Stacked stamping part identification and defect detection method with visual deblurring function

Abstract

The invention relates to a stacked stamping part identification and defect detection method with visual deblurring, and the method comprises the steps: firstly, collecting the types and defect images of a plurality of parts, carrying out the image marking according to the types of the parts and the defects, such as deformation, scratches, unfilled corners, fractures, and the like, and constructing a sample data set; secondly, performing artificial defocusing and motion blur on the acquired image, and simulating oscillation of an industrial machine; then, constructing a deblurring network based on AlexNet and DeblurGan-v2, and carrying out deblurring operation on the artificially blurred image to obtain a clear image; the clear images are input into DETR, and accurate classification of part categories is achieved; and finally, the classified part images are sent to a YOLOv11 network for defect positioning segmentation, and a final identification and defect detection result is obtained. According to the method, classification and defect detection can be carried out on the stamping parts produced on the industrial assembly line in real time, enterprises can be assisted in supervising the production line in time, and the production efficiency of the enterprises is improved.

Description

Technical Field

[0001] The invention relates to the technical field of industrial parts recognition, in particular to a stacked stamping parts recognition and defect detection method with visual deblurring. Background Art

[0002] In recent years, with the improvement of science and technology, machine vision can replace the human eye to extract information from images, process and understand the information, make detection and judgment, and is currently widely used in aviation, shipbuilding, precision measurement, industrial manufacturing and other fields, making great contributions to national development and social life. With the progress and development of society, high-quality products have received more and more attention. In addition to affecting the appearance, the surface defects of the product will also affect the performance of the product, so defect detection is a key link to ensure the production of high-quality products. The stamping parts processing and manufacturing industry is an important part of my country's manufacturing industry. Among all stamping parts, small stamping parts have higher production precision requirements, the defects generated in the production process are smaller, and the difficulty of defect detection is also greater. Therefore, stamping parts identification and defect detection have important application value and practical significance in industrial production.

[0003] At present, the existing methods of industrial parts recognition and defect detection are mainly divided into three methods: manual detection, traditional image processing and special equipment detection. Manual detection is the most traditional and accurate method, which mainly relies on the detection personnel's memory of each part category and the naked eye observation of the surface defects of the parts. Although this method can capture subtle defects intuitively, it is easily affected by the experience and fatigue of the observer, and cannot quickly identify and detect a large number of parts, which is inefficient. Part recognition in traditional image processing methods extracts multiple feature points from different parts, and matches the extracted feature points with the feature points of standard parts to obtain part category information. Part defect detection locates the defect area by locating the defective area by the large difference between the pixel at the defect and the adjacent pixel. This method can realize real-time processing and reduce the influence of human factors, but it has high requirements on image quality and lighting conditions, and the accuracy and robustness of the algorithm still need to be further improved. Special equipment detection usually uses special large equipment to detect whether the parts have defects. Although these instruments can provide accurate measurement data, which helps to more accurately identify the part category and locate defects, the equipment cost is high, the operation may be relatively complicated, and there may be certain limitations for real-time detection. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a stacked stamping parts recognition and defect detection method with visual deblurring, which can classify and detect defects of stamping parts produced on industrial assembly lines in real time, assist enterprises in timely supervision of production lines, and improve enterprise production efficiency.

[0005] To achieve the above object, the present invention adopts the following technical solution: a stacked stamping parts recognition and defect detection method with visual deblurring, comprising the following steps:

[0006] Step 1: First collect the categories and defect images of various parts, and annotate the images according to the categories of parts and deformation, scratches, missing corners, and fracture defects to build a sample data set;

[0007] Step 2: Artificially defocus and motion blur the captured image to simulate industrial machine vibration;

[0008] Step 3: Build a deblurring network based on AlexNet and DeblurGan-v2 to deblur the artificially blurred image and obtain a clear image;

[0009] Step 4: Input the clear image into DETR to achieve accurate classification of part categories;

[0010] Step 5: Send the classified part images into the YOLOv11 network for defect location and segmentation to obtain the final recognition and defect detection results.

