Wood board surface defect detection method based on multi-view feature extraction

Abstract

The invention provides a wood board surface defect detection method based on multi-view feature extraction, and the method comprises the steps: obtaining image data collected by a plurality of industrial cameras on a wood board production line, and obtaining an original image set; extracting local texture and edge information in the original image set through a lightweight convolutional neural network to obtain local features; modeling a global topological relation in the original image set through a sparse image attention mechanism to obtain global features; performing dynamic weighted fusion on the local features and the global features through a channel attention fusion module to obtain final features; and performing optimization training on the initial model through the final features and a pre-constructed loss function to obtain a defect detection model. According to the method, the lightweight convolutional neural network and the sparse graph attention mechanism are applied in parallel, and dynamic weighted fusion is performed on the features output by the two methods, so that the precision and efficiency of defect detection are improved.

Description

Technical Field

[0001] This invention relates to the field of industrial visual inspection technology, and in particular to a method for detecting surface defects in wooden boards based on multi-view feature extraction. Background Technology

[0002] As a core material in furniture manufacturing, construction engineering, transportation, and packaging, the quality of wood panels directly determines the performance and reliability of end products. However, during the production process, wood panels are prone to various surface defects such as cracks, deformation, chipping, holes, and grooves due to factors such as insufficient stability of processing equipment and harsh working environments. These defects not only lead to a decline in the appearance quality of the wood panels and damage their market competitiveness, but may also cause uneven distribution of internal forces, significantly reducing their load-bearing capacity and service life as structural components, and even creating hidden dangers in safety-sensitive scenarios such as construction and transportation.

[0003] Currently, existing technologies suffer from three limitations: First, traditional methods such as manual inspection are inefficient (inspecting a single plank can take several minutes) and have a high false alarm rate, while image processing techniques based on edge detection or threshold segmentation are sensitive to lighting and wood grain interference, resulting in an excessively high false alarm rate. Second, mainstream deep learning models such as YOLOv11 rely on local convolutional features, making it difficult to model cross-regional associated defects (such as through-cracks), and their recall rate for small targets is insufficient. While the Transformer architecture can capture global dependencies, its computational complexity increases quadratically with image size, and inference time for high-resolution images reaches hundreds of milliseconds, failing to meet the real-time requirements of production lines. Third, graph neural networks (such as GCN-YOLO) suffer from computational redundancy due to fully connected adjacency matrices, and deep networks are prone to "oversmoothing" phenomena. Furthermore, they lack an effective fusion mechanism with local texture features, resulting in significantly insufficient robustness in complex scenarios such as uneven industrial lighting and dust interference. Therefore, it is essential to design a method for detecting surface defects in wooden boards based on multi-view feature extraction. Summary of the Invention

[0004] The purpose of this invention is to provide a method for detecting defects on the surface of wooden boards based on multi-view feature extraction. By using a lightweight convolutional neural network and a sparse graph attention mechanism in parallel, and dynamically weighting and fusing the features output by the two methods, the accuracy and efficiency of defect detection can be improved.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A method for detecting surface defects in wooden boards based on multi-view feature extraction includes the following steps:

[0007] The image data collected by multiple industrial cameras on the wood panel production line is obtained to obtain the original image set;

[0008] Local features are obtained by extracting local texture and edge information from the original image set using a lightweight convolutional neural network;

[0009] Global features are obtained by modeling the global topological relationships in the original image set through a sparse graph attention mechanism.

[0010] The channel attention fusion module dynamically weights and fuses local and global features to obtain the final features;

[0011] The initial model is optimized and trained using the final features and a pre-built loss function to obtain the defect detection model.

[0012] Optionally, a sparse graph attention mechanism is used to model the global topological relationships in the original image set to obtain global features, including:

[0013] The original image set is divided into several image patches, and the image patches are mapped to graph structure data to obtain the initial graph;

[0014] The adjacency matrix in the initial graph is sparsified using the PageRank algorithm to obtain a sparse adjacency matrix.

