Laser cut sheet defect image detection method and device
Through the deep learning model SSD-GCN, combined with SSD and GCN, the accuracy and leakage detection rate problems of traditional technology when detecting laser cutting defects are solved, efficient and accurate defect identification is achieved, and the quality and efficiency of battery cell production are improved.
Patent Information
- Application Number
- CN202411682843.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-22
- Publication Date
- 2025-03-11
AI Technical Summary
When detecting laser cutting defects, traditional image recognition technology relies on manual design feature extraction methods, resulting in low recognition accuracy, difficulty in adapting to different scenarios and defect types, and high missed detection rate.
The deep learning model SSD-GCN is adopted, combining the SSD object detection model and graph convolutional network (GCN), and the SSD-GCN model is trained to identify defect categories by preprocessing and normalizing the image data.
It improves the rapid and accurate identification of laser cutting defects, reduces production costs, enhances the quality and efficiency of battery cell production, and reduces safety hazards caused by production defects.
Smart Images
Figure CN119600354A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the field of image detection technology, and in particular to a method and device for detecting defects in laser cutting images. Background Art
[0002] Laser cutting is one of the key links in the production process of battery cells. The quality of the slices directly affects the performance and safety of the final battery products. Some defects often occur in laser cutting, such as dark marks, breakage, splicing, sticking and other problems. These production defects will not only reduce the performance and quality of the battery cells, but may also cause safety hazards such as short circuit and leakage of the battery, which will have a serious impact on the stability and reliability of battery production.
[0003] Traditional methods usually use manual visual inspection or simple image processing technology to detect production defects. Although manual visual inspection is simple and intuitive, it is inefficient, easily affected by subjective factors, and cannot meet the needs of high-speed production lines. Methods based on simple image processing technology can often only detect some surface defects, and cannot accurately detect deep and tiny defects.
[0004] The specific problems of existing technologies are as follows:
[0005] Problem 1: Traditional image recognition technology relies on manually designed feature extraction methods, which may not be sufficient to fully describe production defects, resulting in low recognition accuracy. In addition, traditional image recognition technology uses fixed feature extraction and classification strategies, which are difficult to adapt to different scenarios and production defect types, and lack flexibility and robustness.
[0006] Problem 2: Traditional image recognition algorithms usually rely on manually designed feature extraction methods, which may not fully reveal the characteristics of complex defects, resulting in a high missed detection rate. Laser slicing defects have various forms, and defects of various sizes, shapes, and positions may appear. Traditional image recognition algorithms often find it difficult to cover all possible defect forms, especially some special forms of defects are easily missed. In addition, film defects may be obscured or confused by other components, making it difficult for traditional image recognition algorithms to accurately identify defect areas, thereby increasing the missed detection rate. Summary of the invention
[0007] In response to the above problems, the present disclosure provides a laser cutting defect image detection method and device, which are used to realize laser cutting defect image detection by introducing a deep learning model (SSD-GCN) to improve production efficiency and product quality.
[0008] In a first aspect, a method for detecting laser cutting defect images is provided, the method comprising:
[0009] Preprocessing and normalizing the laser slice image data to obtain processed data;
[0010] The SSD-GCN model is established and trained through historical processing data to obtain the trained SSD-GCN model; the SSD-GCN model includes an SSD module and a GCN module, in which the SSD target detection model uses ResNet as the backbone network.
[0011] The processed data collected in real time are input into the SSD-GCN model in sequence, and the SSD-GCN model is used to identify the defect category of the processed data and finally output the prediction results.
[0012] Specifically, data preprocessing refers to a series of processing performed on image data to make it suitable for training and prediction of deep learning models. Preprocessing includes image resizing, data enhancement, grayscale processing and denoising. Image resizing is to adjust the images in the data set to the same size to ensure data consistency and ensure the stability and performance of model training; data enhancement increases the diversity of training data by randomly transforming training images, enhances the generalization ability of the model, and reduces overfitting; grayscale processing is to convert color images into grayscale images, reduce the dimension of data, and reduce computational complexity; denoising is to use filters and other methods to remove noise in the image, improve image quality and model recognition accuracy. Normalization is to scale the pixel values of image data to a specific range, which is set to between 0 and 1 in this disclosure. This scaling method can speed up the training of the model and contribute to the stability of the model.