[0011] In a preferred embodiment, in step 1, the steps of collecting the categories and defect images of various parts are as follows:

[0012] (1) Five defect-free parts and their two corresponding defective parts were placed together, one good and one bad, and photographed. A total of 5 photos were taken. The placement and angle were random and there was no stacking. Ten photos were collected for each part, a total of 50 photos for the five parts, and 10 photos were taken for each of the remaining seven defect-free parts, for a total of 120 photos;

[0013] (2) Twelve defect-free parts were placed together, with random positions and angles and no stacking, for a total of 90 images; parts with ten types of defects were photographed in the same manner, for a total of 90 images, for a total of 180 images;

[0014] (3) Twelve defect-free parts and ten defective parts are placed together, with random placement and angle without stacking. Ten images are collected for each, for a total of 100 images;

[0015] (4) All parts were photographed with or without some stacking, with random placement and angle, for a total of 100 photos;

[0016] The part recognition and defect detection data set consists of 500 images, which total 1500 images after rotation and brightness change. The training set, validation set, and test set are divided in a ratio of 8:1:1, that is, 1200:150:150, to construct a part data set for training DETR and YOLOv11.

[0017] In a preferred embodiment, step 3 specifically includes: first, using the AlexNet network to perform binary classification of blurred and clear images, and then sending the blurred image to the DeblurGan-v2 network for deblurring; the AlexNet network structure is composed of 5 convolutional layers and 3 fully connected layers, and the AlexNet network structure also has fully connected layers FC7 and FC8, FC6 and FC7 contain Dropout, discarding some neurons to prevent overfitting, and the activation function in FC8 is Sigmoid, outputting the sample classification probability; DeblurGan-v2 uses the blurred image as input, generates a clear image through the generator G, and then inputs it into the discriminator D together with the real clear image for discrimination; the generator G then adjusts its own network parameters according to the result of the discrimination of the discriminator D, and generates data closer to the real in the next round of training; its generator is composed of the backbone network MobileNetv2 as feature extraction, a feature fusion module for top-down feature map, and a hollow space convolution pooling pyramid ASPP that enriches contextual information, and the discriminator is composed of a network layer composed of multiple convolution and activation functions.

[0018] In a preferred embodiment, step 4 specifically includes: first, it is necessary to make a blur judgment on the captured image, and then input the blurred image into the subsequent model after clarifying it; then, a part recognition network based on DETR is constructed. For a picture, DETR predicts a certain number of objects, and directly outputs a set of predictions based on the global information of the object, that is, Transformer decodes the prediction results of all objects in the picture at one time; finally, the classified part image is input into YOLOv11 for defect location. As an improvement of YOLOv8, the original C2f is replaced with C3K2, and C2PSA is added to the last layer of the backbone network, and the classification detection head in the head decoupling head is changed to DWConv to reduce the amount of calculation.

[0019] In a preferred embodiment, the DETR-based part recognition network specifically includes:

[0020] (1) CNN Backbone

[0021] CNN in Backbone is the feature extraction part; convolutional neural network ResNet-50;

[0022] (2) Transformer Encoder

[0023] After Backbone, the number of channels of the output feature map is 2048, and 1×1 convolution is used for dimensionality reduction before inputting into the Transformer Encoder.

[0024] (3) Transformer Decoder

[0025] The decoder has two inputs, one is the object queries, and the other is the global features transmitted from the image end. The two inputs are repeatedly self-attentioned and then output to the fully connected layer. In each decoder, the object query self-attention operation is performed first.

[0026] (4)Prediction Heads

[0027] The detection head, which is composed of a fully connected layer, makes two predictions after receiving the features passed in by the Decoder. One is the prediction of the object category, and the other is the prediction of the box. There are four values ​​in total, namely the center point of the box and the height and width of the box.

[0028] In a preferred embodiment, the network structure of YOLOv11 consists of three parts: Backbone, Neck, and Head; Backbone is responsible for extracting high-level semantic features from the input image, including multiple convolutional layers and pooling layers, and YOLOv11 replaces the previous C2f with C3K2; Neck is used to further integrate and adjust the features extracted by Backbone. Neck upsamples the deep features and splices them with the shallow features, fusing features at different levels to provide more global and local feature information; Head is the last layer of the model.

