obtain information such as defect types, defect positions, defect depths, defect areas, defect area ratios, length and width of a minimum circumscribed rectangle of a defect area and the like in an image so as to obtain corresponding defect evaluation index values;

and the defect evaluation module is used for forming a mapping relation between the evaluation index and the evaluation score based on the acquired defect evaluation index value and in combination with a preset defect evaluation score criterion, so that the whole detection and evaluation process is completed.

Images