A deep learning method for plate defect detection based on dynamic receptive field and adaptive fusion

Abstract

The application discloses a kind of based on dynamic receptive field and adaptive feature fusion plate defect detection deep learning method, to improve the detection precision and robustness of complex background and small scale defect under industrial environment.The proposed method is based on YOLOv10, dynamic receptive field module and adaptive feature fusion module are introduced in its architecture.Dynamic receptive field module uses local features and context information to dynamically adjust the size of convolution receptive field, to enhance the perception ability of different scale target;Adaptive feature fusion module realizes the weighted fusion of multi-layer features through multi-layer perception mechanism, improves feature expression ability and target recognition performance.The experimental results show that the method of the application is superior to existing mainstream algorithms in detection accuracy and real-time performance, has good engineering adaptability and industrial application prospect, and is suitable for furniture manufacturing, building material processing and other plate detection scenarios.

Description

Technical Field

[0001] This research belongs to the field of computer vision and target detection technology, especially the research on defect detection methods for sheet metal. Background Technology

[0002] With the rapid development of industrial automation and intelligent manufacturing, sheet metal, as a basic component, is widely used in various fields such as construction, furniture, and shipbuilding. Its surface quality directly affects product performance and production safety. Traditional sheet metal defect detection relies heavily on manual visual inspection or traditional image processing methods, which are not only inefficient but also easily affected by factors such as operator experience and lighting conditions, resulting in high rates of missed and false detections. Especially when faced with complex backgrounds, changing lighting, or minute defects (such as cracks, bubbles, and scratches), existing methods generally suffer from insufficient recognition accuracy and poor real-time performance, making it difficult to meet the demands of modern industry for high-precision and high-efficiency detection.

[0003] In recent years, deep learning technology has demonstrated powerful feature extraction and representation capabilities in the field of object detection, becoming an important research direction in industrial visual inspection. The YOLO series models are widely used in practical scenarios due to their fast detection speed and high accuracy; however, they still have limitations in detecting complex backgrounds and small objects. Therefore, there is an urgent need to propose a defect detection method with stronger perception and feature fusion capabilities to adapt to the detection needs of multi-scale and multi-morphological defects and improve the system's stability and generalization ability under complex working conditions. This patent proposes an improved YOLOv10 model that combines a dynamic receptive field with an adaptive feature fusion mechanism to achieve efficient and accurate detection of surface defects in sheet materials. Summary of the Invention

[0004] YOLOv10, the latest version in the YOLO series, represents a significant improvement in object detection compared to previous versions, particularly in terms of accuracy and real-time performance. YOLOv10 employs several new technologies, such as a more efficient network structure, stronger feature extraction capabilities, and a more precise object localization mechanism, enabling it to perform well in various complex scenarios. Its core idea remains the same: transforming the object detection problem into a regression problem, using convolutional neural networks for full-image object detection and classification, achieving high accuracy while maintaining high speed. Therefore, YOLOv10 is widely used in various real-time detection tasks, especially for tasks like board defect detection, where both accuracy and speed are crucial, and YOLOv10's advantages are particularly prominent.

[0005] In this study, based on the YOLOv10 architecture, two improvements are proposed to enhance the accuracy and robustness of sheet metal defect detection. The first improvement is the introduction of a dynamic receptive field module, which dynamically adjusts the size of the receptive field based on local features and contextual information of the image. Sheet metal defects often present as small local areas, and the scale and shape of the defects vary considerably. Traditional fixed receptive fields may not be able to effectively capture these subtle features. By adjusting the dynamic receptive field, the network can flexibly select an appropriate receptive field size based on the characteristics of each region, thereby improving the detection accuracy of small defects and the detection capability in complex backgrounds.

[0006] The second improvement is the adaptive feature fusion module, which aims to combine multi-level feature information to enhance the network's adaptability to defects at different scales. The diversity of board defects requires the model to extract useful information from features at different levels; low-level features capture details, while high-level features provide global semantics. Through adaptive feature fusion, the network can dynamically select appropriate feature combinations according to the needs of the current task, thereby improving the model's robustness to multi-scale defects and complex backgrounds. This module, by weighted fusion of feature maps from different levels, preserves local details while effectively utilizing global contextual information, enhancing detection performance. Its overall model architecture is as follows: Figure 1 As shown.

