A veneered wood-based panel surface defect detection method based on an improved YOLO11 model

Abstract

This application discloses a method for detecting surface defects in veneer engineered wood panels based on an improved YOLO11 model. The method includes: processing a first veneer engineered wood panel sample image using adaptive Mosaic enhancement technology based on bounding box position information to obtain a second veneer engineered wood panel sample image; replacing the neck network in the original YOLO11 model with a bidirectional feature pyramid network containing a normalized weight fusion mechanism to obtain an improved YOLO11 model; inputting the second veneer engineered wood panel sample image into the improved YOLO11 model to obtain detection results; performing backpropagation training on the improved YOLO11 model based on the detection results, corresponding real target detection labels, and a preset loss function until convergence conditions are met to obtain a trained improved YOLO11 model; and inputting the veneer engineered wood panel image to be detected into the trained improved YOLO11 model to obtain the target detection result of the veneer engineered wood panel image to be detected. This method can improve the feature extraction capability for minor defects, reduce the false negative and false positive rates, and improve the accuracy and efficiency of target detection.

Description

Technical Field

[0001] This application relates to the fields of defect detection and artificial intelligence technology for decorative engineered wood panels, specifically to a method for detecting surface defects in decorative engineered wood panels based on an improved YOLO11 model. Background Technology

[0002] Decorative engineered wood panels are a primary material for furniture manufacturing and interior decoration. As a type of melamine-impregnated paper, they are one of the mainstream products of melamine-impregnated paper-faced particleboard. According to the national recommended standard GB / T 15102-2017 "Impregnated Paper-Faced Fiberboard and Particleboard," decorative engineered wood panels are prone to surface defects such as dry and wet spots, stains, and cracks during production. Machine vision recognition enables online detection to replace manual inspection. Currently, there is a significant mismatch between the efficiency of defect detection on the production line and the high-speed production pace of decorative panels. Traditional methods, primarily relying on manual visual inspection, are susceptible to subjective judgment interference, have unstable detection accuracy, and their detection speed is difficult to match the production line rhythm, thus hindering the efficient and continuous production of high-quality decorative panels.

[0003] With the continuous development of industrial imaging and deep learning algorithms, surface defect detection based on machine vision has been widely applied and rapidly promoted in industrial production. YOLO11, released in 2024, is a new generation of object detection models in the YOLO series. It inherits the series' characteristics of fast detection speed and high recognition accuracy, and further optimizes network structure, task scalability, and engineering deployment efficiency. Its core improvements are reflected in deep optimization of the network structure: the backbone network uses more efficient C3k2 modules to replace part of the C2f structure; feature extraction capabilities are improved by optimizing the convolution kernel scale and channel dimension; and a dynamic attention mechanism is introduced to achieve adaptive focusing on key target regions. The neck PAN-FPN structure has been lightweighted. The detection head adopts a unified design, combined with deformable convolution to enhance the ability to capture irregular targets, while simplifying the training and inference processes. In terms of functional expansion, YOLO11 not only supports conventional object detection but also integrates multi-task capabilities such as instance segmentation, pose estimation, and object tracking. The unified network architecture and training strategy lower the threshold for practical development and deployment.

[0004] Although YOLO11 achieves an average precision improvement of approximately 2% to 3% on the COCO public dataset compared to YOLOv8 and possesses multi-task scalability, its key metrics such as precision and recall still have room for improvement when applied to scenarios involving the detection of small targets, such as surface defects in engineered wood panels. Furthermore, in actual industrial production line deployments, edge devices generally face limitations in computing power and resource constraints. Balancing inference speed with model complexity while ensuring detection accuracy remains a critical issue that urgently needs to be addressed for the practical application of industrial defect detection models.

[0005] In summary, using machine vision to detect surface defects in decorative engineered wood panels has promising application prospects. However, existing methods still have two shortcomings: first, they are weak in extracting features from minute defects, and shallow details of small targets are easily lost during feature transfer, leading to high rates of missed and false detections; second, in resource-constrained industrial production line environments, detection models struggle to simultaneously balance inference efficiency and detection accuracy. Summary of the Invention

[0006] In view of this, this application provides a method for detecting surface defects of decorative artificial boards based on an improved YOLO11 model. This method can improve the feature extraction capability of minute defects, reduce the rate of missed detection and false detection, and improve detection efficiency and accuracy.

[0007] The surface defect detection method for decorative engineered wood panels based on the improved YOLO11 model provided in this application includes: An adaptive Mosaic enhancement technique based on bounding box location information is used to process the first veneer panel sample image to obtain the second veneer panel sample image. The neck network in the original YOLO11 model is replaced from a single top-down path aggregation network to a bidirectional feature pyramid network containing a normalized weight fusion mechanism, resulting in an improved YOLO11 model. The second veneer panel sample image is then input into the improved YOLO11 model to obtain the detection results. The normalized weight fusion mechanism uses normalized weight parameters to weight and fuse feature maps at different scales. Each scale of feature map corresponds to a weight parameter, which is determined during the training of the improved YOLO11 model. Based on the detection results, the corresponding real target detection labels, and a preset loss function, the improved YOLO11 model is trained through backpropagation until the convergence condition is met, resulting in a trained improved YOLO11 model. The veneer panel image to be detected is then input into the trained improved YOLO11 model to obtain the target detection results for the veneer panel image.

[0008] Optionally, the first decorative panel sample image includes multiple images; the processing of the first decorative panel sample image using adaptive Mosaic enhancement technology based on bounding box position information includes: Each first-surface engineered wood sample image is enlarged and cropped at a random position to obtain a sub-image; based on the geometric center coordinates of the defect annotation boxes in the four randomly selected sub-images, as well as the width and height of the defect annotation boxes, the theoretical center coordinates of the set of all defect annotation boxes in the four sub-images are calculated; using the theoretical center coordinates as the splicing center, the four randomly selected sub-images are cropped and spliced ​​to obtain the second-surface engineered wood sample image.

