A method for detecting surface defects of particle board based on an improved YOLOv11 model

Abstract

A kind of method for detecting surface defects of shaving board based on improved YOLOv11 model belongs to the technical field of shaving board surface defect detection.The present application introduces prior knowledge for automatic recognition and measurement of shaving board surface defects according to the characteristics of shaving board surface defects, improves the network structure of the algorithm model, introduces a newly designed C3K2-MFC structure, so that the network model can better cope with different scale crack defects and improve the detection capability of the model for multi-scale targets.Multiple scale progressive channel attention blocks and frequency-space attention modules are integrated to better capture the boundary features of oil stain defects and avoid the problem of blurred boundaries.Reduce the calculation redundancy in feature interaction, so that the model is more efficient and accurate when processing multi-scale defects.Use PSConv convolution instead of traditional convolution to effectively expand the field of view of feature extraction, making it easier to identify small target defects, and use the newly designed RSCD detection head to effectively reduce the parameter quantity and improve the inference speed and applicability of the model.

Description

Technical Field

[0001] This invention belongs to the field of particleboard surface defect detection technology, specifically relating to a particleboard surface defect detection method based on an improved YOLOv11 model. Background Technology

[0002] Surface defects in particleboard are a key indicator of its quality. Surface defects not only affect the product's appearance but also reduce its mechanical strength and adversely impact subsequent secondary processing. These defects pose safety hazards to users and cause economic losses to manufacturers. Therefore, surface defect detection has become an indispensable quality control step in particleboard production.

[0003] With the continuous improvement of automation in the particleboard industry, traditional surface defect detection methods are no longer sufficient to meet the needs of high-speed, large-scale production. Under conditions where production line speeds reach 1500mm / s–2000mm / s, traditional manual inspection methods have significant shortcomings: low efficiency, poor reliability, and susceptibility to subjective judgment, leading to inconsistent inspection standards. Furthermore, due to the diverse types and varied forms of surface defects in particleboard, and the fact that some defect features are not obvious, traditional machine vision methods also face difficulties in identifying complex defects. Summary of the Invention

[0004] To address the aforementioned problems, this invention utilizes computer vision technology to design a particleboard surface defect detection network based on an improved YOLOv11 model, enabling accurate identification of surface defects. This method effectively avoids the problems of low efficiency, high false negative rate, and poor judgment accuracy associated with manual inspection, thereby reducing economic losses caused by missed defect detection.

[0005] The technical solution adopted in this invention is:

[0006] A method for detecting surface defects in particleboard based on an improved YOLOv11 model includes the following steps:

[0007] S1. Dataset Creation: Collect image data of particleboard surface defects and divide them into training set, validation set and test set according to a certain ratio; label the images in the training set and validation set according to the defect category;

[0008] S2. Image preprocessing: Data enhancement is performed on defects in the independently acquired particleboard surface images;

[0009] S3. Improved Network Model: The core structure of the YOLOv11 object detection model is improved. The improvement consists of improvements to the backbone C3K2 structure, the backbone convolutional module, the neck structure, and the detection head structure.

[0010] S4. Model Training: Input the training set from S1 into the optimized defect detection model constructed in S3 for supervised training. During the training process, input the validation set into the trained intermediate model for validation to obtain the trained model.

[0011] S5. Model Inference: Load the model parameters trained in S4 into the defect detection model in S3, and input the images of particleboard surface defects from the test set in S1 into the defect detection model for inference to obtain the output results of each block image. Finally, output the result of each image.

[0012] S6. Output Results: The board is initially judged based on the number and size of defects on the particleboard surface, and the board quality is then graded based on the judgment.

[0013] Compared with the prior art, the present invention has the following advantages:

[0014] This invention, based on the characteristics of particleboard surface defects, introduces prior knowledge for the automated identification and measurement of particleboard surface defects, and improves the network structure of the algorithm model. Firstly, a newly designed C3K2-MFC structure is introduced, enabling the network model to better handle crack defects of different scales and improving the model's detection capability for multi-scale targets. The integration of multi-scale progressive channel attention blocks and frequency-spatial attention modules better captures the boundary features of oil stain defects, avoiding boundary ambiguity. Computational redundancy in feature interaction is reduced, making the model more efficient and accurate in handling multi-scale defects. PSConv convolution is used instead of traditional convolution, effectively expanding the feature extraction field of view, making small target defects such as pits and black spots easier to identify. A newly designed RSCD detection head effectively reduces the number of parameters. This improves the model's inference speed and applicability. Attached Figure Description