[0007] Overall, the YOLOv10-based board defect detection algorithm, through the introduction of these two modules, significantly improves the detection accuracy of small and multi-scale defects while maintaining real-time performance. These improvements enable the algorithm to more accurately detect various defects on the board surface and reduce false positives and false negatives that may occur with traditional YOLO models in complex backgrounds.

[0008] In object detection tasks, especially in applications like board defect detection, defects in images often exhibit varying scales, shapes, and background noise, posing a challenge to traditional convolutional neural networks (CNNs). Traditional convolutional layers use fixed-size kernels. While this structure has been successful in many tasks, when processing complex images with varying scales and rich background information, kernels with fixed receptive fields may fail to capture crucial information, leading to inaccurate detection of small defects or objects with complex backgrounds. To address this issue, this paper proposes a Dynamic Receptive Field (DRF) module as an effective improvement.

[0009] The core idea of ​​the dynamic receptive field module is to dynamically adjust the size of the receptive field based on the local features and contextual information of the image, enabling the network to adapt to different scale requirements when processing different regions. In this way, the network can not only capture small-scale details more meticulously, but also retain sufficient contextual information, thereby improving detection accuracy and robustness.

[0010] First, input the image. After preprocessing, including standardization and normalization, we obtain Then, the image undergoes feature extraction through several layers of a convolutional neural network (CNN). Let the feature map obtained after the convolutional layers be... ,in This represents the index of the convolutional layer. For feature maps... a certain position in Its characteristics are represented as Traditional convolution operations use a fixed-size receptive field to process features at each location. However, this fixed receptive field cannot adapt to the needs of different regions, especially for defects with different scales.

[0011] To overcome this problem, the dynamic receptive field module calculates an adaptive receptive field size. The receptive field size considers not only local information of the current pixel but also contextual information from the global image. This contextual information is extracted through a self-attention mechanism, ensuring that the network can adaptively adjust the receptive field size to effectively capture defects at different scales. Its model framework is as follows: Figure 2 As shown.

[0012] The purpose of the self-attention mechanism is to calculate the position of each pixel. This involves understanding the relationship between the feature map and other locations globally, thereby extracting contextual information. This paper first calculates the self-attention weight matrix. It reflects the dependencies between each location in the feature map. The self-attention mechanism calculation formula is:

[0013]

[0014] in, Representing the eigenvector and Similarity between components is typically assessed using dot product similarity or cosine similarity. Self-attention weights... The position was calculated and location The correlation between them. This matrix. It can help the network capture global contextual information.

[0015] Furthermore, the contextual information obtained through the self-attention mechanism can be integrated into the features of each location through a weighted summation. Specifically, the location... Context information It can be calculated using the following formula:

[0016]

[0017] here, Indicates position The contextual information is the weighted sum of all positions in the feature map, with the weights determined by the self-attention weight matrix. The decision is made. In this way, the network can obtain global context information for each location.

[0018] With the context information, the next step is to calculate the dynamic receptive field size at each location. To achieve this, this chapter designs a receptive field calculation function that is based on position. Local features and context information To calculate the dynamic receptive field size, the formula for calculating the receptive field size is:

[0019]

[0020] in, and Representing positions respectively The norm of local features and contextual information, This is the adjustment coefficient, used to balance the influence of local features and contextual information. Through this formula, the network can dynamically adjust the size of the receptive field based on the magnitude of local features and contextual information.

[0021] When the receptive field size at each location is calculated Subsequently, during convolutional operations, the size of the receptive field will no longer be fixed, but will dynamically change based on the features and contextual information at each location. In convolutional operations, it is assumed that the location... The kernel size at the point is Then the output of the convolution operation It will be given by the following formula:

[0022]

[0023] in, These are the weights of the convolution kernel. It is the input feature map The position in the middle The eigenvalues ​​at that location. Dynamic receptive field size. The range of the convolution kernel is determined, which affects the convolution calculation for each pixel.

[0024] After processing by the dynamic receptive field module, the feature maps obtained by the network more accurately capture details in the image, especially in areas with small defects and complex backgrounds. Finally, the network passes the processed feature maps to the subsequent detection head for object detection, outputting detection results including class labels and bounding box locations. Due to the introduction of the dynamic receptive field module, the model can better adapt to defects of different scales, improving detection accuracy and robustness.