[0029] Compared with the prior art, the present invention has the following beneficial effects: This patent proposes a deblurring network to address the problem of image blur in actual industrial production, which can effectively improve the subsequent part recognition rate and defect location accuracy; at the same time, the proposed DETR part recognition network uses a combination of Transformer and CNN, which can more accurately and quickly identify part categories compared to ordinary CNN; and the YOLOv11 defect detection network effectively locates the defect position, improving the accuracy of positioning and segmentation of part defect areas. The method of the present invention can classify and detect defects of stamping parts produced on industrial assembly lines in real time, which can assist enterprises in timely supervision of production lines and improve enterprise production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A diagram showing the parts categories of a preferred embodiment of the present invention;

[0031] Figure 2 A defuzzification network structure diagram of a preferred embodiment of the present invention;

[0032] Figure 3 The overall network structure diagram of the preferred embodiment of the present invention;

[0033] Figure 4 A YOLOv11 network structure diagram of a preferred embodiment of the present invention;

[0034] Figure 5 is an overall flow chart of a preferred embodiment of the present invention;

[0035] Figure 6 Figure 1 is a partial data set diagram of a preferred embodiment of the present invention, wherein (a) is a motion blurred image, (b) is a part image, (c) is a stacked part image, and (d) is a defective part image;

[0036] Figure 7 is the AlexNet training index of the preferred embodiment of the present invention, wherein (a) is the AlexNet training loss, and (b) is the training accuracy;

[0037] Figure 8 is a preferred embodiment of the present invention, wherein (a) is a slightly blurred image, (b) is a deblurred image, (c) is a clear image, (d) is a moderately blurred image, (e) is a deblurred image, (f) is a clear image, (g) is a heavily blurred image, (h) is a deblurred image, and (i) is a clear image;

[0038] Fig. 9 DETR network training indicators of the preferred embodiment of the present invention, wherein (a) is the training loss curve, and (b) is the verification loss curve;

[0039] Fig.10 The comparison of the recognition results of the two methods of the preferred embodiment of the present invention, wherein (a) is the recognition of the network part without deblurring, (b) is the recognition of the image part after deblurring, and (c) is the recognition of the original clear image part;

[0040] Fig.11 It is the YOLOv11 network training index of the preferred embodiment of the present invention, where (a) is the training Box loss curve, (b) is the training Class loss curve, (c) is the training DFL loss curve, (d) is the verification Box loss curve, (e) is the verification Class loss curve, and (f) is the verification DFL loss curve. DETAILED DESCRIPTION

[0041] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0042] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0043] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application; as used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or their combinations.

[0044] A stamping parts recognition and defect detection method based on deep learning, reference Figure 1-11 First, collect the categories and defective images of various parts, and annotate the images according to the categories of parts and defects such as deformation, scratches, missing corners, and fractures to build a sample data set; secondly, artificially defocus and motion blur the collected images to simulate industrial machine vibrations; then build a deblurring network based on AlexNet and DeblurGan-v2 to deblur the artificially blurred images to obtain clear images; then input the clear images into DETR to achieve accurate classification of part categories; finally, send the classified part images to the YOLOv11 network for defect location and segmentation to obtain the final recognition and defect detection results. The method of the present invention can classify and detect defects of stamping parts produced on industrial assembly lines in real time, which can assist enterprises in timely supervision of production lines and improve enterprise production efficiency. The detailed technical scheme is as follows:

[0045] 1. Parts collection and data set construction

[0046] The stamping parts of this patent are provided by Fuzhou Shengyu Door Control Intelligent Technology Co., Ltd. There are twelve types of parts and ten types of defects. Five types of parts have two defects each, so there are a total of twenty-two types of parts in the data set. For images taken in a real simulated assembly line, a green background is used to simulate a conveyor belt. The acquisition steps are as follows: (1) Five defect-free parts and their two corresponding defective parts are placed together in a good and bad pattern, and a total of 5 pictures are taken. The placement and angle are random and there is no stacking. Therefore, 10 pictures can be collected for each part, a total of 50 pictures for the five parts, and 10 pictures are taken for each of the remaining seven defect-free parts, for a total of 120 pictures; (2) The twelve defect-free parts are placed together, the placement and angle are random and there is no stacking, and a total of 90 pictures are collected. Parts with ten types of defects were photographed in the same way, and 90 pictures were collected, for a total of 180 pictures; (3) Twelve defect-free parts and ten defective parts were placed together, with random positions and angles and no stacking, and 10 pictures were collected for each, for a total of 100 pictures; (4) All parts were photographed with some stacking and no stacking, with random positions and angles, for a total of 100 pictures. In summary, the part recognition and defect detection dataset has a total of 500 pictures, and after rotation and brightness changes, a total of 1500 pictures are taken. The training set, validation set, and test set are divided according to the ratio of 8:1:1, that is, 1200:150:150, to construct a part dataset for training DETR and YOLOv11.