[0009] Optionally, the theoretical center coordinates of all defect annotation boxes in the four cropped sub-figures. The calculation formula is expressed as: ; in, , They represent the first Zhang extracted the center of the annotation box in the sub-image. coordinates and coordinate, , They represent the first Zhang extracted the width and height of the annotation box in the sub-image. This is a random coordinate offset, with a value range of [1, 256].

[0010] Optionally, the improved YOLO11 model includes a backbone network, a bidirectional feature pyramid network with a normalized weight fusion mechanism, and a head network; the bidirectional feature pyramid network with a normalized weight fusion mechanism includes three multi-scale feature fusion branches, each of which has a different resolution of the initial input feature map. The step of inputting the sample image of the second decorative artificial board into the improved YOLO11 model to obtain the detection results includes: The sample image of the second veneer artificial board is input into the backbone network to obtain a first feature map, a second feature map, and a third feature map. The resolution of the first feature map is higher than that of the second feature map, and the resolution of the second feature map is higher than that of the third feature map. The first feature map is used as the initial input of the first multi-scale feature fusion branch, the second feature map is used as the initial input of the second multi-scale feature fusion branch, and the third feature map is used as the initial input of the third multi-scale feature fusion branch. The first feature map, the second feature map, and the third feature map are fused by combining cross-branch fusion and normalized weight fusion mechanisms to obtain the target fused feature map. The target fused feature map is input into the head network to obtain the detection result.

[0011] Optionally, each multi-scale feature fusion branch includes a first upsampling module, a first weighted feature fusion module, a first feature extraction module, a first feature concatenation module, a second feature extraction module, a downsampling module, a second weighted feature fusion module, a second feature concatenation module, a third feature extraction module, a second upsampling module, a third weighted feature fusion module, a third feature concatenation module, and a fourth feature extraction module; The processing steps of the first multi-scale feature fusion branch include: The first feature map is input to the first upsampling module to obtain an upsampled feature map. The upsampled feature map is then weighted by the first weighted feature fusion module based on the first normalized weights corresponding to the first feature map, resulting in a first weighted feature map. Features are extracted from the first weighted feature map using the first feature extraction module. The extracted feature map and the second feature map are then input to the first feature concatenation module for concatenation. The feature map is then sequentially passed through the second feature extraction module and the downsampling module to obtain a first reference feature map. The first reference feature map is then weighted by the second weighted feature fusion module based on the second normalized weights corresponding to the second feature map, resulting in a second weighted feature map. Finally, the second weighted feature map, the upsampled feature map, and the first weighted feature map are combined. The images are fused to obtain a first fused feature map. The first fused feature map and the third feature map are then concatenated using a second feature stitching module and input into a third feature extraction module to obtain a second reference feature map. The second reference feature map is then input into a second upsampling module, and the sampled feature map is weighted by a third weighted feature fusion module based on the third normalized weight corresponding to the third feature map to obtain a third weighted feature map. The third weighted feature map, the first fused feature map, and the second reference feature map are then fused to obtain a second fused feature map. The second fused feature map and the third feature map are then concatenated using a third feature stitching module, and features are extracted using a fourth feature extraction module to obtain the feature map output by the first multi-scale feature fusion branch.

[0012] Optionally, the preset loss function is constructed based on the distribution focus loss function, the classification loss function, and the localization loss function based on target scale weights. Represented as: ; in, and Let represent the distribution focus loss function, the classification loss function, and the localization loss function based on target scale weights, respectively. and These are the weighting factors for the three loss functions, with values ​​of 1, 0.5, and 7.5 respectively.

[0013] Optionally, a localization loss function based on target scale weights. Represented as: ; in, , representing the target scale weight factor, This is an adjustment coefficient used to control the rate of increase in weight. This indicates the area of ​​the current annotation box. This represents the area of ​​the smallest bounding box in the dataset. This represents the ratio of the intersection to the union of the predicted target detection bounding box and the labeled bounding box. This indicates the distance deviation between the center of the predicted bounding box and the center of the labeled bounding box. This represents the aspect ratio deviation, used to measure the difference in aspect ratio between the predicted target detection box and the labeled box. Indicates directional consistency deviation, used to measure the deviation of the directional trend between the predicted target detection box and the labeled box.

[0014] In this embodiment, an adaptive Mosaic enhancement technique based on bounding box location information is used to process the first veneer panel sample image to obtain a second veneer panel sample image. The neck network in the original YOLO11 model is replaced by a single top-down path aggregation network with a bidirectional feature pyramid network containing a normalized weight fusion mechanism to obtain an improved YOLO11 model. The second veneer panel sample image is input into the improved YOLO11 model to obtain the detection result. Based on the detection result, the corresponding real target detection label, and the preset loss function, the improved YOLO11 model is backpropagated and trained until the convergence condition is met to obtain the trained improved YOLO11 model. The veneer panel image to be detected is input into the trained improved YOLO11 model to obtain the target detection result of the veneer panel image to be detected. Thus, the adaptive Mosaic enhancement technique uses the bounding box position information as a kind of "pseudo-attention," which ensures that the stitching center falls in the area with dense defect targets. This results in the enhanced second-finish artificial board sample image containing multiple effective targets, increasing the effective sample size during training and improving the sensitivity of the YOLO11 model to low-contrast defects, thereby improving the target detection accuracy of the improved YOLO11 model. By replacing the neck network in the original YOLO11 model with a single top-down path aggregation network and a bidirectional feature pyramid network with a normalized weight fusion mechanism, the improved YOLO11 model can simultaneously capture high-level abstract semantics and low-level fine details, effectively fusing features at different scales, reducing information loss, thereby improving the feature extraction capability for small defects, reducing false negative and false positive rates. Furthermore, the normalized weight fusion mechanism only needs to handle simple scalar division, with almost no increase in the number of parameters and computational load, thus improving both the accuracy and efficiency of target detection. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1A schematic flowchart illustrating a surface defect detection method for decorative engineered wood panels based on an improved YOLO11 model, provided for embodiments of this application; Figure 2 A schematic diagram illustrating the process of processing a first-faced artificial board sample image using an adaptive Mosaic enhancement technique based on annotation box position information provided in this application; Figure 3 A schematic diagram of the structure of a YOLO11 model provided in this application; Figure 4 A schematic diagram of the structure of a multi-scale feature fusion branch of a bidirectional feature pyramid network with a normalized weight fusion mechanism provided in this application; Figure 5 A comparison chart of target detection results of YOLO11s and an improved YOLO11 model provided for this application; Figure 6 A comparison of target detection results between another YOLO11s and an improved YOLO11 model provided in this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0018] It should be noted that the terms "first, second, third, and fourth" used in the embodiments of the present invention are only used to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third, and fourth" can be interchanged in a specific order or sequence where permitted, so that the embodiments of the present invention described herein can be implemented in an order other than that illustrated or described herein.