[0025] In the adaptive feature fusion module, feature maps processed by the dynamic receptive field module are first acquired. These feature maps contain receptive field sizes adaptively adjusted based on local features and global context information. To further enhance the expressive power of the features, the module aims to adaptively fuse multi-level features, enabling features from different layers to dynamically adjust their weights according to task requirements, thereby better reflecting multi-scale targets in the image. The core of the adaptive mechanism is to utilize a multilayer perceptron (MLP) to generate weights, which are used to adjust the fusion method of each feature. Its overall network architecture is as follows: Figure 3 As shown.

[0026] First, let the feature map after processing by the dynamic receptive field module be... It contains feature information calculated through dynamic receptive fields. To further enhance the expressive power of the features, the network needs to fuse feature maps from different levels. Let... This represents the low-level feature map of the input. This is the high-level feature map after processing by the dynamic receptive field module. Next, this section will perform weighted fusion of these two different levels of features so that they can fully leverage their respective advantages in the final feature representation.

[0027] During the fusion process, the low-level feature maps are first... and high-level feature maps splicing them together to form a new feature map :

[0028]

[0029] Then, this paper uses a multilayer perceptron (MLP) to generate weighting coefficients for each feature map. Specifically, the input to the MLP is the concatenated feature map. After a series of linear transformations and nonlinear activation function processing, the weights of each feature are obtained. and This is used to control the degree of fusion between low-level and high-level features. The calculation formula for MLP is:

[0030]

[0031] here, and These represent the weights of low-level and high-level features, respectively. The output of the MLP is a scalar used to adjust the importance of each layer's features in the fusion process.

[0032] Next, using these weights, the low-level features and high-level features are weighted and summed to obtain the final fused feature map. The formula for weighted fusion is:

[0033]

[0034] This weighted summation process allows the network to dynamically adjust the contribution of different levels of features to the final output features. For example, when lower-level features have a greater advantage in capturing details, the network will give higher weights to lower-level features; while when global context information is more important, higher-level features will have relatively larger weights.

[0035] Furthermore, to ensure that the fused features can be better used for subsequent object detection tasks, it is usually necessary to process the fused feature map. Additional convolution processing is performed to extract more semantic information. At this point, the output of the convolution operation... It can be represented as:

[0036]

[0037] in, These are the weights of the convolution kernel. It is the location of the fused feature map. The value at that location. After convolution processing, the feature map... It will contain richer high-level semantic information, providing support for subsequent object detection tasks.

[0038] In this way, the adaptive feature fusion module can not only combine features from different levels, but also automatically adjust the importance of features from different levels according to the needs of features at each location and the global context, thereby improving detection accuracy and robustness. This adaptive weighting mechanism enables the model to flexibly adjust the feature fusion strategy when dealing with complex backgrounds and multi-scale targets, thus enhancing the model's adaptability to different targets. Attached Figure Description

[0039] Figure 1 This is a diagram illustrating the overall algorithm architecture of an embodiment of the present invention.

[0040] Figure 2 This is a diagram illustrating the architecture of the dynamic receptive field module according to an embodiment of the present invention.

[0041] Figure 3 This is a diagram of the adaptive weight module architecture according to an embodiment of the present invention;

[0042] Figure 4 This is a visual detection diagram of an embodiment of the present invention; Detailed Implementation

[0043] In this embodiment, the acquired surface image of the board material is first preprocessed, including brightness normalization and noise suppression, to ensure the stability of the subsequent feature extraction process. Then, the image is input into the improved YOLOv10 backbone network. In the initial stage, convolutional layers are used to extract multi-scale basic features, providing input for the dynamic receptive field module and the adaptive feature fusion module. The dynamic receptive field module can adjust the receptive field size in real time based on local texture features and global context information, enabling the model to capture key region features even when facing defects of different sizes and varied shapes, thereby effectively improving the perception ability of small-scale and complex background defects.

[0044] After acquiring the dynamic receptive field features, the feature maps are fed into an adaptive feature fusion module. This module concatenates features from different levels and introduces a multilayer perceptron to generate adaptive weights, automatically adjusting the fusion ratio between low-level detail features and high-level semantic features. This ensures that the model can detect subtle scratches and cracks using low-level features, while also leveraging high-level features to obtain the overall texture distribution and semantic information of the board material, thereby enhancing the model's discriminative power in multi-scale scenes. The fused features processed by this module are then refined through convolution to obtain a more robust high-level semantic representation, providing input for the final detection head.

