Material bundle surface defect identification system and method based on improved YOLOv10

Abstract

The invention relates to a material bundle surface defect identification system based on improved YOLOv10. The image acquisition system is used for capturing and processing a high-resolution material beam image on the surface of the material beam; the data analysis system is used for carrying out identification and defect classification processing on the preprocessed high-resolution material beam image; the dynamic model improvement system adopts an improved YOLOv10 network model; and the data visualization presentation system is used for displaying statistical information of the high-resolution material beam image in real time. The invention also discloses a material bundle surface defect identification method based on the improved YOLOv10. According to the method, the fusion capability of multi-scale features is enhanced, so that the YOLOv10 can still effectively perform defect detection and classification in a complex noise interference environment, and particularly, the detection of tiny defects on the surface of a material bundle is more accurate; the method has stronger adaptability and robustness, and can improve the detection precision and reduce the omission ratio.

Description

Technical Field

[0001] The present invention relates to the technical field of defect detection and classification of material beam industrial product surfaces, and in particular to a material beam surface defect recognition system and method based on improved YOLOv10. Background Art

[0002] At the end of the material bundle industrial production chain, ensuring that product quality meets standards is a crucial link. The existence of defective products not only threatens the stability of downstream equipment operation, but also directly reduces the yield rate and overall competitiveness of manufacturers. At present, although traditional manual inspection methods play a role as gatekeepers to a certain extent, their inherent limitations are becoming increasingly prominent.

[0003] First, manual inspection relies heavily on human resources, which is not only costly but also inefficient. Long-term and intensive inspection work can easily lead to worker fatigue, which in turn increases the risk of missed inspections and poses a hidden danger to product quality.

[0004] Secondly, in the face of the increasing diversity of industrial products, workers need to be retrained every time a new product is introduced, which is time-consuming and laborious. In addition, the automation level of manual inspection is low, making it difficult to adapt to the rapidly changing production needs. At the same time, its accuracy is easily affected by human factors, making it difficult to ensure stable high standards.

[0005] In short, the traditional manual inspection method has become inadequate in the industrial production of material bundles, and a more efficient, accurate and automated solution is urgently needed to replace it in order to improve production efficiency, reduce labor costs, and ensure that each product meets the best quality standards. Summary of the invention

[0006] In order to solve the problems of slow detection speed, insufficient small target detection accuracy, repeated detection and false detection in the existing material beam defect detection technology, the primary purpose of the present invention is to provide a material beam surface defect recognition system and method based on improved YOLOv10, which can effectively perform defect detection and classification, improve detection accuracy and reduce missed detection rate.

[0007] To achieve the above object, the present invention adopts the following technical solution: a material beam surface defect recognition system based on improved YOLOv10, comprising:

[0008] An image acquisition system is used to capture a high-resolution material beam image of the material beam surface. After the high-resolution material beam image is digitized and noise removed, a pre-processed high-resolution material beam image is obtained and sent to a data analysis system;

[0009] The data analysis system is used for real-time quality inspection and control on the production line. It identifies and classifies defects in the pre-processed high-resolution material bundle image, obtains the analyzed high-resolution material bundle image and sends it to the dynamic model improvement system and data visualization presentation system;

[0010] The dynamic model improvement system uses the improved YOLOv10 network model to process the analyzed high-resolution material beam images by automatically training and adjusting the model, and feeds the updated model information back to the data analysis system;

[0011] The data visualization system is used to display the statistical information of the high-resolution material beam image in real time, and after integration, it is provided to the operator for viewing and use.

[0012] The image acquisition system comprises:

[0013] Industrial cameras, used to capture clear, high-resolution industrial images, ensuring stability and accuracy in high-speed, harsh environments;

[0014] A conveying device is used to maintain the continuous fluidity of the material beam data and ensure that the image acquisition system can capture images in different positions or states; the conveying device includes an acrylic plate, two limiters for limiting the movement range of the material beam, two tractors for dragging the material beam, and two storage boxes for storing the material beam, the acrylic plate is laid on the entire workbench surface, two limiters are installed at both ends of the workbench surface, a tractor is arranged next to each limiter, and a storage box is arranged next to each tractor;