[0019] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0020] This application provides a method for detecting surface defects in decorative engineered wood panels based on an improved YOLO11 model, such as... Figure 1 The diagram shown is a flowchart illustrating a surface defect detection method for decorative engineered wood panels based on an improved YOLO11 model, according to an embodiment of this application. The method includes: S101. The first decorative artificial board sample image is processed using adaptive Mosaic enhancement technology based on the position information of the annotation box to obtain the second decorative artificial board sample image.

[0021] In some embodiments, the first decorative engineered wood panel sample image is a pre-collected training data sample. The first decorative engineered wood panel sample includes four main categories: light pigment color, dark pigment color, light texture, and dark texture, which can comprehensively cover the possible conditions of different batches on the decorative engineered wood panel production line. After obtaining the first decorative engineered wood panel sample image, the defects in the first decorative engineered wood panel sample image are labeled using the professional image annotation tool LabelImg.

[0022] In some embodiments, considering the need to improve the balance between detection accuracy and inference speed of the YOLO11 model, this application crops the pre-acquired original veneer panel images into 512×512 pixel sample images and performs grayscale processing on the cropped sample images. Furthermore, since surface defects of the same type of veneer panel can vary in area, shape, and color, each grayscale-processed sample image is preprocessed using methods such as vertical flipping, horizontal flipping, and random 90° rotation to obtain multiple first veneer panel sample images.

[0023] In some embodiments, Mosaic enhancement technology can fuse information from multiple independent scenes in a single image, improving sample diversity. Mosaic enhancement involves randomly selecting four different images, randomly scaling each image, and cropping it at a random location to obtain a sub-image. This sub-image is then stitched together with three other processed sub-images in a 2×2 grid to form a new image. However, the native Mosaic enhancement technology built into YOLO11 has significant limitations due to the characteristics of surface defects in veneer engineered wood panels. Because surface defects in veneer engineered wood panels are often small, the random sampling mechanism of traditional Mosaic enhancement technology may significantly reduce the number of defects in the stitched image, or even result in a completely defect-free image after stitching together four normal areas. Therefore, this application proposes an adaptive Mosaic enhancement technology based on bounding box position information. This adaptive Mosaic enhancement technology utilizes the existing bounding box position information of the first veneer engineered wood panel sample image as a "pseudo-attention" mechanism, ensuring that the stitching center falls in the area with the highest concentration of defect targets. This ensures that the enhanced second veneer engineered wood panel image contains all defects (dried flowers, wet flowers, stains, cracks, etc.).

[0024] In some embodiments, the first decorative panel sample image is processed using an adaptive Mosaic enhancement technique based on the position information of the annotation boxes, including: enlarging each first decorative panel sample image and cropping it at a random position to obtain a sub-image; calculating the theoretical center coordinates of the set of all defect annotation boxes in the four randomly selected sub-images based on the geometric center coordinates of each defect annotation box in the four randomly selected sub-images, as well as the width and height of the defect annotation boxes; and cropping and stitching the four randomly selected sub-images with the theoretical center coordinates as the stitching center to obtain the second decorative panel sample image.

[0025] In some embodiments, the process of processing the first veneer panel sample image using adaptive Mosaic enhancement technology based on bounding box position information is as follows: Figure 2 As shown. The geometric center coordinates of the defect annotation boxes in the cropped sub-images are the pixel coordinates of the geometric center of the defect annotation boxes. The theoretical center coordinates of the set of all defect annotation boxes in the four cropped sub-images are... The calculation formula can be expressed by formula (1): (1); in, , They represent the first Zhang extracted the center of the annotation box in the sub-image. coordinates and coordinate, , They represent the first Zhang extracted the width and height of the annotation box in the sub-image. This is a random coordinate offset, with a value range of [1, 256].

[0026] It should be noted that standard Mosaic enhancement techniques stitch together multiple cropped sub-images, and direct weighting can lead to inconsistent positional distributions. Normalization methods can adjust the coordinates by weighted scaling, bringing them back to a reasonable numerical range and maintaining a uniform distribution, eliminating the offsets and anomalies caused by weight accumulation. Therefore, this application employs an adaptive Mosaic enhancement technique based on bounding box position information, transforming the coordinates of the original image to their corresponding coordinates in the new image, and fixing the generated geometric center within the image range, ensuring it does not exceed the boundaries. Here, a normalization calculation method (using the M-function for normalization) is used, and the M-function also includes weights defined based on area relationships. The smaller the target bounding box area, the greater the weight; the larger the target bounding box area, the smaller the weight. This weighting can balance the contribution of targets of different sizes to the center coordinates, avoid large targets dominating the coordinate calculation, and is more suitable for situations where targets are truncated or have inconsistent sizes in adaptive Mosaic enhancement technology.

[0027] Assume the original theoretical center coordinates of the set of all defect annotation boxes in the four extracted sub-images are: Then the normalized coordinates of the theoretical center are: That is, formula (1), where, This represents the M-function.

[0028] S102. Replace the neck network in the original YOLO11 model with a single top-down path aggregation network and a bidirectional feature pyramid network that includes a normalized weight fusion mechanism to obtain an improved YOLO11 model. Input the sample image of the second veneer artificial board into the improved YOLO11 model to obtain the detection results.