[0045] During the detection phase, the improved YOLOv10 detection head performs target localization and classification based on the fused features, outputting the defect category and bounding box. Experimental comparisons show that this method significantly outperforms the original YOLOv10 and its improved versions in key metrics such as Precision, Recall, and mAP, while maintaining stable performance even under complex lighting, texture variations, and noisy environments. The final detection results accurately label the surface defects of the board material, such as scratches, bubbles, and color differences, and the intuitive detection results further validate the feasibility and promotional value of the model in practical industrial applications.

[0046] Here, we first present the experimental results compared with other models, as shown in Table 1.

[0047] Model Precision Recall mAP@50 mAP@50:95 YOLOv5 0.61 0.55 51.7 33.8 SSD 0.59 0.56 50.4 32.7 YOLOv8 0.63 0.60 55.3 36.5 YOLOX 0.65 0.62 57.6 37.9 LD-YOLOv10 0.67 0.64 61.2 41.2 YOLOv10 0.68 0.66 63.9 43.7 BGF-YOLOv10 0.70 0.68 65.7 44.9 MSOAR-YOLOv10 0.66 0.63 58.8 39.4 Um-YOLOv10 0.72 0.69 67.8 46.5 This invention 0.74 0.71 70.2 48.3

[0048] Experimental results show that the algorithm presented in this paper outperforms existing mainstream models and improved versions in all detection performance metrics. Precision, Recall, mAP@50, and mAP@50:95 reach 0.74, 0.71, 70.2%, and 48.3%, respectively, significantly exceeding the original YOLOv10 and its multiple variants. This performance improvement is mainly attributed to the dynamic receptive field module and adaptive feature fusion mechanism introduced in this invention, which effectively enhances the model's ability to perceive small-scale defects and complex backgrounds, achieving higher accuracy and robustness in industrial sheet metal surface defect detection tasks. This demonstrates significant technological advancement and practical value, meeting the innovation and significance requirements for patent authorization.

[0049] Visual test results as follows Figure 4 As shown:

[0050] As can be seen from the visual detection results, the algorithm presented in this paper can stably and accurately identify multiple types of surface defects under complex backgrounds, various lighting conditions, and different board textures. It exhibits high detection box positioning accuracy, a reasonable confidence distribution, and effectively identifies defect features such as fine scratches, scabs, and color differences, while also demonstrating excellent multi-target detection capabilities. This fully verifies the adaptive advantages of the dynamic receptive field mechanism proposed in this invention in feature region perception, as well as the robustness of the adaptive feature fusion structure to defects of different scales. It provides a practical technical guarantee for efficient and accurate board defect identification in industrial applications, demonstrating significant practicality and industrial transformation potential.

[0051] The parts of this invention not described in detail are well-known to those skilled in the art.

Claims

1. A method for detecting defects in sheet metal, characterized in that: The method includes the following steps: (1) standardizing and normalizing the input board image; (2) constructing a backbone detection network structure based on YOLOv10; (3) introducing a dynamic receptive field module in the backbone network and adjusting the size of the receptive field at each position through the context self-attention mechanism; (4) sending the feature maps output by multiple network layers into the adaptive feature fusion module to perform multi-scale feature fusion; (5) outputting the defect category and localization box through the detection head to complete the identification and classification of the defect area.

2. The method for detecting defects in sheet metal according to claim 1, characterized in that: The dynamic receptive field module constructs a context relevance matrix based on a self-attention mechanism, extracts global context information at each location in the feature map, and uses it to dynamically calculate the size of the convolutional receptive field.

3. The method for detecting defects in sheet metal according to claim 1, characterized in that: The function used by the dynamic receptive field module to calculate the receptive field size is: ,in Indicates position Local features, This indicates the corresponding context information. This is the balance coefficient.

4. The method for detecting defects in sheet metal according to claim 1, characterized in that: The adaptive feature fusion module adopts a multilayer perceptron (MLP) structure. The input is the concatenation of high-level and low-level features, and the outputs are weighted coefficients. and To perform fusion computing.

5. The method for detecting defects in sheet metal according to claim 1, characterized in that: The adaptive fusion calculation method is as follows: ,in These are feature maps for the lower and higher levels, respectively. .

6. The method for detecting defects in sheet metal according to claim 1, characterized in that: Before being fed into the detection head, the fused feature map needs to undergo an additional 3×3 convolution operation and a BN normalization layer to enhance its semantic representation capability.

7. A defect detection system for implementing the method of claims 1-6, comprising an image acquisition module, a preprocessing module, an improved YOLOv10 detection module, and a result output module, wherein the features The extraction and detection module incorporates a dynamic receptive field and an adaptive fusion mechanism.

Images