[0015] Light sources are used to enhance image contrast, reduce light interference, and improve detail clarity, ensuring that industrial cameras can capture high-quality images;

[0016] Image acquisition module, used with industrial cameras and light sources to collect high-resolution, clear image data and perform preprocessing;

[0017] The data analysis system comprises:

[0018] The signal input module is used to receive the digital image signal and external control signal from the image acquisition module, ensure the synchronization of the image and external events, and perform signal preprocessing and monitoring to provide high-quality and synchronous data input for the data analysis system;

[0019] Defect detection module, which analyzes image data to automatically identify and classify defects in the production process;

[0020] The defect handling module is used to respond and handle detected defects in real time, and record the defects and handling results.

[0021] The dynamic model improvement system adopts a model update module, which is used to continuously train and optimize the improved YOLOv10 network model using the latest data set;

[0022] The data visualization presentation system comprises:

[0023] Display module, used to display defect detection images and result statistics in real time, and supports user interaction;

[0024] The host computer is used to coordinate the operation of each module, process high-resolution material beam image data, and provide a user interface for operators to monitor, adjust and intervene.

[0025] Another object of the present invention is to provide a recognition method of a material beam surface defect recognition system based on an improved YOLOv10, the method comprising the following steps in sequence:

[0026] (1) Image preprocessing: First, the input material bundle image I orig Resize the input material bundle image I orig The size is adjusted to H×W, and then normalized to map the pixel values ​​to the interval [0, 1]. At the same time, spatial-channel decoupled downsampling is used to optimize the computational efficiency to obtain the standardized material beam image I preprocessed ;

[0027] (2) Feature extraction: An adaptive deep network is introduced to replace the backbone network of the YOLOv10 network model, and multi-scale feature maps are extracted as F extracted ; After extracting the multi-scale feature map F extracted Finally, the separation and enhancement attention module is added to the neck network of the YOLOv10 network model to enhance the features and obtain the enhanced feature map F enhanced ;

[0028] (3) Feature fusion: The convolution module in the adaptive deep network is replaced by a convolution adaptive channel network module, which combines the shallow high-resolution features with the deep low-resolution semantic information to output a fused feature map F with rich information. fused ;

[0029] (4) Detection head detection: The rejection loss is introduced in the detection head. The detection head is used to fuse the feature map F. fused Perform target detection on it and get the prediction result set D predictions ;

[0030] (5) Output features: For the prediction result set D predictions Perform post-processing and output the detected target R final .

[0031] The step (1) specifically refers to: for the input material beam image I orig , after preprocessing, the standardized material bundle image I is obtained preprocessed :

[0032] I preprocessed =Norm(SpatialChannelDecoupling(Resize(I orig , H, W)))

[0033] Where Norm represents image normalization, SpatialChannelDecoupling represents spatial-channel decoupling downsampling, Resize represents image adjustment, H represents image height, and W represents image width.

[0034] The step (2) specifically refers to: in the feature extraction stage, an adaptive deep network is introduced to replace the backbone network of the YOLOv10 network model, and when the standardized material bundle image I is input preprocessed When , assuming that the weight of the self-attention mechanism is A, the extracted multi-scale feature map is F extracted :

[0035] F extracted =Attention(A)·ADNet(I preprocessed )

[0036] In the formula, Attention represents the attention mechanism, and ADNet represents the adaptive deep network, which is used to enhance the feature representation capability;

[0037] After extracting the multi-scale feature map F extracted Finally, the separation and enhancement attention module is added to the neck network of the YOLOv10 network model to enhance the features and obtain the enhanced feature map F enhanced :

[0038] F enhanced =SEAM(F extracted )

[0039] SEAM stands for separation and enhancement of attention module; the separation and enhancement of attention module includes a channel separation attention module and a spatial enhancement module. The channel separation module selectively processes the features of different channels, and the spatial enhancement module enhances the sensitivity of the model to local spatial features. The separation and enhancement of attention module decouples different spatial and channel information, so that the model has stronger representation capabilities when dealing with small targets and occlusion conditions.