[0015] The key nodes and edges in the sparse adjacency matrix are dynamically filtered using the Top-KAST sparse strategy to obtain a sparse graph.

[0016] Feature extraction of sparse graphs is performed using an adaptive receptive field adjustment mechanism to obtain global features.

[0017] Optionally, a sparse graph is obtained by dynamically filtering key nodes and edges in the sparse adjacency matrix using a Top-KAST sparse strategy, including:

[0018] Calculate the attention score between nodes in the sparse adjacency matrix;

[0019] Retain the K neighbor nodes with the highest attention scores to obtain the updated sparse adjacency matrix;

[0020] The node features in the initial graph are updated based on the updated sparse adjacency matrix to obtain the sparse graph.

[0021] Optionally, the local and global features are dynamically weighted and fused using a channel attention fusion module to obtain the final features, including:

[0022] The local and global features are concatenated along the channel dimension to obtain a fused feature map;

[0023] A spatial weight map is obtained by weighting the fused feature map according to its spatial dimensions using a spatial attention mechanism.

[0024] The fused feature map is weighted according to the channel dimension by using a channel attention mechanism to obtain a channel weight map;

[0025] The spatial weight map and the channel weight map are multiplied element-wise to obtain the final feature.

[0026] Optionally, the specific steps for pre-constructing the loss function include:

[0027] The spatial deviation between the predicted bounding box and the ground truth bounding box in the initial model is modeled by the probability-weighted intersection-union loss, thus obtaining the localization loss;

[0028] The classification loss is obtained by modeling the category semantic embedding and global context in the initial model through meta-learning dynamic focusing factors;

[0029] The confidence loss is obtained by modeling the prediction confidence in the initial model using binary cross-entropy loss.

[0030] The loss function is obtained by weighted summation of the localization loss, classification loss, and confidence loss.

[0031] Optionally, the localization loss, classification loss, and confidence loss are weighted and summed to obtain a loss function, including:

[0032] Based on the current number of training steps and the preset decay parameters, calculate the dynamic weights of localization loss, classification loss, and confidence loss;

[0033] The dynamic weights are multiplied by the localization loss, classification loss, and confidence loss respectively to generate a weighted loss term;

[0034] The weighted loss terms are summed to generate the total loss function.

[0035] According to specific embodiments provided by the present invention, the following technical effects are disclosed: The present invention provides a method for detecting surface defects in wood boards based on multi-view feature extraction. This method includes: acquiring image data collected by multiple industrial cameras on a wood board production line to obtain an original image set; extracting local texture and edge information from the original image set using a lightweight convolutional neural network to obtain local features; modeling the global topological relationships in the original image set using a sparse graph attention mechanism to obtain global features; dynamically weighting and fusing the local and global features using a channel attention fusion module to obtain final features; and optimizing and training the initial model using the final features and a pre-constructed loss function to obtain a defect detection model. This method improves the accuracy and efficiency of defect detection by using a lightweight convolutional neural network and a sparse graph attention mechanism in parallel and dynamically weighting and fusing the features output by the two methods. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart of the method for detecting defects on the surface of wood panels according to the present invention;

[0038] Figure 2 This is a schematic diagram of a lightweight convolutional neural network structure according to an embodiment of the present invention;

[0039] Figure 3 This is a schematic diagram of the Top-KAST sparse strategy in an embodiment of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] like Figure 1 As shown, this invention provides a method for detecting surface defects in wooden boards based on multi-view feature extraction, comprising the following steps:

[0043] Step 100: Acquire image data from multiple industrial cameras on the wood panel production line to obtain the original image set;

[0044] Specifically, in this embodiment, raw images are acquired from the wood board production line. Two industrial cameras are deployed on one side of the line, and under suitable lighting conditions, images are captured at a high frame rate as the wood boards move along the production line. A total of 4,000 images with six types of defects are acquired. The software LabelImg is used to label the defect types and coordinates in these images to obtain the raw image set, which is then divided into a training set (3,200 images) and a test set (800 images).