[0015] Figure 1 This is an overall flowchart of the present invention;

[0016] Figure 2 This is an overall structural diagram of the particleboard surface defect detection invention;

[0017] Figure 3 This is a structural diagram of the MANet of the present invention;

[0018] Figure 4 This is a diagram of the FasterNet structure of the present invention;

[0019] Figure 5 This is a structural diagram of the ConvGLU of the present invention;

[0020] Figure 6 This invention relates to Pinwheel-shaped Conv convolution;

[0021] Figure 7 This is the multi-scale progressive channel attention block of the present invention;

[0022] Figure 8 This is the frequency-spatial attention module of the present invention;

[0023] Figure 9 This invention relates to the Conv_GN structure and RSCD detection head structure. Detailed Implementation

[0024] To better understand the purpose, structure, and function of this invention, the invention will be described in further detail below with reference to the accompanying drawings.

[0025] With the development of mechanization and automation, vision-based particleboard surface defect detection systems are gradually being applied to production lines, achieving unmanned and non-destructive testing. Particleboard surface defects are numerous and complex in pattern. Traditional machine vision algorithms require manual feature design, making them ill-suited to complex and ever-changing defect data, exhibiting poor robustness and generalization. The continuous development of deep learning-based particleboard surface defect detection technology has demonstrated significant advantages over ultrasonic, X-ray, and laser detection technologies in terms of detection accuracy, speed, and reduced labor costs. Deep learning algorithms, trained on a large number of particleboard surface images, can accurately identify various types of defects, such as cracks, glue stains, and oil contamination. Its automated inspection process also greatly improves inspection efficiency, making quality control in the production process more rigorous and efficient.

[0026] Therefore, this invention is an improvement on the deep learning-based particleboard surface defect detection technology.

[0027] like Figure 1 As shown, a method for detecting surface defects in particleboard based on an improved YOLOv11 model includes the following steps:

[0028] Step 1: Dataset Creation: Collect images of surface defects in particleboard and divide them into training, validation, and test sets in a 0.8:0.1:0.1 ratio. Label the images in the training and validation sets according to the defect categories.

[0029] Defect categories include shavings, black spots, cracks, oil stains, and dents.

[0030] Step 2, Image Preprocessing: Data augmentation techniques are used to enhance the defects in the self-collected particleboard surface images, enriching and diversifying the dataset to improve the robustness and accuracy of the model. In Step 2, rotation, noise addition, scaling, and cropping are used to expand the number of images in the dataset.

[0031] Step 3: Optimize the model, such as Figure 2 As shown ( Figure 2 The diagram shows the overall structure of the improved YOLOv11 model. The MANet network, FasterNet network module, and ConvGLU are built-in sub-units of the C3K2-MFC module, and their detailed structures correspond to... Figure 3 , Figure 4 , Figure 5 This paper proposes an improved method for detecting surface defects in particleboard based on the improved YOLOv11 model. The improved method consists of four parts: improved C3K2 structure of the model backbone, improved convolution of the model backbone, improved feature fusion and upsampling block (neck structure improvement), and improved lightweight detection head.

[0032] 1. Improved C3K2 Backbone Structure: The backbone improvement involves replacing the original C3K2 modules in the backbone structure with a newly designed improved C3K2-MFC structure. The specific innovations of the C3K2-MFC module are as follows:

[0033] (1) A new cross-scale feature interaction sub-branch is added, which integrates the 1×1 convolution, depthwise separable convolution and C2f module of the MANet network. After splicing and 1×1 convolution compression, 2c channel features are generated to improve the ability to capture small-sized defects in particleboard.

[0034] (2) Embed the PWConv and PConv structures of the FasterNet network, replacing the standard 3×3 convolution, and press "c> And c- > "Divide channels to reduce computational load and adapt to real-time detection;"

[0035] (3) ConvGLU is introduced as a channel mixer. Its gate signal originates from its own neighboring features. Combined with 3×3 depth separable convolution, a dynamic channel adjustment mechanism is constructed. At the same time, residual connections of feature element-wise addition are added to avoid gradient vanishing and optimize the differentiation effect of complex defects.