[0013] The core idea of SSD is to use feature maps of different scales to detect different defect targets, generate candidate areas through a single forward propagation, and implement classification and regression. This method greatly improves the detection speed. In this disclosure, the SSD module uses ResNet as a feature extraction network, and generates feature maps of different scales by adding several additional convolutional layers for detecting targets of different sizes. On each feature map, the added convolutional layer is used to predict the category and bounding box of the target. The SSD module defines multiple default boxes on each feature map. These default boxes have different scales and aspect ratios. By regressing the offsets of these default boxes, the SSD module can predict the bounding box of the target.
[0014] GCN is a deep learning model for processing graph structured data. It can effectively learn graph data by propagating information between the nodes and their neighbors. A graph consists of nodes and edges, which are used to represent objects and their relationships. The convolution operation in GCN is to perform a weighted summation of the features of each node and its neighboring nodes, thereby updating the node features. The present invention captures the high-order neighbor information of a node by stacking multiple GCNs.
[0015] In the defect detection task, there may be certain relationships between the defect labels, and these relationships can be represented as a graph structure. The present disclosure regards the target detection box as a node and the spatial relationship between the detection boxes as an edge. The GCN module extracts these graph structure information to enhance the performance of the defect detection model. Specifically, the GCN module performs a weighted fusion of the extracted high-level relationship features and the convolutional features extracted by the SSD module. During the training process, the parameters of the SSD and GCN parts are optimized at the same time, so that the model can better learn the target features and their relationships in the image.
[0016] Furthermore, Multibox_Loss is used as the loss function for the target detection task to train the detector to predict the category and location of the target. This loss function calculates the loss by comparing the similarity between the model's predicted box and the actual labeled box, while considering the classification error of the target category and the regression error of the bounding box to comprehensively evaluate the accuracy of the detection result.
[0017] The loss function L(x,c,l,g) includes: classification loss L conf (x,c) and the localization loss L loc (x,l,g), the formula is as follows:
[0018] L(x,c,l,g)=L conf (x,c)+αL loc (x,l,g)
[0019] Among them, x represents the variable matrix, which is used to determine which predicted boxes are positive samples; c represents the predicted category probability distribution, which is a vector, each element of which represents the probability that the predicted box belongs to a certain category. When calculating the category classification loss, this probability distribution is used to compare with the true category; l represents the predicted bounding box parameters, including the center coordinates and size of the bounding box. These parameters are predicted by regression methods; g represents the true bounding box parameters, including the center coordinates and size of the actual object; α represents the loss balance coefficient, which is usually set to 1, but can be adjusted according to the specific application scenario to adapt to different task requirements.
[0020] Furthermore, the positioning loss is used to measure the accuracy of the model in locating the target position. Based on the smooth L1 loss function, the difference between the position of each anchor box predicted by the model and the actual annotation box is compared;
[0021] The position difference between the predicted box and the target box is calculated through the positioning loss to guide the model to learn to accurately predict the position of the target. The formula is as follows:
[0022]
[0023] in, represents an indicator variable, which takes the value of 1 when the i-th predicted box matches the j-th real box, otherwise it takes the value of 0; l i Represents the predicted bounding box parameters (such as center coordinates and width and height); g i Represents the actual bounding box parameters; smooth L1 (x) represents the smooth L1 loss function, which is defined as:
[0024]
[0025] Furthermore, the classification loss is used to measure the classification accuracy of the model for the target category. Based on the cross entropy loss function, the difference between the probability distribution of the category to which each anchor box belongs predicted by the model and the target category in the actual annotation box is compared;
[0026] For each anchor box, the classification loss calculates the gap between the predicted category probability and the actual target category to guide the model to learn the ability to correctly classify targets. The formula is as follows:
[0027]
[0028] Among them, c i Represents the predicted category probability distribution; Pos represents a positive sample, that is, a predicted box that matches the real box; Neg represents a negative sample, that is, a predicted box that does not match the real box.
[0029] Furthermore, the error back propagation algorithm BP is used to train the SSD-GCN model to find the optimal weight parameters that minimize the loss function.