[0040] The step (3) specifically refers to: replacing the convolution module in the adaptive deep network with a convolution adaptive channel network module to enhance the ability of cross-stage feature fusion; combining the enhanced feature maps F from different resolution levels enhanced Fusion is performed to ensure the detection capability of objects of different sizes; through fusion, the shallow high-resolution features are combined with the deep low-resolution semantic information to output a fusion feature map F with rich information. fused :

[0041] F fused =FusionModule(KACNConv(F enhanced ))

[0042] In the formula, FusionModule represents the multi-scale feature fusion module, and KACN represents the convolutional adaptive channel network module.

[0043] The step (4) specifically refers to: introducing a rejection loss in the detection head, and the detection head fusion feature map F fused The detection head uses a one-to-many matching strategy and a one-to-one matching strategy. The one-to-many matching strategy is suitable for the regression and classification of the pre-selected box, and the one-to-one matching strategy uses the Hungarian algorithm to ensure unique matching, and obtains the prediction result set D predictions :

[0044] D predictions ={(b′ i , p′ i , c′ i )|i=1,2,...,N}

[0045] In the formula, b′ i is the pre-selected box without post-processing, p′ i is the category label without post-processing, c′ i is the confidence score without post-processing, i is the index, and N is the number of detected targets;

[0046] At this point, the total loss Total Loss is expressed as:

[0047] Total Loss=Localization Loss+Classification Loss+Repulsion Loss

[0048] In the formula, Localization Loss is the location loss, which is used to measure the difference between the pre-selected box and the real box; Classification Loss is the classification loss, which is used to measure the accuracy of category prediction; Repulsion Loss is the rejection loss, which is used to enhance the detection ability of small objects and dense objects.

[0049] In step (5), the post-processing includes pre-selection box filtering, pre-selection box merging, cluster analysis and adaptive thresholding; first, by setting a confidence threshold, the prediction results corresponding to the pre-selection boxes below the confidence threshold are discarded, and then the similar pre-selection boxes are merged. If the overlapping area of ​​two pre-selection boxes exceeds the set limit value, the pre-selection box with the lower prediction probability is discarded; cluster analysis is performed; finally, the threshold is dynamically adjusted according to the actual detection environment to adapt to the detection requirements of different targets, thereby obtaining the detection target R final ; The detection target R final Including the prediction box b i , category label p i and the confidence score c i :

[0050] R final =PostProcess(D predictions )=(b i , p i , c i )|i=1,2,...,N

[0051] Where PostProcess represents post-processing, i is the index, and N is the number of detected targets.

[0052] It can be seen from the above technical scheme that the beneficial effects of the present invention are: first, by introducing the convolutional adaptive channel network module, the fusion capability of multi-scale features is enhanced, so that YOLOv10 can still effectively detect and classify defects in a complex noise interference environment, especially the detection of tiny defects on the surface of the material beam is more accurate; second, the use of an adaptive deep network significantly improves the feature extraction capability of YOLOv10, and when facing defects of different sizes and shapes, it has stronger adaptability and robustness, which can improve detection accuracy and reduce missed detection rate; third, by introducing the separation and enhancement attention module, the detection capability of YOLOv10 for small targets under occlusion is enhanced, so that targets can still be effectively separated in complex scenes and accurately detected; fourth, the addition of exclusion loss can effectively reduce confusion between overlapping targets and improve the matching accuracy of the detection head, thereby improving the accuracy of YOLOv10 for The recognition and positioning accuracy of the target, these improvements effectively improve the accuracy and efficiency of material beam defect detection, enhance the adaptability of the model in different application scenarios, and provide a more efficient and low-cost solution for the detection of surface defects of industrial product material beams; Fifth, the present invention has the functions of defect detection, positioning and classification, and obtains the product surface conditions in real time through the image acquisition module, and performs corresponding defect detection and classification tasks; Sixth, with the help of feature fusion technology, it can effectively detect defects of different sizes, and quickly adapt to new defect types or product types through transfer learning; Seventh, the present invention can be trained on the COCO dataset, and does not require a large number of datasets for defect detection, only a small amount of product defect data is needed to obtain excellent results; Eighth, the host computer software is used to display the current product status in real time and save necessary data, achieving the advantages of high automation, high accuracy and low cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a system structure block diagram of the present invention;