[0036] 1.1 The cross-scale feature interaction sub-branch of the C3K2-MFC module uses the MANet network as the core feature fusion unit. That is, the MANet network is a key component of the C3K2-MFC module for integrating multiple types of convolutional features. Through its unique hybrid aggregation mechanism, the MANet network can efficiently capture feature information at different scales and locations, providing richer feature representations for subsequent object detection tasks. For example... Figure 3The network architecture shown integrates three typical convolutional variants: 1×1 convolution, used for channel-level feature recalibration, effectively adjusts the relationships between channels, enhances the semantic depth of features, and thus improves the model's classification and localization capabilities; Depthwise separable convolution (DSConv), used for efficient spatial feature processing, improves model efficiency and accuracy by splitting traditional convolution operations into two independent steps, thereby enhancing the model's expressive power; and the C2f module enhances feature hierarchical integration by fusing features from different stages through cross-stage residual connections. This design helps improve the model's non-linear representation capabilities, thus better handling complex image features and obtaining richer feature representations.

[0037] The calculation formula for the MANet module is shown below:

[0038]

[0039]

[0040]

[0041]

[0042]

[0043]

[0044]

[0045] …

[0046]

[0047] Where X mid The number of channels is 2c, and each X1, X2, ..., X 4+n The number of channels is c for each feature. Finally, these three types of features are fused through a concatenation operation and compressed using a 1×1 convolution to generate an output feature with 2c channels. The specific formula is shown below:

[0048]

[0049] 1.2 The C3K2-MFC module adopts the FasterNet network module as a lightweight convolutional core unit to optimize the computational efficiency of the C3K2-MFC module, thereby supporting the performance adaptation of the YOLOv11 model in the real-time detection of particleboard surface defects. The structure of the FasterNet network module (FasterNet Block) is as follows: Figure 4 The proposed Partial Convolution (PConv) operation, combined with Pointwise Convolution (PWConv), is used to efficiently extract spatial features. Furthermore, a stepwise convolutional structure is designed, consisting of a 3×3 PConv layer and two 1×1 PWConv layers. This structure not only reduces FLOPs but also significantly enhances the model's inference performance.

[0050] The FasterNet network module uses a two-step convolution structure of PWConv and PConv, and its FLOPs calculation formula can be expressed as:

[0051] .

[0052] Where h and w represent the height and width of the feature map, respectively. This refers to the number of the first or last consecutive channels in a contiguous memory access pattern, while k represents the size of the convolutional kernel. The formula satisfies c > 0. , and c- > .

[0053] 1.3 The C3K2-MFC module uses ConvGLU as a feature propagation enhancement component to improve the feature stability of the C3K2-MFC module in deep networks, thereby optimizing the YOLOv11 model's feature extraction capability for complex defects in particleboard. ConvGLU...

[0054] Convolutional Gated Linear Unit (ConvGLU) is a novel channel mixer that combines the ideas of Gated Linear Unit (GLU) and convolution, designed to better adapt to the needs of visual tasks. Compared to the Squeeze-and-Excitation (SE) module, ConvGLU's gated signal originates from its own neighboring features, rather than global average pooling. This mechanism provides greater flexibility in channel attention and enables more effective capture of local information. Compared to traditional GLUs,

[0055] like Figure 5As shown, ConvGLU achieves lower computational complexity by introducing 3×3 depthwise separable convolutions, while maintaining the same number of parameters as MLP and GLU, thus improving the model's efficiency and robustness. When the feature map is input, it first undergoes two linear transformations, followed by depthwise separable convolutions, an activation function, and a final linear transformation. Finally, an addition operator adds the output of the current layer to the result of the previous layer, forming a residual that is connected to the final output.

[0056] 2. Improved Backbone Convolution: The backbone convolution module is improved to Pinwheel-shaped Convolution (PSConv). PSConv significantly expands the receptive field through asymmetric padding and a specific convolution kernel design, better adapting to the pixel distribution characteristics of small targets. The method applies batch normalization (BN) and sigmoid linear units (SiLU) after each convolution. Figure 6 As shown, the first layer of PConv performs parallel convolutions:

[0057]

[0058]

[0059]

[0060]

[0061] in: It is a convolution operator. It is an output channel of Convolution kernel. Padding parameters. This represents the number of pixels padded in the left, right, top, and bottom directions. After the first layer of interleaved convolution, the height of the output feature map is... ,width and number of channels The relationship between the input feature map and the input feature map is shown in the following formula:

[0062]

[0063] in This refers to the number of channels in the final output feature map of the PSConv module. It is the convolution stride. The results of the first layer of interleaved convolutions are concatenated. The output calculation is as follows:

[0064]

[0065] Finally, the concatenated tensors are passed through a convolution kernel. Normalization is performed, no padding required. The height and width of the output feature map are adjusted to preset values. and This allows PSConv to be interchanged with Conv layers and serves as a channel attention mechanism to analyze the contribution of different convolutional directions to the output. Final output. The calculation formula is as follows:

[0066]

[0067]

[0068] 3. Feature Fusion and Upsampling Block Improvement (Neck Structure Improvement): The neck structure improvement replaces the feature fusion and upsampling operations of the neck structure in the YOLOv11 model structure with a Multi-scale PCA-FSA module (Multi-scale Progressive Channel Attention Module). The model is able to better capture the details and boundary features for detecting oil stain defects.

[0069] FSA and MPCA together form the Multi-scale PCA-FSA module, which is used to replace the feature fusion and upsampling operations in the neck of YOLOv11. The two complement each other. MPCA is responsible for cross-scale channel feature selection and integration, and outputs feature maps with multi-scale adaptability, but it does not fully explore spatial and frequency details. Therefore, its output is used as the input of FSA, and FSA optimizes the details through Fourier transform and spatial attention.

[0070] 3.1, such as Figure 7 As shown, the MPCA module has two inputs, and the current encoder outputs a feature map. and the next encoder output feature map .Then, and Dimensionality reduction and feature extraction are performed using two independent Global Average Pooling (GAP) and 1×1 convolutional layers, followed by concatenation. Finally, another 1×1 convolutional layer is used to fuse the two feature maps, resulting in multi-scale channel attention. This process can be defined by the following formula:

[0071]

[0072]

[0073] In the formula: Conv1×1 represents a 1×1 convolutional layer, and Concate represents concatenation along the channel dimension. The A layer fuses channel features from two adjacent scales and completes cross-scale information exchange. Then, A is divided into two parts. and Its dimensions and and Maintain consistency. Then... and will be respectively with the corresponding and Multiply to obtain a weighted feature map. and The feature map is then processed using a transposed convolutional layer. Upsampled by 2 times, and with Matching the number of channels yields... Finally, regarding and Perform element-wise addition to achieve feature fusion and obtain This is the output of the MPCA module, and the process can be represented by the following formula:

[0074]

[0075] =

[0076]

[0077] and These represent element-wise multiplication and element-wise addition, respectively, with UP referring to the transposed convolutional layer. Finally, the output of each MPCA block is used for dual-domain feature learning (sum and frequency domains) through the corresponding FSA block.

[0078] 3.2 The main function of the FSA module (Frequency-Spatial Attention Module) is to learn features in both the spatial and frequency domains simultaneously, thereby extracting image information more comprehensively and improving segmentation accuracy.

[0079] like Figure 8 As shown, the output of the MPCA block is given first. Then, a two-dimensional discrete Fourier transform (2D-DFT) is used to convert it to the frequency domain. Next, the frequency features are divided into high-frequency features. and low frequency This requires the use of two masks. and .for A square with side length n is placed at the center of the mask and assigned a value of 1, while the rest of the area is assigned a value of 0. Then, for... and Perform element-wise multiplication to obtain the low-frequency components. .for The small square region is assigned a value of 0, and the remaining regions are assigned a value of 1. Then, for and Perform element-wise multiplication to obtain high-frequency components. Next, regarding Apply learnable global filters Filtering is performed. Then, the original high-frequency features are... With the filtered Reintegrate to obtain the feature map after global learning. Subsequently, the feature map is transformed back to the spatial domain using the two-dimensional inverse discrete Fourier transform (2D-IDFT), resulting in... Meanwhile, for the output from MPCA It also uses spatial attention in the spatial branch to facilitate feature learning. The final output of the FSA module can be obtained through... get.

[0080] Lightweight Detection Head Improvement: The improved detection head structure introduces an RSCD detection head, which is as follows... Figure 9 As shown, the RSCD detection head structure includes a Conv_GN module and a DBB module. It first receives input features from layers P3, P4, and P5 of the improved YOLOv11 model's neck structure and processes them through a 3×3 Conv_GN convolution. The BN (batch normalization) layer in the CBS module is modified to a GN (group normalization) layer. Next, information is integrated through shared convolutional units of the DBB convolution, aiming to reduce information redundancy and increase the learning opportunities of adjacent feature information. Finally, the information extracted by the shared convolution is input into the classification and regression heads, followed by scaling layers to enhance the preservation of multi-scale features.