[0030] Furthermore, Adam is used as the optimizer during the training process. First, a part of the data is randomly selected from the training data set, and the gradient of each weight parameter is calculated through the loss function. Then, based on these gradients as clues, the corresponding parameters are fine-tuned along the direction of the gradient. Through continuous iteration, the optimal weight parameters are finally obtained.
[0031] Furthermore, BN batch normalization is used to accelerate the training process of the neural network and improve the robustness and generalization ability of the model. Each feature in each batch of training data is standardized so that its mean is 0 and its variance is 1, making the network more stable.
[0032] In a second aspect, a laser cutting defect image detection device includes: a data processing unit, a model building unit and a recognition unit;
[0033] A data processing unit, used for preprocessing and normalizing the laser slice image data to obtain processed data;
[0034] The model building unit is used to build the SSD-GCN model and train it through historical processing data to obtain the trained SSD-GCN model; the SSD-GCN model includes an SSD module and a GCN module, wherein the SSD target detection model uses ResNet as the backbone network.
[0035] The recognition unit is used to input the real-time collected processing data into the SSD-GCN model in sequence, identify the defect category of the processed data through the SSD-GCN model, and finally output the prediction result.
[0036] In a third aspect, an electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus;
[0037] a memory storing a computer program;
[0038] The processor is used to implement the above-mentioned laser cutting defect image detection method when executing the computer program stored in the memory.
[0039] In a fourth aspect, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned laser cutting defect image detection method is implemented.
[0040] The present disclosure has at least the following beneficial effects:
[0041] The present invention discloses research and development of an efficient and accurate laser cutting defect image detection system. The system can realize rapid and accurate identification of manufacturing defects, improve the quality and efficiency of battery cell production, and reduce safety hazards caused by manufacturing defects.
[0042] Compared with traditional methods, the present invention aims to reduce production costs while ensuring accurate identification of defect information, provide companies with economical solutions, improve production efficiency and optimize product quality, and have significant improvements over traditional methods. Therefore, the invention requires not only to be innovative in technology, but also to provide an economically feasible solution, bringing substantial progress to the field of battery cell manufacturing.
[0043] The SSD-GCN model proposed in this disclosure can fully utilize both spatiotemporal features and sequence dependencies, thereby more comprehensively understanding the evolution of laser film defect image detection. The depth of the entire model enables the system to abstract the features of the input data layer by layer, thereby better capturing the abstract and complex relationships in the defect image recognition task. By integrating the SSD and GCN networks, a more comprehensive and accurate technical solution for detecting film defect images is provided, which has an important driving effect on battery production and application fields.
[0044] Other features and advantages of the present disclosure will be described in the following description, and partly become apparent from the description, or be understood by implementing the present disclosure. The purpose and other advantages of the present disclosure can be realized and obtained by the structures indicated in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0046] Figure 1 This is a flow chart of the detection method according to the embodiment of the present disclosure;
[0047] Figure 2 This is a schematic diagram of the SSD model structure disclosed in the present invention;
[0048] Figure 3 This is a schematic diagram of the GCN model structure disclosed in this disclosure;
[0049] Figure 4 This is a schematic diagram of the public training process;
[0050] Figure 5 This is a schematic diagram of the structure of the detection device according to the embodiment of the present disclosure;
[0051] Figure 6 This is a schematic diagram of the training process of the public model. DETAILED DESCRIPTION
[0052] In order to make the purpose, technical solution and advantages of the embodiments of the present disclosure clearer, the technical solution in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.
[0053] The SSD-GCN model proposed in the present disclosure is a hybrid neural network structure, which adds a graph convolution module (GCN) on the basis of the traditional SSD target detection model. SSD uses a single forward propagation to complete the target detection task, which greatly reduces the calculation time and improves the detection speed compared with the traditional two-stage method. The model effectively realizes the detection of targets at different scales by performing target detection on feature maps of multiple different scales, and simultaneously processes small targets and large targets in a single model. Features at different levels are fused through a feature pyramid network to extract richer semantic information to improve the accuracy of target detection. The model uses multiple predefined prior boxes to cover targets of different scales and aspect ratios, thereby improving the model's adaptability to target position and scale. GCN can perform effective feature extraction on feature maps of different scales, so that SSD can better capture the semantic information and spatial structure of the target, thereby improving the accuracy of target detection. GCN can model the semantic relevance between targets, so that the model can better understand the relationship between targets, thereby improving the accuracy and robustness of target detection. In addition, GCN can effectively fuse global context information, so that the model can better understand the semantic information of the entire image, thereby improving the accuracy and robustness of target detection.