[0047] In addition, the above dataset was randomly defocused and motion blurred to generate a total of 1500 blurred-clear image pairs, and the blurred dataset 1 was constructed. The training set, validation set, and test set were divided according to the ratio of 8:1:1, that is, 1200:150:150, for training DeblurGan-v2. And 500 blurred data of different degrees were added to the original clear dataset to construct the blurred dataset 2. The training set, validation set, and test set were divided according to the ratio of 8:1:1, that is, 1600:200:200 for training AlexNet. Twenty-two parts such as Figure 1 shown.

[0048] 2. Part image deblurring based on AlexNet and DeblurGan-v2

[0049] First, the artificially blurred part image needs to be deblurred. In this patent, the AlexNet network is first used to classify the blurred and clear images, and then the blurred images are sent to the DeblurGan-v2 network for deblurring. The network structures of the two networks are as follows: Figure 2As shown in the figure, the AlexNet network structure consists of 5 convolutional layers and 3 fully connected layers. In addition, AlexNet also has fully connected layers FC7 and FC8. FC6 and FC7 contain Dropout, which discards some neurons to prevent overfitting. The activation function in FC8 is Sigmoid, which outputs the sample classification probability. DeblurGan-v2 uses a blurred image as input, generates a clear image through the generator G, and then inputs it into the discriminator D together with the real clear image for discrimination. The generator G then adjusts its own network parameters according to the result of the discriminator D, and generates data closer to the real in the next round of training. Its generator consists of three parts: MobileNetv2 as the backbone network for feature extraction, a feature fusion module for top-down feature maps, and an atrous spatial convolutional pooling pyramid (ASPP) that enriches contextual information. The discriminator consists of a network layer composed of multiple convolution and activation functions.

[0050] 3. Overall network model for real-time part recognition and defect detection

[0051] Figure 3 The figure shows the overall network structure of this patent. In order to effectively identify the part category and locate the defective area, it is necessary to first perform blur judgment on the captured image, and then input the blurred image into the subsequent model after clarifying it; then. Build a part recognition network based on DETR. For a picture, DETR predicts a certain number of objects, and directly outputs a set of predictions based on the global information of the object, that is, Transformer decodes the prediction results of all objects in the picture at one time. This parallel feature makes DETR very efficient; finally, the classified part image is input into YOLOv11 for defect location. As an improvement of YOLOv8, the original C2f is replaced by C3K2, and C2PSA is added to the last layer of the backbone network. The two Con in the classification detection head in the head decoupling head are changed to DWConv, which reduces the amount of calculation.

[0052] 3.1 DETR-based Part Recognition Network Model

[0053] DETR consists of a CNN backbone, Transformer encoder and decoder, and a detection head consisting of a fully connected layer. It regards the target detection task as a set prediction problem and proposes a new paradigm for target detection, SetPrediction.

[0054] (1) CNN Backbone

[0055] The CNN in Backbone is the feature extraction part, which is used to reduce the size of the input part image. The convolutional neural network used in this patent is ResNet-50. Usually, the output channel of Backbone is 2048, and the image height and width are reduced to 1 / 32 of the original.

[0056] (2) Transformer Encoder

[0057] After Backbone, the number of channels of the output feature map is 2048, and 1×1 convolution is used for dimensionality reduction before it is input into the Transformer Encoder. At this time, the self-attention mechanism performs a global analysis on the feature map. Since the last feature map is more friendly to large objects, performing Self-Attention will help the network better extract the connection between parts in different positions.

[0058] (3) Transformer Decoder

[0059] There are two inputs in the decoder, one is object queries, which is a learnable embedding, and the other is the global features transmitted from the image end. The two inputs are repeatedly self-attentioned and output to the fully connected layer. In each decoder, a self-attention operation of the object query is performed first. The purpose of the self-attention operation of the object query is to remove redundant frames, because after communicating with each other, the possible frames for each query are known, and then duplicate frames are avoided.

[0060] (4)Prediction Heads

[0061] The detection head, which is composed of a fully connected layer, makes two predictions after receiving the features passed in by the Decoder. One is the prediction of the object category, and the other is the prediction of the box. There are four values ​​in total, namely the center point of the box and the height and width of the box.