[0081] Step 4: Model Training: Input the training set into the optimized defect detection model constructed in Step 3 for supervised training. During the training process, input the validation set into the trained intermediate model for validation to obtain the trained model parameters.

[0082] During training, the optimizer for the training parameters was AdamW, and the learning rate was set to 0.01.

[0083] Step 5, Model Inference: Load the model parameters trained in Step 4 into the network in Step 3, and input the images of particleboard surface defects from the test set in Step 1 into the defect detection model for inference to obtain the output results of each block image. Finally, output the result of each image.

[0084] Step 6, Output Results: Make a preliminary judgment on the board based on the number and size of defects on the particleboard surface, and then grade the board quality based on the judgment.

[0085] The specific method is as follows: First, set several particleboard quality grades; then set a reference table, the header of which includes the actual size range, quantity range and particleboard quality grade; obtain the particleboard quality grade by finding that the actual size and quantity of the particleboard surface defects to be judged fall within the actual size range and quantity range of the reference table respectively.

[0086] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for detecting surface defects in particleboard based on an improved YOLOv11 model, characterized in that: Includes the following steps: S1. Dataset Creation: Collect image data of particleboard surface defects and divide them into training set, validation set and test set according to a certain ratio; label the images in the training set and validation set according to the defect category; S2. Image preprocessing: Data enhancement is performed on defects in the independently acquired particleboard surface images; S3. Improved Network Model: The core structure of the YOLOv11 object detection model is improved. The improvement consists of improvements to the backbone C3K2 structure, the backbone convolutional module, the neck structure, and the detection head structure. S4. Model Training: Input the training set from S1 into the optimized defect detection model constructed in S3 for supervised training. During the training process, input the validation set into the trained intermediate model for validation to obtain the trained model. S5. Model Inference: Load the model parameters trained in S4 into the defect detection model in S3, and input the images of particleboard surface defects from the test set in S1 into the defect detection model for inference to obtain the output results of each block image. Finally, output the result of each image. S6. Output Results: Based on the quantity and size of surface defects in the particleboard, a preliminary judgment is made on the particleboard quality. Subsequent assessments are then used to grade the particleboard quality. In S3, the C3K2 backbone structure is improved: the original C3K2 module in the backbone structure is replaced by a C3K2-MFC module. The C3K2-MFC module adds a cross-scale feature interaction sub-branch with the MANet network as the core feature fusion unit, adopts the FasterNet network module as the lightweight convolutional core unit, and introduces ConvGLU as the channel mixer. The MANet network architecture integrates the following three typical convolutional variants. 1×1 convolution: used for feature recalibration at the channel level; Depthwise separable convolution: used for efficient spatial feature processing; C2f module: used to enhance feature hierarchical integration, fusing features from different stages through cross-stage residual connections; These three types of features are fused through a concatenation operation and compressed using a 1×1 convolution to generate an output feature with 2c channels. The specific formula is shown below: The FasterNet network module adopts a two-step convolution structure of PWConv and PConv, and its FLOPs calculation formula is expressed as: , Where h and w represent the height and width of the feature map, respectively. This represents the number of the first or last consecutive channels in the continuous memory access mode, k represents the size of the convolution kernel, and the formula satisfies c >

0. , and c- > , The ConvGLU introduces a 3×3 depthwise separable convolution, and ConvGLU can be combined with positional encoding. When a feature map is input, it first undergoes two linear transformations, followed by a depthwise separable convolution, an activation function, and a final linear transformation. Finally, an addition operator adds the output of the current layer to the result of the previous layer, forming a residual that is connected to the final output.