[0054] like Figure 1 As shown, a method for detecting defective images of laser cutting is provided, the method comprising:
[0055] S101, preprocessing and normalizing the laser cutting image data to obtain processed data;
[0056] S102, establish an SSD-GCN model, and train it through historical processing data to obtain a trained SSD-GCN model; the SSD-GCN model includes an SSD module and a GCN module, wherein the SSD target detection model uses ResNet as the backbone network,
[0057] S103, input the processed data collected in real time into the SSD-GCN model in sequence, identify the defect category of the processed data through the SSD-GCN model, and finally output the prediction result.
[0058] The specific implementation is as follows:
[0059] The present disclosure provides a method for detecting defective images of laser cutting, which adopts SSD-GCN. The model takes laser cutting images as input data and outputs images with defect information. The model first detects defective targets in the image through the SSD network. Then the specific defect category is identified through the GCN network. Finally, the image defect information is obtained.
[0060] The SSD-GCN model consists of an SSD module and a GCN module. Specifically, the data is first preprocessed and normalized, which can make the neural network operate faster. Then, the data enters the SSD module. This disclosure uses ResNet as the backbone network of the SSD target detection model. Figure 2 As shown, the data is then passed through the GCN module to identify the defect category, such as Figure 3 As shown, the prediction result is finally output.
[0061] This paper adopts Multibox_Loss as the loss function for target detection tasks and trains the detector to predict the category and location of the target. This loss function calculates the loss by comparing the similarity between the model's predicted box and the actual labeled box, while comprehensively considering the classification error of the target category and the regression error of the bounding box to comprehensively evaluate the accuracy of the detection result. Multibox_Loss mainly consists of two parts: classification loss and positioning loss.
[0062] L(x,c,l,g)=L conf (x,c)+αL loc (x,l,g)
[0063] Among them, x represents the variable matrix, which is used to determine which predicted boxes are positive samples; c represents the predicted category probability distribution, which is a vector, each element of which represents the probability that the predicted box belongs to a certain category. When calculating the category classification loss, this probability distribution is used to compare with the true category; l represents the predicted bounding box parameters, including the center coordinates and size of the bounding box. These parameters are predicted by regression methods; g represents the true bounding box parameters, including the center coordinates and size of the actual object; α represents the loss balance coefficient, which is usually set to 1, but can be adjusted according to the specific application scenario to adapt to different task requirements.
[0064] The localization loss is used to measure the accuracy of the model's localization of the target location. It is based on a smooth L1 loss function that compares the difference between the location of each anchor box predicted by the model and the actual annotation box. Specifically, the localization loss calculates the position difference between the predicted box and the target box to guide the model to learn to accurately predict the location of the target.
[0065]
[0066] in, represents an indicator variable, which takes the value of 1 when the i-th predicted box matches the j-th real box, otherwise it takes the value of 0; l i Represents the predicted bounding box parameters (such as center coordinates and width and height); g i Represents the actual bounding box parameters; smooth L1 (x) represents the smooth L1 loss function, which is defined as:
[0067]
[0068] Classification loss is used to measure the accuracy of the model in classifying the target category. It is based on the cross entropy loss function and compares the difference between the probability distribution of the category to which each anchor box belongs predicted by the model and the target category in the actual annotation box. Specifically, for each anchor box, the classification loss calculates the gap between the predicted category probability and the actual target category to guide the model's ability to learn to correctly classify targets.
[0069]
[0070] Among them, c i Represents the predicted category probability distribution; Pos represents a positive sample, that is, a predicted box that matches the real box; Neg represents a negative sample, that is, a predicted box that does not match the real box.
[0071] like Figure 4 As shown, the present disclosure uses the error back propagation algorithm (BP) to train the model and find the optimal weight parameters that minimize the loss function. Adam is used as the optimizer during the training process. First, a portion of data is randomly selected from the training data set, and the gradient of each weight parameter is calculated through the loss function. Then, based on these gradients as clues, the corresponding parameters are fine-tuned along the direction of the gradient. Through continuous iteration, the optimal weight parameters are finally obtained.