[0045] Step 200: Extract local texture and edge information from the original image set using a lightweight convolutional neural network to obtain local features;

[0046] Specifically, the lightweight convolutional neural network in this embodiment adopts the YOLOv11 model. The backbone network of this model mainly consists of DSConv, the lightweight DSC3k2, and the Spatial Pyramid Pooling Fast (SPPF) module, as shown in the specific structure below. Figure 2 As shown, this model uses two DSConv modules and multiple DSC3k modules to form a lightweight DSC3k2 module. To further reduce computational requirements and simplify the model structure, the residual connections in C3k2 are also reduced. Furthermore, the SPPF module is used to improve the model's ability to handle multi-scale features and further enhance detection performance while maintaining the model's lightweight nature. The SPPF module effectively aggregates features of different scales through fast pooling operations, thereby enhancing spatial context information and enabling the model to perform well when handling targets of different sizes.

[0047] Step 300: Model the global topological relationships in the original image set using a sparse graph attention mechanism to obtain global features;

[0048] Specifically, this embodiment constructs a lightweight multi-view fusion detection model, SAGANet. This model integrates the efficient local feature extraction capabilities of YOLOv11 with the advantages of SAGAT in modeling global topology, constructing a dual-branch multi-view feature extraction framework. YOLOv11 is responsible for extracting fine-grained texture and edge features from the image, improving the model's ability to distinguish small targets and similar defects; the SAGAT branch constructs a graph structure through ViG, establishing sparse connections between defect regions and enhancing the model's ability to model complex contexts.

[0049] Furthermore, SAGANet introduces two key modules in its structure: the first is the Self-Attentional Graph Head, which combines graph neural structure with self-attention mechanism to strengthen semantic connections between key nodes, thereby improving global semantic aggregation capability; the second is the Channel Attention Fusion (CAF) module, which suppresses redundant features and enhances key response regions by dynamically weighting and fusing the output features of the two branches, thereby significantly improving detection accuracy and robustness.

[0050] More specifically, step 300 in this embodiment is implemented using the SAGAT model. This model first models the input image's graph structure based on ViG graph structure modeling, then sparsifies the graph structure using the PageRank algorithm and Top-KAST sparsity strategy, and finally extracts features from the sparsified graph structure. Specifically, the two-dimensional image is first divided into several image patches and mapped to graph structure data. After extracting local features, graph nodes are constructed, and the PageRank algorithm is introduced to sparsify the adjacency matrix, generating a sparse graph Gs. While maintaining key structural information, this significantly reduces redundant computation and improves the processing efficiency of graph neural networks on large-scale data. The expression is:

[0051] ;

[0052] Where PR(i) is the node v i The PageRank is defined as follows: N(i) is the set of neighboring nodes of node i, |V| is the total number of nodes in the graph, |N(j)| is the number of neighboring nodes of node j, and d is the damping factor, which is 0.85 in this embodiment.

[0053] It should be noted that SAGAT effectively filters and strengthens key nodes and edges by employing a sparsity strategy and an adaptive adjustment mechanism, reducing interference from irrelevant information. This mechanism not only improves computational efficiency but also enhances the accuracy of feature representation, enabling the model to adaptively adjust network depth and complexity across various tasks, thereby improving robustness, stability, and generalization ability.

[0054] like Figure 3 As shown, then according to the Top-KAST sparsity strategy, only the K neighbor nodes with the highest attention scores of each node are retained to achieve forward sparse propagation, backward sparse training, and dynamic reparameterization update of attention. This limits the computational scope of attention, thereby reducing computational complexity, while ensuring the aggregation ability of key nodes and improving the accuracy of graph representation. During forward propagation, in each training layer t, only the top K weight values ​​of the current network weight θt, sorted by attention value, are used. This embodiment assumes that the feature vector of the current node is hi. Through sparsity operation, only the top K neighbor nodes with the highest weights are considered, expressed as: , where L t Let j represent the attention weight set after Top-K sparsification, and θ represent the node. t j This represents the current network weights. The forward propagation process relies only on the weights of these selected neighboring nodes and calculates the output, expressed as: .