[0054] Figure 2 The present invention is a flow chart of the method. DETAILED DESCRIPTION

[0055] like Figure 1 As shown, a material beam surface defect recognition system based on improved YOLOv10 includes:

[0056] An image acquisition system is used to capture a high-resolution material beam image of the material beam surface. After the high-resolution material beam image is digitized and noise removed, a pre-processed high-resolution material beam image is obtained and sent to a data analysis system;

[0057] The data analysis system is used for real-time quality inspection and control on the production line. It identifies and classifies defects in the pre-processed high-resolution material bundle image, obtains the analyzed high-resolution material bundle image and sends it to the dynamic model improvement system and data visualization presentation system;

[0058] The dynamic model improvement system uses the improved YOLOv10 network model to process the analyzed high-resolution material beam images by automatically training and adjusting the model, and feeds the updated model information back to the data analysis system;

[0059] The data visualization system is used to display the statistical information of the high-resolution material beam image in real time, and after integration, it is provided to the operator for viewing and use.

[0060] The image acquisition system comprises:

[0061] Industrial cameras, used to capture clear, high-resolution industrial images, ensuring stability and accuracy in high-speed, harsh environments;

[0062] A conveying device is used to maintain the continuous fluidity of the material beam data and ensure that the image acquisition system can capture images in different positions or states; the conveying device includes an acrylic plate, two limiters for limiting the movement range of the material beam, two tractors for dragging the material beam, and two storage boxes for storing the material beam. The acrylic plate is laid on the entire workbench surface, and two limiters are installed at both ends of the workbench surface. A tractor is arranged next to each limiter, and a storage box is arranged next to each tractor; the material beam is placed on the storage box, and the material beam passes through the tractor and the limiter in turn to reach the acrylic plate, and the industrial camera takes a picture of the material beam on the acrylic plate;

[0063] Light sources are used to enhance image contrast, reduce light interference, and improve detail clarity, ensuring that industrial cameras can capture high-quality images. Through carefully designed light sources, shadows, reflections, and other light interference can be minimized to ensure image consistency and clarity. The effective combination of industrial cameras and light sources significantly improves image quality and ensures the accuracy of subsequent image processing and analysis, thereby improving the accuracy and reliability of detection.

[0064] Image acquisition module, used with industrial cameras and light sources to collect high-resolution, clear image data and perform preprocessing;

[0065] The data analysis system comprises:

[0066] The signal input module is used to receive digital image signals and external control signals from the image acquisition module, ensure the synchronization of images with external events, and perform signal preprocessing and monitoring to provide high-quality, synchronous data input for the data analysis system; the module also integrates signals from multiple sensors, supports multi-camera or multi-sensor collaboration, and coordinates the processing order of multi-channel signals. At the same time, the signal input module monitors the status of the input signal to ensure the normal operation of the system, and promptly alarms or takes corresponding measures when abnormal signals are found. In order to exchange data with other devices, the signal input module provides standard communication interfaces such as USB, GigE Vision, Camera Link, etc., and supports software and hardware trigger modes. Finally, the acquired image information is transmitted to the defect detection module.

[0067] The defect detection module is used to analyze image data and automatically identify and classify defects in the production process, such as distortion, structural inconsistency, surface contamination, etc. The module first uses feature extraction, edge detection and other technologies to analyze the image to find potential defect areas, and then uses the improved YOLOv10 network model to judge and mark these abnormal features. Subsequently, the defects will be classified and evaluated. The defect detection module feeds back the detection results to the defect handling module to trigger the corresponding operation, and records the results for quality control analysis and model training. Due to the high efficiency of the production line, the module also has the ability to process large amounts of image data in real time, ensuring that the product can instantly complete defect identification when passing through the inspection equipment, thereby improving product quality, reducing scrap rates, and improving production efficiency.

[0068] The defect handling module is used to respond to and handle detected defects in real time, and record the defects and processing results. When the defect detection module identifies defects in the product, the defect handling module decides how to handle these defects. If the material bundle is too twisted to cause the width to be unqualified or there are too many defects, a voice prompt will be given and mechanical equipment will be controlled to remove unqualified products, mark products to be repaired, or adjust production parameters to prevent further defects. The module also records defects and their processing results, generates reports for quality management and production optimization, and feeds back the processing results to the host computer or other system components for necessary adjustments. The defect handling module works in conjunction with other system modules to ensure the continuity and consistency of the entire production process, thereby improving production efficiency, reducing scrap rates, and ensuring that the final product quality meets standards.