[0029] Among them, the normalized weight fusion mechanism performs weighted fusion of feature maps of different scales through normalized weight parameters. Each scale of feature map corresponds to a weight parameter, and each weight parameter is determined during the training process of the improved YOLO11 model.

[0030] In some embodiments, the original YOLO11 model includes a backbone network, a neck network, and a head network. In this application, the neck network in the original YOLO11 model is replaced with a bidirectional feature pyramid network that includes a normalized weight fusion mechanism, while the backbone network and the head network remain unchanged.

[0031] In some embodiments, the bidirectional feature pyramid network with normalized weight fusion mechanism includes three multi-scale feature fusion branches, each with a different resolution of the initial input feature map. The process involves inputting the second-finished engineered wood sample image into the improved YOLO11 model to obtain detection results, including: inputting the second-finished engineered wood sample image into the backbone network to obtain a first feature map, a second feature map, and a third feature map; using the first feature map as the initial input of the first multi-scale feature fusion branch, the second feature map as the initial input of the second multi-scale feature fusion branch, and the third feature map as the initial input of the third multi-scale feature fusion branch, and combining cross-branch fusion and normalized weight fusion mechanisms to fuse the first feature map, the second feature map, and the third feature map to obtain a target fused feature map; and inputting the target fused feature map into the head network to obtain the detection results.

[0032] Among them, the resolution of the first feature map is higher than that of the second feature map, and the resolution of the second feature map is higher than that of the third feature map.

[0033] In some embodiments, cross-branch fusion represents the use of feature maps at different scales (with different resolutions) as intermediate inputs to multi-scale feature fusion branches. For example, the first feature map serves as the initial input to the first multi-scale feature fusion branch, while the second and third feature maps serve as intermediate inputs. Similarly, the second feature map serves as the initial input to the second multi-scale feature fusion branch, while the first and third feature maps serve as intermediate inputs. Finally, the third feature map serves as the initial input to the third multi-scale feature fusion branch, while the first and second feature maps serve as intermediate inputs. This allows feature maps at different scales to be fused using a normalized weight fusion mechanism, resulting in richer feature information.

[0034] Furthermore, to address the issue of unequal contributions of features at different levels to the final detection, the bidirectional feature pyramid network with normalized weight fusion mechanism assigns a learnable weight to each input feature map. By training and improving the YOLO11 model, it can automatically learn the importance of features at different scales, allowing the model to optimize the loss of shallow feature maps, such as the localization error of small targets, on the basis of overall convergence. Without affecting the detection of large targets, it significantly enhances the learning ability for small targets. Moreover, the normalized weight fusion mechanism of the bidirectional feature pyramid network uses normalized weight parameters for feature fusion, requiring only simple scalar division. Compared with other feature fusion algorithms, it can reduce the computation by approximately 25%, greatly adapting to the real-time inference needs of edge computing nodes on the production line. The mathematical expression corresponding to the normalized weight fusion mechanism is shown in formula (2): (2); in, Represents the target fusion feature map. Indicates the first Feature maps of each input layer Indicates the first A learnable weight parameter, This represents the c-th learnable weight parameter. This indicates the number of feature maps in the input layer. This is the numerical stability constant, usually taken as 0.0001, to prevent the denominator from being 0.

[0035] It's important to note that the learnable weight parameters are randomly initialized by the network during training, and their update process consists of two phases: forward propagation and back propagation. During forward propagation, the learnable weight parameters, after activation and normalization, participate in the calculation of the detection loss. During back propagation, the gradient is propagated back from the total loss to the fusion node. This gradient represents the impact of changes in the current learnable weight parameters on the loss, and the optimizer automatically adjusts the size of the learnable weight parameters based on this gradient. If increasing a learnable weight parameter reduces the loss, the optimizer increases its value; conversely, it decreases it, thus giving greater weight to feature branches more useful for the detection task. Furthermore, the learnable weight parameters are merely numerical values, adding almost no increase to the number of parameters or computational cost. This is a key reason why improving the YOLO11 model can enhance object detection efficiency.

[0036] In some embodiments, each multi-scale feature fusion branch includes a first upsampling module, a first weighted feature fusion module, a first feature extraction module, a first feature concatenation module, a second feature extraction module, a downsampling module, a second weighted feature fusion module, a second feature concatenation module, a third feature extraction module, a second upsampling module, a third weighted feature fusion module, a third feature concatenation module, and a fourth feature extraction module; the processing of the first multi-scale feature fusion branch includes: The first feature map is input to the first upsampling module to obtain an upsampled feature map. The upsampled feature map is then weighted by the first weighted feature fusion module based on the first normalized weights corresponding to the first feature map, resulting in a first weighted feature map. Features are extracted from the first weighted feature map using the first feature extraction module. The extracted feature map and the second feature map are then input to the first feature concatenation module for concatenation. The feature map is then sequentially passed through the second feature extraction module and the downsampling module to obtain a first reference feature map. The first reference feature map is then weighted by the second weighted feature fusion module based on the second normalized weights corresponding to the second feature map, resulting in a second weighted feature map. Finally, the second weighted feature map, the upsampled feature map, and the first weighted feature map are combined. The images are fused to obtain a first fused feature map. The first fused feature map and the third feature map are then concatenated using a second feature stitching module and input into a third feature extraction module to obtain a second reference feature map. The second reference feature map is then input into a second upsampling module, and the sampled feature map is weighted by a third weighted feature fusion module based on the third normalized weight corresponding to the third feature map to obtain a third weighted feature map. The third weighted feature map, the first fused feature map, and the second reference feature map are then fused to obtain a second fused feature map. The second fused feature map and the third feature map are then concatenated using a third feature stitching module, and features are extracted using a fourth feature extraction module to obtain the feature map output by the first multi-scale feature fusion branch.