[0055] During backpropagation, gradient calculation only involves the top K neighbor nodes with the strongest correlation to the current node. In the gradient calculation, the superset Mt includes the top K+P important neighbor nodes, where P is an additional redundancy to improve stability, expressed as: Based on this, gradient updates are performed only on these key neighbor nodes, thereby reducing the influence of unimportant nodes. Furthermore, to maintain sparsity, after each gradient update, the top K most important nodes are selected to reconstruct the sparse mask Lt, thereby updating the network weights θt, expressed as: , where η is the learning rate and L is the loss function.

[0056] Finally, to prevent the receptive field from being too small or too large, this embodiment dynamically adjusts the convolutional layer strategy to control the output of the Top-KAST sparse strategy, thereby obtaining global features. In graph neural networks, the receptive field is determined by the depth of the convolutional layers, and the size of the receptive field directly affects the ability of each node to capture information. The formula for calculating the receptive field size R is: Where q is the average degree of the graph, i.e., the average number of edges connected to each node in the graph, |E| is the number of edges, |V| is the number of nodes, and L is the number of convolutional layers. Simultaneously, the optimal number of convolutional layers L is adjusted according to the actual structure of the output graph of the Top-KAST sparse strategy, expressed as: .

[0057] Step 400: Dynamically weight and fuse local and global features using the channel attention fusion module to obtain the final features;

[0058] Specifically, the Channel Attention Fusion (CAF) module, through spatial and channel attention mechanisms, can dynamically adjust the importance of feature dimensions, thereby achieving effective alignment and weighted fusion of multi-view features. The specific implementation process of the CAF module in this embodiment is as follows:

[0059] (1) Feature Input: The input of the CAF module is the feature X of SAGAT. SA Features X of YOLOv11 YOLO The expression is: , Where B represents the number of samples in a batch, C represents the number of channels, i.e. the dimension of the feature map, H represents the height of the feature map, and W represents the width of the feature map.

[0060] (2) Feature alignment: To ensure feature consistency, X needs to be aligned. SA Make adjustments to make it consistent with X YOLO To maintain consistency in shape for subsequent splicing and merging, the expression for the adjusted result is: .

[0061] (3) Feature splicing: With X YOLO Concatenation is performed along the channel dimension. Channel-level concatenation allows the fused features to simultaneously contain local details and global relational information, enriching the expressive power of the features. The expression is: .

[0062] (4) Spatial Attention Generation: Spatial attention maps are used to characterize the importance of features at different spatial locations. In this embodiment, the spatial attention is normalized using the Sigmoid activation function to ensure that the weights at each location are reasonable. The concatenated features are then processed through a 1×1 convolution operation to generate a spatial attention map A. HW The expression is: 1×1 convolution operations avoid the computational overhead of excessively large convolution kernels while effectively extracting important information in the spatial dimension.

[0063] (5) Channel attention calculation: respectively for and X YOLO Channel attention is calculated using the following formula: , The introduction of channel attention mechanism helps to highlight key features and suppress redundant information, thereby improving the effectiveness of feature representation.

[0064] (6) Spatial-channel fusion weighting: and respectively with A HW Element-wise multiplication generates a weighted feature, expressed as: , By using this element-wise multiplication method, we can ensure the full integration of spatial and channel information while avoiding excessive smoothing of information and loss of detail.

[0065] (7) Fusion Output: and By adding element by element, we obtain the final output feature, expressed as: The fused features not only integrate the multidimensional information from YOLOv11 and SAGAT, but also optimize the expressive power of the features through an attention mechanism. The fused features possess both global structural understanding and local detail awareness capabilities, significantly improving detection accuracy in challenging scenarios such as complex backgrounds, small defects, and occluded areas.

[0066] It should be noted that by combining X YOLO With X SAThis approach effectively integrates information from different perspectives, enriching the diversity of feature representation. Furthermore, the complementarity of multi-view features allows the model to maintain high detection accuracy when identifying occluded parts or deformed defect areas. The CAF module dynamically adjusts the importance of features through an attention mechanism, reducing noise interference and information loss from a single perspective, thereby lowering the false detection rate and detection ambiguity. In addition, the multi-view feature fusion strategy can more accurately identify detailed information in small object detection, while ensuring the integrity and consistency of features in large object detection.