[0069] The dynamic model improvement system adopts a model update module, which is used to continuously train and optimize the improved YOLOv10 network model using the latest data set; by continuously collecting new samples and feedback information from the defect detection module, the module enables the detection system to have the ability to continuously learn and adapt to new defect types or product types. When facing different application scenarios and product changes, the model update module can adjust the model parameters in time to ensure the flexibility and robustness of the system. In addition, it supports both online and offline update modes, and can update the model without affecting production, or perform batch updates during production downtime. The module also has automated processing capabilities and can automatically complete the process of data preparation, model training, verification and deployment, reducing the need for manual intervention. At the same time, the model update module is also responsible for monitoring model performance and deciding whether to update the model based on the evaluation results to ensure that the system is always in the best working state.

[0070] The data visualization presentation system comprises:

[0071] Display module, used to display defect detection images and result statistics in real time, and supports user interaction, allowing users to view detailed information or adjust settings;

[0072] The host computer is used to coordinate the operation of each module, process high-resolution material beam image data, and provide a user interface for operators to monitor, adjust and intervene. As the control center of the entire equipment, the host computer is responsible for coordinating the work of each module to ensure the smooth progress of the entire detection process; it receives data from the image acquisition module, uses the improved YOLOv10 network model to perform image analysis, and processes the detection results; based on the detection results, the host computer can make corresponding decisions and provide a friendly user interface so that operators can easily monitor the system status, adjust parameter settings or execute specific commands; at the same time, it can also store detection results and related data and generate report files. The close integration of the display module and the host computer ensures the smooth transmission of information, improves the overall efficiency and accuracy of the system, and enables operators to monitor the detection process in real time and intervene as needed.

[0073] like Figure 2 As shown, the method comprises the following steps in order:

[0074] (1) Image preprocessing: First, the input material bundle image I orig Resize the input material bundle image I orig The size is adjusted to H×W, and then normalized to map the pixel values ​​to the interval [0, 1]. At the same time, spatial-channel decoupled downsampling is used to optimize the computational efficiency to obtain the standardized material beam image I preprocessed ;

[0075] (2) Feature extraction: An adaptive deep network is introduced to replace the backbone network of the YOL0v10 network model, and multi-scale feature maps are extracted as F extracted ; After extracting the multi-scale feature map F extracted Finally, the separation and enhancement attention module is added to the neck network of the YOLOv10 network model to enhance the features and obtain the enhanced feature map F enhanced ;

[0076] (3) Feature fusion: The convolution module in the adaptive deep network is replaced by a convolution adaptive channel network module, which combines the shallow high-resolution features with the deep low-resolution semantic information to output a fused feature map F with rich information. fused ;

[0077] (4) Detection head detection: The rejection loss is introduced in the detection head. The detection head is used to fuse the feature map F. fused Perform target detection on it and get the prediction result set D predictions ;

[0078] (5) Output features: For the prediction result set D predictions Perform post-processing and output the detected target R final .

[0079] The step (1) specifically refers to: for the input material beam image I orig , after preprocessing, the standardized material bundle image I is obtained preprocessed :

[0080] I preprocessed =Norm(SpatialChannelDecoupling(Resize(I orig , H, W)))

[0081] Where Norm represents image normalization, SpatialChannelDecoupling represents spatial-channel decoupling downsampling, Resize represents image adjustment, H represents image height, and W represents image width.