[0072] In order to solve the problems of exponential growth of parameters, overfitting and slow training speed that may occur in model training, this disclosure introduces some training techniques. These training techniques are designed to avoid the above problems as much as possible and improve the performance and training efficiency of the model.
[0073] Batch Normalization (BN): Batch Normalization (BN) is a regularization method used in deep learning to accelerate the training process of neural networks and improve the robustness and generalization ability of the model. Its basic idea is to standardize each feature in each batch of training data so that its mean is 0 and its variance is 1.
[0074] The main goal of BN is to solve the problem of internal covariate shift in the training process of deep neural networks. Internal covariate shift refers to the change in the distribution of input data at each layer, which makes it difficult for the network to learn. BN eliminates the change in the distribution of input data by standardizing each feature on each batch of data, making the network more stable.
[0075] Adam: This is a commonly used optimization algorithm for training deep neural networks. It combines the advantages of the momentum optimization algorithm and the adaptive learning rate algorithm, and can efficiently update the parameters of the network and accelerate the convergence speed of the model.
[0076] The basic principle of Adam is to consider the first-order moment estimate and the second-order moment estimate of the gradient at the same time during the update process of each parameter. Specifically, Adam uses an exponentially weighted moving average to estimate the first-order moment and the second-order moment of the gradient, and to correct the deviation of these estimates.
[0077] like Figure 5 As shown, a laser cutting defect image detection device includes: a data processing unit 501, a model building unit 502 and a recognition unit 503;
[0078] The data processing unit 501 is used to pre-process and normalize the laser slice image data to obtain processed data;
[0079] The model building unit 502 is used to build an SSD-GCN model and train it through historical processing data to obtain a trained SSD-GCN model; the SSD-GCN model includes an SSD module and a GCN module, wherein the SSD target detection model uses ResNet as the backbone network.
[0080] The identification unit 503 is used to input the processed data collected in real time into the SSD-GCN model in sequence, identify the defect category of the processed data through the SSD-GCN model, and finally output the prediction result.
[0081] In specific implementation, the SSD-GCN model is a hybrid neural network structure, which is characterized by combining the SSD target detection module with the graph convolution module to improve the accuracy and robustness of target detection. SSD is responsible for the target detection task. This is a single forward propagation target detection algorithm that directly outputs the prediction results of the target category and position by applying a series of convolution and pooling operations on the image feature map. SSD can effectively capture targets of different scales and aspect ratios and has a high detection speed. GCN is a neural network based on a graph structure that can effectively process graph data and learn the association information between nodes. The introduction of this module enhances the association between labels and improves the accuracy of target recognition. In the SSD-GCN model, GCN is responsible for modeling the semantic association between targets during target detection. The GCN model can better understand the relationship between targets and improve the accuracy and robustness of target detection.
[0082] Compared with traditional object detection methods, SSD-GCN can learn more representative and abstract feature representations, better capture the semantic relationship and spatial structure between objects, and effectively integrate global context information, so that the model can better understand the semantic information of the entire image, thereby improving the accuracy and robustness of object detection.
[0083] like Figure 6As shown, the present disclosure provides an electronic device, including a processor 601, a communication interface 602, a memory 603 and a communication bus 604, wherein the processor 601, the communication interface 602 and the memory 603 communicate with each other through the communication bus 604;
[0084] Memory 603, storing computer programs;
[0085] The processor 601 is used to implement the above method when executing the computer program stored in the memory 603.
[0086] The present disclosure provides a computer-readable storage medium storing a computer program, wherein the computer program implements the above method when executed by a processor.
[0087] The computer-readable storage medium may be included in the device / apparatus described in the above embodiment; or it may exist independently without being assembled into the device / apparatus. The above computer-readable storage medium carries one or more programs, and when the above one or more programs are executed, the method according to the embodiment of the present disclosure is implemented.
[0088] According to an embodiment of the present disclosure, a computer-readable storage medium may be a non-volatile computer-readable storage medium, for example, may include but is not limited to: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, apparatus, or device.
[0089] Although the present disclosure has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.