[0067] Step 500: Optimize and train the initial model using the final features and the pre-built loss function to obtain the defect detection model.

[0068] Specifically, to address the sensitivity of traditional loss functions to noise, outliers, and class imbalance, this embodiment constructs an Adaptive Robust Balanced Sparse Loss (ARoBoSS) function, expressed as follows: In this function, the error residual term *u* represents the spatial deviation between the predicted bounding box and the ground truth bounding box. When the predicted result deviates significantly from the ground truth bounding box (*u* is large), ARoBoSS will suppress the contribution of that sample to the total loss, thereby reducing the interference of outliers on model training. By integrating a sparsity suppression mechanism and a dynamic weighting strategy, this function can effectively focus on key samples during the training process, significantly improving the stability and generalization ability of the model in real-world complex environments.

[0069] More specifically, traditional IoU-based loss functions (such as GIoU, DIoU, and CIoU) treat bounding boxes as deterministic coordinate predictions, making it difficult to model localization uncertainties in scenarios with occlusion, blur, or low-quality targets. This embodiment introduces probability-weighted IoU, modeling the four boundaries of the predicted bounding box as independent one-dimensional Gaussian distributions, thereby achieving the modeling and optimization of spatial uncertainty. The expressions for the four boundaries of the predicted bounding box are: ,in Let represent the mean and variance of the i-th boundary of the predicted bounding box, used to model prediction uncertainty. Then, the expected intersection-union ratio (PWIoU) between the predicted and ground truth bounding boxes is calculated through sampling, yielding the probabilistic PWIoU index, calculated as follows: B gt These are the ground truth bounding boxes. The final localization loss function is expressed as: The second term is an uncertainty penalty term, used to suppress high variance predictions and improve the robustness and convergence stability of localization. This is a weighting hyperparameter used to control the degree of influence of the uncertainty penalty term on the total loss.

[0070] Furthermore, to achieve dynamic adjustment of the loss weights, this embodiment also constructs a weight adjustment formula. This formula is based on the idea of ​​exponential decay and can automatically adjust the weight of each loss term according to the training progress, ensuring that the model can focus on different tasks according to the needs of the current stage. The expression is:

[0071] ;

[0072] ;

[0073] ;

[0074] Where t is the current training step or training stage, α loc α class and α conf These are the initial weight coefficients for each loss term. Larger values ​​are used in the early stages of training to encourage the model to focus on specific tasks. β loc β class and β conf γ is used to control the rate of weight changes; its magnitude determines the speed at which the weights of each loss term are updated during training. loc γ class and γ conf These are all hyperparameters used to control the smoothness of weight updates. Based on the weight adjustment formula, the final total loss function dynamically adjusts the weights as the training progresses, and the expression is: ,in, Indicates location loss. The classification loss is represented by the cross-entropy loss, which is obtained by performing binary classification on each category. The expression is: Where C represents the number of categories; y (c) p represents the true label (0 or 1) of class c. (c) This represents the probability that the model predicts the class to be c. The confidence loss is expressed as: Where N represents the number of positive samples, C represents the number of categories, and s i This represents the confidence score of the model's prediction. This represents the true value (1 if there is a target, 0 if there is no target).

[0075] Finally, the initial parameters of the YOLOv11 and SAGAT models are initialized. Based on the calculated loss function value, the model parameters are adjusted to minimize the loss function, resulting in a trained defect detection model. The test set images are then input into the trained defect detection model. After multi-view feature extraction, feature fusion, and model inference, the defect detection results, including defect type and location, are output.