[0037] For example, the original YOLO11s model is as follows: Figure 3 As shown, C2PSA (Cross Stage Partial with Position-Sensitive Attention) represents the cross-stage positional connectivity and position-sensitive attention module, SPPF (Spatial Pyramid Pooling Fast) represents the fast spatial pyramid pooling module, C3K2 (Cross Stage Partial with 2 Convolutions) represents the cross-stage local connectivity module (containing two convolutions, supporting optional convolution kernel sizes), and CBS (Conv, BatchNorm, SiLU) represents convolution, batch normalization, and SiLU activation functions. This application replaces the neck network in the YOLO11s model with a bidirectional feature pyramid network that includes a normalized weight fusion mechanism. Figure 4 The diagram shows the structure of the multi-scale feature fusion branch of the bidirectional feature pyramid network with normalized weight fusion mechanism. The following uses this multi-scale feature fusion branch as the first multi-scale fusion branch to illustrate the data processing flow: The first feature map is input into the first upsampling module to obtain an upsampled feature map. The upsampled feature map is then weighted by the first weighted feature fusion module based on the first normalized weights corresponding to the first feature map, resulting in a first weighted feature map. Features are extracted from the first weighted feature map using the first C3K2 module (corresponding to the first feature extraction module). The extracted feature map and the second feature map are then input into the first Concat module (corresponding to the first feature concatenation module) for concatenation. The concatenation is then passed sequentially through the second C3K2 module (corresponding to the second feature extraction module) and the downsampling module to obtain a first reference feature map. The first reference feature map is then weighted by the second weighted feature fusion module based on the second normalized weights corresponding to the second feature map, resulting in a second weighted feature map. Finally, the second weighted feature map, the upsampled feature map, and the first weighted feature map are combined. The images are fused to obtain a first fused feature map. The first fused feature map and the third feature map are concatenated using the second feature concatenation module (the second Concat) and input into the third C3K2 (corresponding to the third feature extraction module) to obtain a second reference feature map. The second reference feature map is input into the second upsampling module (the second Upsample), and the upsampled feature map is weighted by the third weighted feature fusion module based on the third normalization weight corresponding to the third feature map to obtain a third weighted feature map. The third weighted feature map, the first fused feature map, and the second reference feature map are fused to obtain a second fused feature map. The second fused feature map and the third feature map are concatenated using the third feature concatenation module (the third Concat), and features are extracted by the fourth C3K2 (corresponding to the fourth feature extraction module) to obtain the feature map output by the first multi-scale feature fusion branch.

[0038] Understandably, the bidirectional feature pyramid network structure of the normalized weight fusion mechanism eliminates single input nodes with low contribution, adds cross-scale connections between the same level, and combines various multi-scale fusion branches to fuse features at different levels. This can enhance the reusability of features, reduce computational redundancy in the neck network, and more effectively fuse features at different scales, thereby reducing information loss.

[0039] S103. Based on the detection results, the corresponding real target detection labels, and the preset loss function, backpropagation training is performed on the improved YOLO11 model until the convergence condition is met, and the trained improved YOLO11 model is obtained.

[0040] In some embodiments, the preset loss function is constructed based on the distribution focus loss function, the classification loss function, and the localization loss function based on target scale weights. This can be expressed by formula (3): (3); in, and Let represent the distribution focus loss function, the classification loss function, and the localization loss function based on target scale weights, respectively. and These are the weighting factors for the three loss functions, with values ​​of 1, 0.5, and 7.5 respectively.

[0041] In some embodiments, the localization loss function is based on target scale weights. This can be expressed by formula (4): (4); in, , representing the target scale weight factor, As the adjustment coefficient, this application uses the experimental results for the dataset. It is 0.1. Used to control the rate of increase of weights; This indicates the area of ​​the current annotation box. This represents the area of ​​the smallest bounding box in the dataset. This represents the ratio of the intersection to the union of the predicted target detection bounding box and the labeled bounding box. This indicates the distance deviation between the center of the predicted bounding box and the center of the labeled bounding box. This represents the aspect ratio deviation, used to measure the difference in aspect ratio between the predicted target detection box and the labeled box. Indicates directional consistency deviation, used to measure the deviation of the directional trend between the predicted target detection box and the labeled box.

[0042] Understandable, It is a monotonically increasing exponential function, as the area of ​​the annotation box... The smaller the target scale weight factor, the closer its value is to 1, and the smaller the overall localization loss. Conversely, when the bounding box area *s* is larger, due to the inherent properties of the function *f(x) = exp(x)*, the target scale weight factor will increase rapidly, and the overall localization loss will be amplified. If the same loss weight is applied to all targets, the model will tend to focus on those targets that are numerous, large in area, and easier to optimize, while neglecting the accurate detection of small targets, resulting in poor detection performance for small targets. Therefore, this application introduces a target scale weight factor... A novel localization loss function based on target scale weights is constructed. Its core purpose is to solve the problem of target scale imbalance by assigning higher weights to the localization error of small targets. This makes it possible for even slight offsets in the bounding boxes of small targets to lead to a significant increase in localization loss, thereby forcing the improved YOLO11 model to prioritize optimizing the bounding boxes of small targets and thus improving the ability of the improved YOLO11 model to detect small targets.

[0043] The distributed focus loss function, classification loss function, and loss function are consistent with those used in the original YOLO11 model. Distributed focus loss function This can be expressed by formula (5): (5); in, This represents the total number of valid foreground samples (only grids / anchors containing the target are included in the loss calculation). Represents a discrete distance interval index. ∈[0,1,2,…,reg_max], where reg_max represents the maximum distance quantization upper limit (YOLO11 defaults to reg_max=16, which discretizes the boundary distance of the bounding box into 17 intervals). The discrete target distribution representing the true distance (a one-hot type distribution label obtained by piecewise linear interpolation of the true boundary distance, with non-zero weights only in the intervals near the true distance). The model predicts the first... The sample, the first The Softmax normalized probability of each distance interval (the network output is obtained as a discrete distribution after softmax). This represents the maximum probability value of the predicted distribution for the l-th sample. Indicates the focus coefficient (YOLO11 default). =0.5).

[0044] Classification loss function This can be expressed by formula (6): (6); in, This represents the total number of categories in the dataset. The model predicts the first... The sample belongs to the first The Sigmoid normalized probability of the class. This indicates a quality fusion label. This indicates that a category indicator function exists, and the true category of the sample is... The background category is set to 1, and the background category to 0 for all other background categories. Indicates the first Intersection over Union (IoU) of predicted bounding boxes and ground truth bounding boxes for each sample.