Claims
1. A laser cutting defect image detection method, characterized in that: The method comprises: Preprocessing and normalizing the laser slice image data to obtain processed data; Establish an SSD-GCN model and train it through historical processing data to obtain a trained SSD-GCN model; the SSD-GCN model includes an SSD module and a GCN module; The processed data collected in real time are input into the SSD-GCN model in sequence, and the SSD-GCN model is used to identify the defect category of the processed data and finally output the prediction results.
2. A laser cutting defect image detection method according to claim 1, characterized in that: Multibox_Loss is used as the loss function for target detection tasks, and the detector is trained to predict the category and location of the target. This loss function calculates the loss by comparing the similarity between the model's predicted box and the actual labeled box, while considering the classification error of the target category and the regression error of the bounding box to comprehensively evaluate the accuracy of the detection result. The loss function L(x,c,l,g) includes: classification loss L conf (x,c) and the localization loss L loc (x,l,g), the formula is as follows: L(x,c,l,g)=Lconf(x,c)+αLLoc(x,l,g) Among them, x represents the variable matrix, which is used to determine which prediction boxes are positive samples; c represents the predicted category probability distribution, c is a vector, and each element in c represents the probability that the predicted box belongs to a certain category; when calculating the category classification loss, the probability distribution c is used to compare with the true category; l represents the predicted bounding box parameters, including the center coordinates and size of the bounding box; g represents the true bounding box parameters, including the center coordinates and size of the actual object; α represents the loss balance coefficient.
3. A laser cutting defect image detection method according to claim 2, characterized in that: The positioning loss is used to measure the accuracy of the model in positioning the target. Based on the smooth L1 loss function, the difference between the position of each anchor box predicted by the model and the actual annotation box is compared; The position difference between the predicted box and the target box is calculated through the positioning loss to guide the model to learn to accurately predict the position of the target. The formula is as follows: in, represents an indicator variable, which takes the value of 1 when the i-th predicted box matches the j-th real box, otherwise it takes the value of 0; l i represents the predicted bounding box parameters; g i Represents the actual bounding box parameters; smooth L1 (x) represents the smooth L1 loss function, which is defined as:
4. The method for detecting laser cutting defect images according to claim 2, characterized in that: The classification loss is used to measure the classification accuracy of the model for the target category. Based on the cross entropy loss function, the difference between the probability distribution of the category to which each anchor box belongs predicted by the model and the target category in the actual annotation box is compared; For each anchor box, the classification loss calculates the gap between the predicted category probability and the actual target category to guide the model to learn the ability to correctly classify targets. The formula is as follows: Among them, c i Represents the predicted category probability distribution; Pos represents a positive sample, that is, a predicted box that matches the real box; Neg represents a negative sample, that is, a predicted box that does not match the real box.
5. The method for detecting laser cutting defect images according to claim 1, characterized in that: The SSD-GCN model is trained using the error back propagation algorithm BP to find the optimal weight parameters that minimize the loss function.
6. A laser cutting defect image detection method according to claim 5, characterized in that: Adam is used as the optimizer during the training process. First, a part of the data is randomly selected from the training data set, and the gradient of each weight parameter is calculated through the loss function. Then, based on these gradients as clues, the corresponding parameters are fine-tuned along the direction of the gradient. Through continuous iteration, the optimal weight parameters are finally obtained.
7. The method for detecting laser cutting defect images according to claim 1, characterized in that: Batch normalization is used to speed up the training process of the neural network and improve the robustness and generalization ability of the model. Each feature in each batch of training data is standardized to have a mean of 0 and a variance of 1.
8. A laser cutting defect image detection device, characterized in that: include: Data processing unit, model building unit and recognition unit; A data processing unit, used for preprocessing and normalizing the laser slice image data to obtain processed data; A model building unit, used to build an SSD-GCN model and train it through historical processing data to obtain a trained SSD-GCN model; The SSD-GCN model includes the SSD module and the GCN module. The SSD target detection model uses ResNet as the backbone network. The recognition unit is used to input the real-time collected processing data into the SSD-GCN model in sequence, identify the defect category of the processed data through the SSD-GCN model, and finally output the prediction result.
9. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; a memory storing a computer program; The processor is used to implement a laser cutting defect image detection method according to any one of claims 1 to 7 when executing a computer program stored in a memory.
10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, a laser cutting defect image detection method according to any one of claims 1 to 7 is implemented.