[0076] The beneficial effects of this invention are as follows:

[0077] 1) Local texture and edge information are extracted by a lightweight convolutional neural network, and global topological relationships are modeled by combining a sparse graph attention mechanism, thus achieving collaborative representation of local and global features;

[0078] 2) The channel attention fusion module dynamically weights and fuses local and global features, suppressing redundant information interference and enhancing the response of key features;

[0079] 3) The adaptive robust balance loss function is used to optimize the localization, classification and confidence tasks, which mitigates the impact of noisy data, outliers and class imbalance.

[0080] 4) By using the Top-KAST sparse strategy and adaptive receptive field adjustment mechanism, the computational complexity is reduced and the training and inference efficiency of the model is improved;

[0081] 5) The spatial uncertainty of the predicted box boundary is modeled by using a probability-weighted intersection-union loss function, which optimizes the detection capability of small defects and fuzzy boundaries.

[0082] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0083] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for detecting surface defects in wooden boards based on multi-view feature extraction, characterized in that, Includes the following steps: The image data collected by multiple industrial cameras on the wood panel production line is obtained to obtain the original image set; Local features are obtained by extracting local texture and edge information from the original image set using a lightweight convolutional neural network; Global features are obtained by modeling the global topological relationships in the original image set using a sparse graph attention mechanism. The local features and the global features are dynamically weighted and fused using a channel attention fusion module to obtain the final features. The initial model is optimized and trained using the final features and the pre-built loss function to obtain the defect detection model.

2. The method for detecting surface defects of wooden boards based on multi-view feature extraction according to claim 1, characterized in that, Global topological relationships in the original image set are modeled using a sparse graph attention mechanism to obtain global features, including: The original image set is divided into several image patches, and the image patches are mapped to graph structure data to obtain the initial graph; The adjacency matrix in the initial graph is sparsified using the PageRank algorithm to obtain a sparse adjacency matrix. The key nodes and edges in the sparse adjacency matrix are dynamically filtered using the Top-KAST sparse strategy to obtain a sparse graph. The global features are obtained by extracting features from the sparse graph through an adaptive receptive field adjustment mechanism.

3. The method for detecting surface defects of wooden boards based on multi-view feature extraction according to claim 2, characterized in that, The key nodes and edges in the sparse adjacency matrix are dynamically filtered using a Top-KAST sparse strategy to obtain a sparse graph, including: Calculate the attention score between nodes in the sparse adjacency matrix; The K neighbor nodes with the highest attention scores are retained to obtain the updated sparse adjacency matrix; The node features in the initial graph are updated based on the updated sparse adjacency matrix to obtain the sparse graph.

4. The method for detecting surface defects of wooden boards based on multi-view feature extraction according to claim 1, characterized in that, The local features and the global features are dynamically weighted and fused using a channel attention fusion module to obtain the final features, including: The local features and the global features are concatenated along the channel dimension to obtain a fused feature map; The fused feature map is weighted by spatial dimensions using a spatial attention mechanism to obtain a spatial weight map; The fused feature map is weighted according to the channel dimension using a channel attention mechanism to obtain a channel weight map; The spatial weight map and the channel weight map are multiplied element-wise to obtain the final feature.

5. The method for detecting surface defects of wooden boards based on multi-view feature extraction according to claim 1, characterized in that, The specific steps for pre-constructing the loss function include: The spatial deviation between the predicted bounding box and the ground truth bounding box in the initial model is modeled by a probability-weighted intersection-union loss, thus obtaining the localization loss; The classification loss is obtained by modeling the category semantic embedding and global context in the initial model through meta-learning dynamic focusing factors. The prediction confidence in the initial model is modeled using binary cross-entropy loss to obtain the confidence loss; The loss function is obtained by weighted summation of the localization loss, the classification loss, and the confidence loss.

6. The method for detecting surface defects of wooden boards based on multi-view feature extraction according to claim 5, characterized in that, The loss function is obtained by weighted summing of the localization loss, the classification loss, and the confidence loss, including: Based on the current number of training steps and the preset decay parameters, calculate the dynamic weights of the localization loss, the classification loss, and the confidence loss; The dynamic weights are multiplied by the localization loss, the classification loss, and the confidence loss respectively to generate a weighted loss term; The weighted loss terms are summed to generate the total loss function.

Images