[0062] 3.2 Part defect detection and positioning network model based on YOLOv11

[0063] In this patent, the YOLOv11 network is used for part defect detection. The network structure of YOLOv11 is as follows Figure 4As shown in the figure, it consists of three parts: Backbone, Neck, and Head. Backbone is mainly responsible for extracting high-level semantic features from the input image. It usually contains multiple convolutional layers and pooling layers, and YOLOv11 replaces the previous C2f with C3K2 to extract features more effectively. Neck is mainly used to further integrate and adjust the features extracted by Backbone. It upsamples the deep features and splices them with the shallow features, integrating features at different levels to provide more global and local feature information, which helps to capture the details and contextual relationships of the target and improve the accuracy of segmentation. Head is the last layer of the model, and its structure will vary depending on different tasks. At the same time, YOLOv11 proposes a deep separable convolution DWConv, which can reduce the number of parameters and calculations.

[0064] Specific embodiments and descriptions

[0065] 1. Experimental environment + data set construction introduction

[0066] Since a large amount of image processing is required, GPU is used to increase the training speed during training. Therefore, this patent uses RTX3060Ti (16GB) and Intel (R) 12490fCPU processors, uses Python programming language, and uses Cuda version 12.6. The deep learning framework used is Pytorch2.3.1.

[0067] The stamping parts of this patent are provided by Fuzhou Shengyu Door Control Intelligent Technology Co., Ltd. There are twelve types of parts and ten types of defects, five of which have two defects each, so the data set has a total of twenty-two types of parts. It is an image taken in a real simulated assembly line, with a green background to simulate a conveyor belt. The acquisition steps are as described above. Finally, a total of 1,500 parts data sets were constructed for training DETR and YOLOv11, 1,500 pairs of blur data sets 1 for training DeblurGan-v2, and blur data sets 2 for training AlexNet. Some data sets are as follows: Figure 6 shown.

[0068] 2. Deblurring network training process diagram and indicators

[0069] When training the AlexNet network, the training process is as follows: Figure 7 shown.

[0070] When training the DeblurGan-v2 network, the model metrics are shown in Table 1.

[0071] Table 1 Model objective evaluation indicators

[0072]

[0073] After training the deblurring network, the clear image is as follows Figure 8 shown.

[0074] 3.DETR recognition network training process diagram and indicators

[0075] During the training of the DETR part recognition network, the loss curve of the model is as follows: Fig. 9 shown.

[0076] The part recognition experiment was conducted using the ablation experiment method, comparing the method used in this patent with the method that does not use the defuzzification network. Fig.10 Shown are the part recognition results of the two methods in the case of heavy stacking.

[0077] The recognition rates of the two methods for part recognition are shown in Table 2. As can be seen from the table, the patented part recognition method has a higher accuracy rate for multi-part recognition.

[0078] Table 2 Recognition rate of DETR target detection model in different situations

[0079]

[0080]

[0081] 4. YOLOv11 defect location network training process diagram and indicators

[0082] During the training of the YOLOv11 defect localization network, the loss curve of the model is as follows: Fig.11 shown.

[0083] For the evaluation of defect location detection and segmentation quality, four indicators, TP, TN, FP, and FN, are first introduced. TP (True Positive) refers to pixels that are positive samples and predicted as positive samples, TN (True Negative) refers to pixels that are negative samples and predicted as negative samples, FP (False Positive) refers to pixels that are negative samples predicted as positive samples, and FN (False Negative) refers to pixels that are positive samples predicted as negative samples.

[0084] IOU is the intersection-over-union ratio, which is the ratio of the overlapped area between the prediction and the real area to the total area between the prediction and the real area. The best segmentation result is 1. The three are defined as follows:

[0085]

[0086] The defect detection experiment was conducted using the ablation experiment method, comparing the method used in this patent with the method that does not use the deblurring network. The average indicators of the two methods for defect location detection in blurred part images are shown in Table 2. The algorithm proposed in this patent has high segmentation accuracy under different degrees of blur, indicating that the algorithm is highly robust. Table 2 Model defect location detection indicator IOU

[0087]

[0088]

Claims

1. A stacked stamping parts recognition and defect detection method with visual deblurring, characterized in that: The following steps are involved: Step 1: First, collect the categories and defect images of various parts, and annotate the images according to the categories of parts and deformation, scratches, missing corners, and fracture defects to build a sample data set; Step 2: Artificially defocus and motion blur the captured image to simulate industrial machine vibration; Step 3: Build a deblurring network based on AlexNet and DeblurGan-v2 to deblur the artificially blurred image and obtain a clear image; Step 4: Input the clear image into DETR to achieve accurate classification of part categories; Step 5: Send the classified part images into the YOLOv11 network for defect location and segmentation to obtain the final recognition and defect detection results.