[0082] The step (2) specifically refers to: in the feature extraction stage, an adaptive deep network is introduced to replace the backbone network of the YOLOv10 network model, and when the standardized material bundle image I is input preprocessed When , assuming that the weight of the self-attention mechanism is A, the extracted multi-scale feature map is F extracted :

[0083] F extracted =Attention(A)·ADNet(I preprocessed )

[0084] In the formula, Attention represents the attention mechanism, and ADNet represents the adaptive deep network, which is used to enhance the feature representation capability;

[0085] After extracting the multi-scale feature map F extracted Finally, the separation and enhancement attention module is added to the neck network of the YOLOv10 network model to enhance the features and obtain the enhanced feature map F enhanced :

[0086] F enhanced =SEAM(F extracted )

[0087] SEAM stands for separation and enhancement of attention module; the separation and enhancement of attention module includes a channel separation attention module and a spatial enhancement module. The channel separation module selectively processes the features of different channels, and the spatial enhancement module enhances the sensitivity of the model to local spatial features. The separation and enhancement of attention module decouples different spatial and channel information, so that the model has stronger representation capabilities when dealing with small targets and occlusion conditions.

[0088] The step (3) specifically refers to: replacing the convolution module in the adaptive deep network with a convolution adaptive channel network module to enhance the ability of cross-stage feature fusion; combining the enhanced feature maps F from different resolution levels enhanced Fusion is performed to ensure the detection capability of objects of different sizes; through fusion, the shallow high-resolution features are combined with the deep low-resolution semantic information to output a fusion feature map F with rich information. fused :

[0089] F fused =FusionModule(KACNConv(F enhanced ))

[0090] In the formula, FusionModule represents the multi-scale feature fusion module, and KACN represents the convolutional adaptive channel network module.

[0091] The step (4) specifically refers to: introducing a rejection loss in the detection head, and the detection head fusion feature map F fused The detection head uses a one-to-many matching strategy and a one-to-one matching strategy. The one-to-many matching strategy is suitable for the regression and classification of the pre-selected box, and the one-to-one matching strategy uses the Hungarian algorithm to ensure unique matching, and obtains the prediction result set D predictions :

[0092] D predictions ={(b′ i , p′ i , c′ i)|i=1,2,...,N}

[0093] In the formula, b′ i is the pre-selected box without post-processing, p′ i is the category label without post-processing, c′ i is the confidence score without post-processing, i is the index, and N is the number of detected targets;

[0094] At this point, the total loss Total Loss is expressed as:

[0095] Total Loss=Localization Loss+Classification Loss+Repulsion Loss

[0096] In the formula, Localization Loss is the location loss, which is used to measure the difference between the pre-selected box and the real box; Classification Loss is the classification loss, which is used to measure the accuracy of category prediction; Repulsion Loss is the rejection loss, which is used to enhance the detection ability of small objects and dense objects.

[0097] An improved YOLOv10 network model is obtained by replacing the YOLOv10 network model with an adaptive deep network, adding separation and enhancement attention modules, replacing the convolutional adaptive channel network module, and introducing rejection loss.

[0098] In step (5), the post-processing includes pre-selection box filtering, pre-selection box merging, cluster analysis and adaptive thresholding; first, by setting a confidence threshold, the prediction results corresponding to the pre-selection boxes below the confidence threshold are discarded, and then the similar pre-selection boxes are merged. If the overlapping area of ​​two pre-selection boxes exceeds the set limit value, the pre-selection box with the lower prediction probability is discarded to ensure the uniqueness and accuracy of the detection results; cluster analysis is performed to help understand the spatial relationship between different targets; finally, the threshold is dynamically adjusted according to the actual detection environment to adapt to the detection requirements of different targets, thereby obtaining the detection target R final ; The detection target R final Including the prediction box b i , category label p i and the confidence score c i :

[0099] R final =PostProcess(D predictions )=(b i , p i , ci )|i=1,2,...,N

[0100] Where PostProcess represents post-processing, i is the index, and N is the number of detected targets.

[0101] In summary, the present invention enhances the fusion capability of multi-scale features by introducing a convolutional adaptive channel network module, so that YOLOv10 can still effectively perform defect detection and classification in a complex noise interference environment, especially the detection of tiny defects on the surface of material beams is more accurate; the adaptive deep network is adopted to significantly improve the feature extraction capability of YOLOv10, and it has stronger adaptability and robustness when facing defects of different sizes and shapes, which can improve the detection accuracy and reduce the missed detection rate; by introducing the separation and enhancement attention module, the detection capability of YOLOv10 for small targets under occlusion is enhanced, so that targets can still be effectively separated in complex scenes and accurately detected; adding the rejection loss can effectively reduce the confusion between overlapping targets and improve the matching accuracy of the detection head, thereby improving the recognition and positioning accuracy of YOLOv10 for targets. These improvements effectively improve the accuracy and efficiency of material beam defect detection, enhance the adaptability of the model in different application scenarios, and provide a more efficient and low-cost solution for the detection of surface defects of industrial product material beams.