2. The method for detecting surface defects in particleboard based on an improved YOLOv11 model according to claim 1, characterized in that: In S3, the backbone convolution module is improved to PSConv. The first layer of PConv performs the following parallel convolutions: in: It is a convolution operator. It is an output channel of Convolution kernel, padding parameters This indicates the number of pixels filled in the left, right, top, and bottom directions. After the first layer of interleaved convolution, the height of the output feature map is... ,width and number of channels The relationship between the input feature map and the input feature map is shown in the following formula: in: This refers to the number of channels in the final output feature map of the PSConv module. It is the convolution stride; the results of the first layer of interleaved convolutions are concatenated. The output calculation is as follows: Finally, the concatenated tensors are passed through a convolution kernel. Normalization is performed, and the height and width of the output feature map are adjusted to preset values. and This allows PSConv to be interchanged with Conv layers and serves as a channel attention mechanism to analyze the contribution of different convolutional directions to the output, ultimately producing the output. The calculation formula is as follows: 。 3. The method for detecting surface defects in particleboard based on an improved YOLOv11 model according to claim 1, characterized in that: In S3, the neck structure improvement uses a Multi-scale PCA-FSA module to replace the feature fusion and upsampling operations of the neck structure in the YOLOv11 model structure; the Multi-scale PCA-FSA module consists of FSA and MPCA. The MPCA module has two inputs: the current encoder output feature map. and the next encoder output feature map ,Then, and Dimensionality reduction and feature extraction are performed using two independent global average pooling and 1×1 convolutional layers, followed by concatenation. Finally, another 1×1 convolutional layer is used to fuse the two feature maps to obtain multi-scale channel attention. The above process can be defined by the following formula: In the formula: Conv1×1 represents a 1×1 convolutional layer, Concate represents concatenation along the channel dimension, A fuses the channel features of two adjacent scales and completes cross-scale information exchange, and then A is divided into two parts. and Its dimensions and and Maintain consistency, then and will be respectively with the corresponding and Multiply to obtain a weighted feature map. and The feature map is then processed using a transposed convolutional layer. Upsampled by 2 times, and with Matching the number of channels yields... Finally, regarding and Perform element-wise addition to achieve feature fusion and obtain This serves as the output of the MPCA module and also as the input of the FSA module. The FSA and MPCA modules together form the Multi-scale PCA-FSA module, used to replace the feature fusion and upsampling operations in the YOLOv11 neckline. Their functions are complementary: MPCA is responsible for cross-scale channel feature selection and integration, outputting a feature map with multi-scale adaptability, but it doesn't fully exploit spatial and frequency details; therefore, its output is used as the input to FSA, which optimizes details through Fourier transform and spatial attention. Thus, the output of the MPCA module is also the input to the FSA module. This process can be expressed by the following formula: = and These represent element-wise multiplication and element-wise addition, respectively. UP refers to the transposed convolutional layer. Finally, the output of each MPCA block will be used for dual-domain feature learning of sum and frequency through the corresponding FSA block.

4. The method for detecting surface defects in particleboard based on an improved YOLOv11 model according to claim 3, characterized in that: The FSA module is used to learn features in both the spatial and frequency domains simultaneously; first, the output of the MPCA block is given. Then, a two-dimensional discrete Fourier transform is used to convert it to the frequency domain, and the frequency characteristics are divided into high-frequency features. and low frequency Using two masks, and , for A square with side length n is placed at the center of the mask and assigned a value of 1, while the rest of the area is assigned a value of 0. Then, for... and Perform element-wise multiplication to obtain the low-frequency components. , for The small square region is assigned a value of 0, and the remaining regions are assigned a value of 1. and Perform element-wise multiplication to obtain high-frequency components. , right Apply learnable global filters Perform filtering, and then extract the original high-frequency features. With the filtered Reintegrate to obtain the feature map after global learning. Subsequently, the feature map is transformed back to the spatial domain using a two-dimensional inverse discrete Fourier transform, resulting in... Meanwhile, for the output from MPCA Spatial attention is used in the spatial branch to facilitate feature learning, and the final output of the FSA module can be obtained through... get.

5. The method for detecting surface defects in particleboard based on an improved YOLOv11 model according to claim 1, characterized in that: In S3, the improved detection head structure introduces an RSCD detection head, which includes a Conv_GN module and a DBB module. The RSCD detection head structure first receives input features from the P3, P4, and P5 layers of the neck structure of the improved YOLOv11 model and processes them by inputting them into a Conv_GN convolution with a 3×3 kernel. The BN layer in the CBS module is modified to a GN layer. Next, information is integrated through the shared convolutional units of the DBB convolution. Finally, the information extracted by the shared convolution is input into the classification and regression heads, and then the features are scaled by a Scale layer.

6. The method for detecting surface defects in particleboard based on an improved YOLOv11 model according to claim 1, characterized in that: The specific method of S6 is as follows: set several particleboard quality grades; Then, a reference table is set up, the header of which includes the actual size range, quantity range, and particleboard quality grade; The corresponding particleboard quality grade is obtained by determining whether the actual size and number of the surface defects to be judged fall within the actual size and quantity range of the reference table.

Images