2. The stacked stamping parts recognition and defect detection method with visual deblurring according to claim 1 is characterized in that: In step 1, the steps for collecting the categories and defect images of various parts are as follows: (1) Five defect-free parts and their two corresponding defective parts were placed together, one good and one bad, and photographed. A total of five photos were taken. The placement and angle were random and there was no stacking. Ten photos were collected for each part, a total of 50 photos for the five parts. Ten photos were taken for each of the remaining seven defect-free parts, for a total of 120 photos. (2) Twelve defect-free parts were placed together, with random positions and angles and no stacking, for a total of 90 images; parts with ten types of defects were photographed in the same manner, for a total of 90 images, for a total of 180 images; (3) Twelve defect-free parts and ten defective parts are placed together, with random placement and angle and no stacking. Ten images are collected for each type, for a total of 100 images; (4) All parts were photographed with or without some stacking, with random positions and angles, for a total of 100 photos; The part recognition and defect detection data set consists of 500 images, which total 1500 images after rotation and brightness change. The training set, validation set, and test set are divided in a ratio of 8:1:1, that is, 1200:150:150, to construct a part data set for training DETR and YOLOv11.

3. The stacked stamping parts recognition and defect detection method with visual deblurring according to claim 1, characterized in that: Step 3 specifically includes: firstly, using the AlexNet network to perform binary classification of blurred and clear images, and then sending the blurred images to the DeblurGan-v2 network for deblurring; the AlexNet network structure is composed of 5 convolutional layers and 3 fully connected layers, and the AlexNet network structure also has fully connected layers FC7 and FC8, FC6 and FC7 contain Dropout, discarding some neurons to prevent overfitting, and the activation function in FC8 is Sigmoid, outputting the sample classification probability; DeblurGan-v2 takes the blurred image as input, generates a clear image through the generator G, and then inputs it into the discriminator D together with the real clear image for discrimination; the generator G then adjusts its own network parameters according to the result of the discrimination of the discriminator D, and generates data closer to the real in the next round of training; its generator is composed of the backbone network MobileNetv2 as feature extraction, a feature fusion module for top-down feature map, and a hollow spatial convolution pooling pyramid ASPP that enriches contextual information, and the discriminator is composed of a network layer composed of multiple convolution and activation functions.

4. The stacked stamping parts recognition and defect detection method with visual deblurring according to claim 1, characterized in that: Step 4 specifically includes: first, it is necessary to make a blur judgment on the captured image, and then input the blurred image into the subsequent model after clarifying it; then, build a part recognition network based on DETR. For a picture, DETR predicts a certain number of objects, and directly outputs the predicted set according to the global information of the object, that is, Transformer decodes the prediction results of all objects in the picture at one time; finally, input the classified part image into YOLOv11 for defect location. As an improvement of YOLOv8, the original C2f is replaced by C3K2, and C2PSA is added to the last layer of the backbone network. The two Con in the classification detection head in the head decoupling head are changed to DWConv to reduce the amount of calculation.

5. The stacked stamping parts recognition and defect detection method with visual deblurring according to claim 4, characterized in that: The DETR-based part recognition network specifically includes: (1) CNN Backbone CNN in Backbone is the feature extraction part; convolutional neural network ResNet-50; (2) Transformer Encoder After Backbone, the number of channels of the output feature map is 2048, and 1×1 convolution is used for dimensionality reduction before inputting into the Transformer Encoder. (3) Transformer Decoder The decoder has two inputs, one is the object queries, and the other is the global features transmitted from the image end. The two inputs are repeatedly self-attentioned and then output to the fully connected layer. In each decoder, the object query self-attention operation is performed first. (4) Prediction Heads The detection head, which is composed of a fully connected layer, makes two predictions after receiving the features passed in by the Decoder. One is the prediction of the object category, and the other is the prediction of the box. There are four values ​​in total, namely the center point of the box and the height and width of the box.

6. The stacked stamping parts recognition and defect detection method with visual deblurring according to claim 4, characterized in that: The network structure of YOLOv11 consists of three parts: Backbone, Neck, and Head. Backbone is responsible for extracting high-level semantic features from the input image, including multiple convolutional layers and pooling layers, and YOLOv11 replaces the previous C2f with C3K2. Neck is used to further integrate and adjust the features extracted by Backbone. Neck upsamples the deep features and concatenates them with the shallow features, fusing features at different levels to provide more global and local feature information. Head is the last layer of the model.

Images