Claims

1. A material beam surface defect recognition system based on improved YOLOv10, characterized by: include; An image acquisition system is used to capture a high-resolution material beam image of the material beam surface. After the high-resolution material beam image is digitized and noise removed, a pre-processed high-resolution material beam image is obtained and sent to a data analysis system; The data analysis system is used for real-time quality inspection and control on the production line. It identifies and classifies defects in the pre-processed high-resolution material bundle image, obtains the analyzed high-resolution material bundle image and sends it to the dynamic model improvement system and data visualization presentation system; The dynamic model improvement system uses the improved YOLOv10 network model to process the analyzed high-resolution material beam images by automatically training and adjusting the model, and feeds the updated model information back to the data analysis system; The data visualization system is used to display the statistical information of the high-resolution material beam image in real time, and after integration, it is provided to the operator for viewing and use.

2. The material beam surface defect recognition system based on improved YOLOv10 according to claim 1, characterized in that: The image acquisition system comprises: Industrial cameras, used to capture clear, high-resolution industrial images, ensuring stability and accuracy in high-speed, harsh environments; A conveying device is used to maintain the continuous fluidity of the material beam data and ensure that the image acquisition system can capture images in different positions or states; the conveying device includes an acrylic plate, two limiters for limiting the movement range of the material beam, two tractors for dragging the material beam, and two storage boxes for storing the material beam, the acrylic plate is laid on the entire workbench surface, two limiters are installed at both ends of the workbench surface, a tractor is arranged next to each limiter, and a storage box is arranged next to each tractor; Light sources are used to enhance image contrast, reduce light interference, and improve detail clarity, ensuring that industrial cameras can capture high-quality images; Image acquisition module, used with industrial cameras and light sources to collect high-resolution, clear image data and perform preprocessing; The data analysis system comprises: The signal input module is used to receive the digital image signal and external control signal from the image acquisition module, ensure the synchronization of the image and external events, and perform signal preprocessing and monitoring to provide high-quality and synchronous data input for the data analysis system; Defect detection module, which analyzes image data to automatically identify and classify defects in the production process; The defect handling module is used to respond and handle detected defects in real time, and record the defects and handling results.

3. The material beam surface defect recognition system based on improved YOLOv10 according to claim 1, characterized in that: The dynamic model improvement system adopts a model update module, which is used to continuously train and optimize the improved YOLOv10 network model using the latest data set; The data visualization presentation system comprises: Display module, used to display defect detection images and result statistics in real time, and supports user interaction; The host computer is used to coordinate the operation of each module, process high-resolution material beam image data, and provide a user interface for operators to monitor, adjust and intervene.

4. The recognition method of the material beam surface defect recognition system based on the improved YOLOv10 according to any one of claims 1 to 3, characterized in that: The method comprises the following steps in order: (1) Image preprocessing: First, the input material bundle image I orig Resize the input material bundle image I orig The size is adjusted to H×W, and then normalized to map the pixel values ​​to the interval [0,1]. At the same time, spatial-channel decoupled downsampling is used to optimize the computational efficiency to obtain the standardized material beam image I preprocessed ; (2) Feature extraction: An adaptive deep network is introduced to replace the backbone network of the YOLOv10 network model, and multi-scale feature maps are extracted as F extracted ; After extracting the multi-scale feature map F extracted Finally, the separation and enhancement attention module is added to the neck network of the YOLOv10 network model to enhance the features and obtain the enhanced feature map F enhanced ; (3) Feature fusion: The convolution module in the adaptive deep network is replaced by a convolution adaptive channel network module, which combines the shallow high-resolution features with the deep low-resolution semantic information to output a fused feature map F with rich information. fused ; (4) Detection head detection: The rejection loss is introduced in the detection head. The detection head is used to fuse the feature map F. fused Perform target detection on it and get the prediction result set D predictions ; (5) Output features: For the prediction result set D predictions Perform post-processing and output the detected target R final .