[0045] In some embodiments, during the training of the improved YOLO11 model, the detection loss value is calculated based on the detection results of the improved YOLO11 model for the second-faced artificial board sample and the corresponding real target detection label, and the preset loss function is combined with formula (3). The network parameters of the improved YOLO11 model are updated by gradient descent until the convergence condition is met. If the detection loss value is less than the detection loss threshold or the number of iterations reaches the preset number, the training can be terminated and the trained improved YOLO11 model is obtained.

[0046] S104. Input the image of the decorative panel to be detected into the trained improved YOLO11 model to obtain the target detection result of the image of the decorative panel to be detected.

[0047] In some embodiments, the image of the veneer panel to be detected can be a real-time acquired image of the veneer panel. The trained improved YOLO11 model can detect defects in the image of the veneer panel to be detected, thereby obtaining the target detection result of the image of the veneer panel to be detected.

[0048] In this embodiment, an adaptive Mosaic enhancement technique based on bounding box location information is used to process the first veneer panel sample image to obtain a second veneer panel sample image. The neck network in the original YOLO11 model is replaced by a single top-down path aggregation network with a bidirectional feature pyramid network containing a normalized weight fusion mechanism to obtain an improved YOLO11 model. The second veneer panel sample image is input into the improved YOLO11 model to obtain the detection result. Based on the detection result, the corresponding real target detection label, and the preset loss function, the improved YOLO11 model is backpropagated and trained until the convergence condition is met to obtain the trained improved YOLO11 model. The veneer panel image to be detected is input into the trained improved YOLO11 model to obtain the target detection result of the veneer panel image to be detected. Thus, the adaptive Mosaic enhancement technique uses the bounding box position information as a kind of "pseudo-attention," which ensures that the stitching center falls in the area with dense defect targets. This results in the enhanced second-finish artificial board sample image containing multiple effective targets, increasing the effective sample size during training and improving the sensitivity of the YOLO11 model to low-contrast defects, thereby improving the target detection accuracy of the improved YOLO11 model. By replacing the neck network in the original YOLO11 model with a single top-down path aggregation network and a bidirectional feature pyramid network with a normalized weight fusion mechanism, the improved YOLO11 model can simultaneously capture high-level abstract semantics and low-level fine details, effectively fusing features at different scales, reducing information loss, thereby improving the feature extraction capability for small defects, reducing false negative and false positive rates. Furthermore, the normalized weight fusion mechanism only needs to handle simple scalar division, with almost no increase in the number of parameters and computational load, thus improving both the accuracy and efficiency of target detection.

[0049] The following describes the experimental process for the surface defect detection method for decorative engineered wood panels based on the improved YOLO11 model provided in this application.

[0050] To verify the effectiveness of the improved YOLO11 model in handling complex wood grain backgrounds and multi-scale defects, this application constructed a high-performance graphics workstation to train and test the improved YOLO11 model.

[0051] In terms of hardware, the central processing unit (CPU) and graphics processing unit (GPU) work asynchronously to support data flow and logical operations. The core processor used in this study is an Intel(R) Core(TM) i9-14900K, which possesses excellent multi-threaded data processing capabilities. The graphics processing unit is configured with an NVIDIA GeForce RTX 4090 (24GB), whose powerful parallel computing capabilities and ample video memory provide physical guarantees for efficient model training and complex feature fusion. In terms of software environment, the system is based on a 64-bit Windows 11 operating system, building a deep learning development environment. Python 3.10 is used as the programming language, and the deep learning framework chosen is PyTorch 2.1.0, coupled with the CUDA 12.1 acceleration library to fully utilize the GPU's hardware performance.

[0052] Based on the computing power of the aforementioned hardware platform, this study set the training parameters for the detection model. The specific training parameters are shown in Table 1: Table 1 Training Parameter Settings

[0053] The image resolution of the input model is uniformly set to 512×512. Larger sizes preserve image details better, but increase computational power and time exponentially; conversely, smaller sizes cause small defects to disappear during downsampling. This size allows for the preservation of subtle crack features while fully utilizing the parallel computing power of the GPU, avoiding information loss due to excessively rapid downsampling.

[0054] Batch size refers to the number of samples used in a single gradient update, and it is responsible for balancing training stability and GPU memory bandwidth utilization. A larger batch size can provide more stable gradients and accelerate training, but it requires more powerful GPU memory. A smaller batch size, such as the 16 set in this application, although it introduces some gradient noise, can sometimes help the detection model escape local optima and perform better in complex texture recognition.

[0055] The number of iterations refers to the number of times the model fully learns the entire dataset, directly determining the sufficiency of the model's learning. Too few iterations will cause the model to stop learning before fully mastering the defect features, leading to underfitting; too many iterations will limit the model's understanding of defect features to the training set, losing its ability to process new materials, leading to overfitting. 300 iterations is usually a balance point to ensure convergence.

[0056] The initial learning rate is arguably the most important hyperparameter in model training. It represents the minimum step size for updating model parameters and controls the rate at which the loss function decreases. An excessively high learning rate causes the model's parameters and weights to oscillate around the optimal solution, while an excessively low learning rate significantly reduces training speed and may even prevent the model from escaping local optima. Considering both model convergence speed and feature complexity, this application sets the initial learning rate to 0.01.

[0057] First, the impact of adaptive Mosaic image enhancement techniques based on bounding box location information on model performance is evaluated. This experiment compares the effect of using only basic image enhancement (mirroring, random rotation, HSV color space dithering) with the effect of combining adaptive Mosaic image enhancement techniques. The experimental results are shown in Table 2. Table 2. Detection results of image enhancement experiment

[0058] As shown in Table 2, applying adaptive Mosaic augmentation based on bounding box location information to the dataset improved the model's precision and recall by 1.8% and 1.9%, respectively, and increased the mAP@0.5 and mAP@0.5:0.95 values ​​by 2.7% and 3.2%, respectively. However, there was no increase in floating-point computation or inference speed. This indicates that applying adaptive Mosaic augmentation based on bounding box location information to the dataset can improve the model's performance in small object detection without consuming more resources.