5. The identification method according to claim 4, characterized in that: The step (1) specifically refers to: for the input material beam image I orig , after preprocessing, the standardized material bundle image I is obtained preprocessed : I preprocessed =Norm(SpatialChannelDecoupling(Resize(I orig ,H,W))) Where Norm represents image normalization, SpatialChannelDecoupling represents spatial-channel decoupling downsampling, Resize represents image adjustment, H represents image height, and W represents image width.

6. The identification method according to claim 4, characterized in that: The step (2) specifically refers to: in the feature extraction stage, an adaptive deep network is introduced to replace the backbone network of the YOLOv10 network model, and when the standardized material bundle image I is input preprocessed When , assuming that the weight of the self-attention mechanism is A, the extracted multi-scale feature map is F extracted : F extracted =Attention(A)·ADNet(I preprocessed ) In the formula, Attention represents the attention mechanism, and ADNet represents the adaptive deep network, which is used to enhance the feature representation capability; After extracting the multi-scale feature map F extracted Finally, the separation and enhancement attention module is added to the neck network of the YOLOv10 network model to enhance the features and obtain the enhanced feature map F enhanced : F enhanced =SEAM(F extracted ) SEAM stands for Separation and Enhancement Attention Module; The separation and enhancement attention module includes a channel separation attention module and a spatial enhancement module. The channel separation module selectively processes the features of different channels, and the spatial enhancement module enhances the sensitivity of the model to local spatial features. The separation and enhancement attention module decouples different spatial and channel information, making the model more capable of representing small objects and occlusions.

7. The identification method according to claim 4, characterized in that: The step (3) specifically refers to: replacing the convolution module in the adaptive deep network with a convolution adaptive channel network module to enhance the ability of cross-stage feature fusion; combining the enhanced feature maps F from different resolution levels enhanced Fusion is performed to ensure the detection capability of objects of different sizes; through fusion, the shallow high-resolution features are combined with the deep low-resolution semantic information to output a fusion feature map F with rich information. fused : F fused =FusionModule(KACNConv(F enhanced )) In the formula, FusionModule represents the multi-scale feature fusion module, and KACN represents the convolutional adaptive channel network module.

8. The identification method according to claim 4, characterized in that: The step (4) specifically refers to: introducing a rejection loss in the detection head, and the detection head fusion feature map F fused For target detection, the detection head adopts one-to-many matching strategy and one-to-one matching strategy; The one-to-many matching strategy is suitable for the regression and classification of the pre-selected box, while the one-to-one matching strategy uses the Hungarian algorithm to ensure a unique match and obtain the prediction result set D predictions : D predictions ={(b′ i ,p′ i ,c′ i )∣i=1,2,…,N} Where b' i is the pre-selected box without post-processing, p' i is the category label without post-processing, c' i is the confidence score without post-processing, i is the index, and N is the number of detected targets; At this point, the total loss Total Loss is expressed as: Total Loss=Localization Loss+Classification Loss+Repulsion Loss In the formula, Localization Loss is the location loss, which is used to measure the difference between the pre-selected box and the real box; Classification Loss is the classification loss, which is used to measure the accuracy of category prediction; Repulsion Loss is the rejection loss, which is used to enhance the detection ability of small objects and dense objects.

9. The identification method according to claim 4, characterized in that: In step (5), the post-processing includes pre-selection box filtering, pre-selection box merging, cluster analysis and adaptive thresholding; first, by setting a confidence threshold, the prediction results corresponding to the pre-selection boxes below the confidence threshold are discarded, and then the similar pre-selection boxes are merged. If the overlapping area of ​​two pre-selection boxes exceeds the set limit value, the pre-selection box with the lower prediction probability is discarded; Perform cluster analysis; finally, dynamically adjust the threshold according to the actual detection environment to adapt to the detection requirements of different targets, so as to obtain the detection target R final ; The detection target R final Including the prediction box b i , category label p i and the confidence score c i : R final =PostProcess(D predictions )=(b i ,p i ,c i )∣i=1,2,…,N Where PostProcess represents post-processing, i is the index, and N is the number of detected targets.

Images