[0059] Then, the target scale weight factor in the localization loss function is determined through experiments. ,right Experiments were conducted with values ​​of 0.01, 0.1, 1, 2, and 10 respectively. Table 3 shows the statistical results of mAP@0.5 under different values. As can be seen from Table 3, when... At that time, the mAP@0.5 value was the largest (95.3), therefore this application will The value is 0.1.

[0060] Table 3. Correspondence between weighting factors for different target scales and mAP@0.5 values

[0061] Furthermore, ablation experiments were used to analyze the impact of the bidirectional feature pyramid network structure, adaptive Mosaic enhancement, and scale weight factor on the model's detection performance. This application uses YOLO11s as a base and introduces at least one of the following: bidirectional feature pyramid network structure, target scale weight factor, and adaptive Mosaic enhancement. The results are compared with the base model, and the ablation experiment results are shown in Table 4. Table 4 Ablation Experiment Results

[0062] As shown in Table 4, adaptive Mosaic enhancement, bidirectional feature pyramid network structure, and target scale weight factor can all improve the detection performance of YOLO11s for small targets to a certain extent. Comparing Group 1 and Group 8, it can be seen that by simultaneously applying adaptive Mosaic enhancement, bidirectional feature pyramid network structure, and target scale weight factor to the model, the precision and recall of the model increased by 6.8% and 6.1%, respectively; the mAP@0.5 value and mAP@0.5:0.95 value increased by 10.0% and 13.0%, respectively; and the floating-point operation and inference speed increased by 20.8% and 15.6%, respectively. This indicates that these three improvement measures can significantly improve the model's performance in detecting small targets. Comparing groups 6 and 8, replacing the original FPN module with a bidirectional feature pyramid network structure in the model's neck part improved the model's precision and recall by 5.7% and 5.1%, respectively. The mAP@0.5 value and mAP@0.5:0.95 value increased by 7.9% and 11.6%, respectively. Floating-point computation and inference speed also increased by 15.1% and 13.3%, respectively. This indicates that introducing the bidirectional feature pyramid network structure can improve the model's performance in small object detection to some extent. Comparing groups 7 and 8, introducing a target scale weight factor into the model's loss function improved the model's precision and recall by 2.8% and 3.1%, respectively. The mAP@0.5 value and mAP@0.5:0.95 value increased by 3.5% and 5.5%, respectively. Floating-point computation increased by 1.6%, while inference speed did not increase significantly. This indicates that introducing a target scale weight factor into the model's loss function can improve the model's performance in small object detection with a slight increase in resource consumption. Comparing the data from groups 5, 6, and 7, we can conclude that when detecting small targets, the three improvement measures enhance the model's detection performance as follows: the bidirectional feature pyramid network structure is superior to the target scale weight factor, and the target scale weight factor is superior to the adaptive Mosaic enhancement. In terms of improving model detection performance and reducing resource consumption, the target scale weight factor is better than the bidirectional feature pyramid network structure, which in turn is better than the adaptive Mosaic enhancement.

[0063] To verify the performance difference of the YOLO11 model before and after the improvement, the object detection results of the existing YOLO11s and the improved YOLO11 model provided in this application were compared. The detection results of the two models were statistically analyzed, and a confusion matrix was plotted with the true type on the x-axis and the predicted type on the y-axis, as shown below. Figure 5 As shown, Figure 5 The left side shows the confusion matrix plotted based on the target detection results obtained from the improved YOLO11 model provided in this application. Figure 5 The right side shows the confusion matrix plotted based on the target detection results obtained from the existing YOLO11s model. The confusion matrix reveals that the improved YOLO11 model significantly improves detection performance compared to YOLO11s, particularly reducing the misclassification of background as dirt. Experimental results show that the improved YOLO11 model correctly detected 89 dry flowers, 392 wet flowers, dirt, and 250 cracks, respectively, representing increases of 4.1%, 9.1%, and 5.8% compared to the unimproved YOLO11s model. This indicates that the improved YOLO11 model significantly improves the detection performance of small targets, especially dirt.

[0064] like Figure 6 The image shown illustrates the detection results of YOLO11s and the improved YOLO11 model provided in this application for a single image (the text above the blue detection box in the image indicates the defect name and detection confidence). Figure 6 As can be seen, YOLO11s exhibits some false negatives when detecting small target defects, failing to identify all defects completely. In contrast, the improved YOLO11 model can more accurately identify target defects under the same detection conditions, demonstrating higher reliability. Therefore, the surface defect detection method for decorative engineered wood panels proposed in this application, based on the improved YOLO11 model, can effectively reduce the false detection rate of the original model, improve the detection rate, and demonstrate stronger robustness in complex environments.

[0065] Understandably, this application effectively enhances the model's ability to detect small targets and improves its generalization and robustness through three measures: Mosaic enhancement of the dataset, replacement of the original FPN module in the YOLO11 neck network with a bidirectional feature pyramid network, and the introduction of a scale weight factor into the loss function. Furthermore, experimental results show that the improved YOLO11 model's inference speed meets the requirements for real-time industrial detection and can be used for online detection of common defects such as dried and wet flowers, stains, and cracks in the production of decorative engineered wood panels, providing technical support for the high-quality development of home building materials.

[0066] It should be noted that, depending on the implementation needs, the steps described in the embodiments of this application can be broken down into more steps, or two or more steps or parts of the steps can be combined into new steps to achieve the purpose of the embodiments of this application.

[0067] The methods described in the embodiments of this application can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code downloaded over a network that is originally stored in a remote recording medium or a non-transitory machine-readable medium and will be stored in a local recording medium. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for executing the methods shown herein.

[0068] Those skilled in the art will recognize that the units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this application.

[0069] The above embodiments are only used to illustrate the embodiments of this application, and are not intended to limit the embodiments of this application. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the embodiments of this application. Therefore, all equivalent technical solutions also fall within the scope of the embodiments of this application, and the patent protection scope of the embodiments of this application should be defined by the claims.

Claims

1. A veneered panel surface defect detection method based on an improved YOLO11 model, characterized in that, include: The first decorative panel sample image was processed using an adaptive Mosaic enhancement technique based on the position information of the annotation box to obtain the second decorative panel sample image. The neck network in the original YOLO11 model was replaced by a single top-down path aggregation network with a bidirectional feature pyramid network that includes a normalized weight fusion mechanism, resulting in an improved YOLO11 model. The sample image of the second veneer artificial board was then input into the improved YOLO11 model to obtain the detection results. The normalized weight fusion mechanism uses normalized weight parameters to perform weighted fusion of feature maps at different scales. Each scale of feature map corresponds to a weight parameter, and each weight parameter is determined during the training process of the improved YOLO11 model. Based on the detection results, the corresponding real target detection labels and the preset loss function, the improved YOLO11 model is backpropagated and trained until the convergence condition is met, and the trained improved YOLO11 model is obtained. The image of the veneer panel to be detected is input into the trained improved YOLO11 model to obtain the target detection result of the veneer panel image.

2. The method of claim 1, wherein, The first decorative panel sample image includes multiple images; the processing of the first decorative panel sample image using adaptive Mosaic enhancement technology based on bounding box position information includes: Each sample image of the first-faced artificial board is enlarged and cropped at a random position to obtain a sub-image; Based on the geometric center coordinates of each defect annotation box in the four randomly selected sub-images, as well as the width and height of the defect annotation boxes, calculate the theoretical center coordinates of the set of all defect annotation boxes in the four sub-images. Using the theoretical center coordinates as the splicing center, four randomly selected sub-images are cropped and spliced ​​to obtain the sample image of the second decorative artificial board.

3. The method according to claim 2, characterized in that, Theoretical center coordinates of all defect labeling box sets in the 4 intercepted subgraphs The calculation formula is represented as: ； wherein, , respectively represent the x-coordinate and y-coordinate of the center of the bounding box in the i-th intercepted sub-graph, , respectively represent the width and height of the bounding box in the i-th intercepted sub-graph, is a random coordinate offset, and the value range is [1, 256].​​​​ 4. The method of claim 1, wherein, The improved YOLO11 model includes a backbone network, a bidirectional feature pyramid network with a normalized weight fusion mechanism, and a head network. The bidirectional feature pyramid network with the normalized weight fusion mechanism includes three multi-scale feature fusion branches, each with a different resolution of the initial input feature map. The step of inputting the sample image of the second decorative artificial board into the improved YOLO11 model to obtain the detection results includes: The sample image of the second decorative artificial board is input into the backbone network to obtain the first feature map, the second feature map and the third feature map; the resolution of the first feature map is higher than that of the second feature map, and the resolution of the second feature map is higher than that of the third feature map. The first feature map is used as the initial input of the first multi-scale feature fusion branch, the second feature map is used as the initial input of the second multi-scale feature fusion branch, and the third feature map is used as the initial input of the third multi-scale feature fusion branch. The first feature map, the second feature map and the third feature map are fused by combining cross-branch fusion and normalized weight fusion mechanisms to obtain the target fused feature map. The target fusion feature map is input into the head network to obtain the detection result.

5. The method of claim 3, wherein, Each multi-scale feature fusion branch includes a first upsampling module, a first weighted feature fusion module, a first feature extraction module, a first feature concatenation module, a second feature extraction module, a downsampling module, a second weighted feature fusion module, a second feature concatenation module, a third feature extraction module, a second upsampling module, a third weighted feature fusion module, a third feature concatenation module, and a fourth feature extraction module; The processing steps of the first multi-scale feature fusion branch include: The first feature map is input into the first upsampling module to obtain the upsampled feature map, and the upsampled feature map is weighted by the first weighted feature fusion module based on the first normalized weight corresponding to the first feature map to obtain the first weighted feature map; The first feature extraction module is used to extract features from the first weighted feature map. The extracted feature map and the second feature map are then input into the first feature concatenation module for concatenation. The feature map is then passed through the second feature extraction module and the downsampling module in sequence to obtain the first reference feature map. The second weighted feature fusion module weights the first reference feature map based on the second normalized weight corresponding to the second feature map to obtain the second weighted feature map. The second weighted feature map, the upsampled feature map and the first weighted feature map are then fused to obtain the first fused feature map. The first fused feature map and the third feature map are stitched together using the second feature stitching module and then input into the third feature extraction module to obtain the second reference feature map; The second reference feature map is input into the second upsampling module. The third weighted feature fusion module weights the sampled feature map based on the third normalized weight corresponding to the third feature map to obtain the third weighted feature map. The third weighted feature map, the first fusion feature map and the second reference feature map are then fused to obtain the second fusion feature map. The second fused feature map and the third feature map are concatenated using the third feature concatenation module, and features are extracted using the fourth feature extraction module to obtain the feature map output by the first multi-scale feature fusion branch.

6. The method according to claim 1, characterized in that, The preset loss function is constructed based on a distribution focal loss function, a classification loss function, and a positioning loss function based on a target scale weight is represented as: ； wherein, and denote the distribution focus loss function, the classification loss function and the positioning loss function based on the target scale weight, respectively, and are weight factors of the three loss functions, and the values are 1, 0.5 and 7.5, respectively.

7. The method according to claim 6, characterized in that, The target scale weight-based positioning loss function is represented as: ； in, , representing the target scale weight factor, This is an adjustment coefficient used to control the rate of increase in weight. This indicates the area of ​​the current annotation box. This represents the area of ​​the smallest bounding box in the dataset. This represents the ratio of the intersection to the union of the predicted target detection bounding box and the labeled bounding box. This indicates the distance deviation between the center of the predicted bounding box and the center of the labeled bounding box. This represents the aspect ratio deviation, used to measure the difference in aspect ratio between the predicted target detection box and the labeled box. Indicates directional consistency deviation, used to measure the deviation of the directional trend between the predicted target detection box and the labeled box.

Images