Hot-pressing light guide plate surface defect detection method based on mixed feature extraction

Abstract

The invention provides a hot-pressing light guide plate surface defect detection method based on mixed feature extraction, and the method comprises the following steps: constructing a hot-pressing light guide plate defect detection model based on a YOLOv11 network architecture, and embedding a parallel texture fusion module used for collaborative modeling of local texture features and global context features in a backbone network unit, a cavity block re-parameter module for expanding a receptive field through a cavity convolution and re-parameterization mechanism is embedded in the neck network unit, and the detection head unit comprises a plurality of re-parameterization detail enhancement detection heads adopting shared convolution and differential convolution; obtaining a surface image of the hot-pressing light guide plate, constructing a defect data set, and training the model; and inputting a to-be-detected hot-pressing light guide plate image into the trained hot-pressing light guide plate defect detection model, and outputting a defect detection result. According to the method, the model calculation cost and parameter quantity are greatly reduced while accurate identification of various complex defects of the hot-pressing light guide plate is ensured.

Description

Technical Field

[0001] This invention relates to a method for detecting surface defects in a hot-pressed light guide plate, specifically a method for detecting surface defects in a hot-pressed light guide plate based on hybrid feature extraction, belonging to the field of industrial machine vision inspection technology. Background Technology

[0002] Light guide plates are core optical components in backlight modules, widely used in LCD displays, laptops, and automotive displays. Hot-pressed light guide plates are manufactured by hot-pressing a fine dot structure onto the surface of acrylic sheets to disrupt total internal reflection, thus converting side-incident light sources into uniform surface light sources. During the hot-pressing process, factors such as fluctuations in process parameters, mold wear, or environmental contaminants can easily cause various defects on the surface of the light guide plate, including scratches, missing etch marks, black / white spots, dark patches, and shadows. These defects not only affect optical performance but can also lead to uneven brightness or visual flaws in the final display product. Therefore, efficient and accurate quality inspection of hot-pressed light guide plates before they leave the factory has become a crucial step in ensuring product reliability.

[0003] However, the dense distribution of light guide dots in hot-pressed light guide plate images creates strong background noise due to the regular texture. Minor defects such as shallow scratches and small stains are often obscured by these dot textures, making it difficult for existing machine vision inspection algorithms to distinguish between dots and defects, resulting in a very high false detection rate. Furthermore, defects vary greatly in scale and shape, ranging from large shadows or clumps of dirt to tiny, pixel-level scratches or punctures. Traditional image processing algorithms often struggle to balance local details with global semantics; many defects exhibit extremely low contrast with the background and blurred edges, making it difficult for conventional convolutional neural network models to extract effective boundary features. In addition, industrial production lines typically operate at high speeds, requiring inspection models with extremely high inference speeds. However, to improve accuracy, existing technologies often continuously increase the depth and width of inspection models, leading to a surge in model parameters and making real-time deployment on edge computing devices difficult. Summary of the Invention

[0004] Based on the above background, the purpose of this invention is to provide a method for detecting surface defects of hot-pressed light guide plates based on hybrid feature extraction. When detecting surface defects of hot-pressed light guide plates, this method can effectively suppress texture background interference, has enhanced multi-scale feature extraction capabilities and lightweight model characteristics, thereby solving the technical problems of insufficient feature extraction capabilities, low sensitivity to low-contrast defects, and excessive computational resource consumption of the detection model in the prior art when using general target detection algorithms for detecting surface defects of hot-pressed light guide plates.

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

[0006] A method for detecting surface defects in a hot-pressed light guide plate based on hybrid feature extraction, the method comprising the following steps:

[0007] A defect detection model for hot-pressed light guide plates is constructed based on the YOLOv11 network architecture. The hot-pressed light guide plate defect detection model includes a backbone network unit, a neck network unit, and a detection head unit. The backbone network unit embeds a parallel texture fusion module for collaboratively modeling local texture features and global context features. The neck network unit embeds a dilated block reparameterization module for expanding the receptive field through dilated convolution and reparameterization mechanisms. The detection head unit includes multiple reparameterized detail enhancement detection heads that employ shared convolution and differential convolution.

[0008] A surface image of the hot-pressed light guide plate is acquired and a defect dataset is constructed. The defect dataset is then used to train a defect detection model for the hot-pressed light guide plate.

[0009] The image of the hot-pressed light guide plate to be detected is input into the trained hot-pressed light guide plate defect detection model, and the defect detection result is output.

[0010] Preferably, the construction steps of the parallel texture fusion module specifically include:

[0011] The input feature map is divided into a first branch and a second branch in the channel dimension. The first branch is used to extract local texture features through convolution operations, and the second branch is used to capture global long-range dependency information through the MHSA-CGLU submodule.

[0012] The output features of the first branch and the output features of the second branch are concatenated and fused, and then channel integration is performed through a convolutional layer before output.

[0013] Preferably, the MHSA-CGLU submodule includes a first two-dimensional normalization layer, a multi-head self-attention mechanism layer, a first random path drop-off layer, a second two-dimensional normalization layer, a convolutional gated linear unit layer, and a second random path drop-off layer connected in sequence; wherein, the convolutional gated linear unit layer includes a parallel main convolutional branch and a gated branch, the gated branch uses the Sigmoid function to generate a weight map, and multiplies it element-wise with the output of the main convolutional branch to achieve dynamic feature selection.

[0014] Preferably, the void block weight parameter module adopts a staged multi-scale context modeling strategy, and the construction steps of the void block weight parameter module specifically include:

[0015] In the stage of constructing regional residual modeling, the original input features are downsampled and basic semantics are extracted through 3×3 convolution, batch normalization and SiLU activation function;

[0016] A semantic-level residual enhancement stage is constructed. The output features of the region-level residual modeling stage are further processed through three parallel convolutional paths. The first convolutional path is a standard convolution with an expansion rate of 1, the second convolutional path is a dilated reparameter block with an expansion rate of 2 and a kernel size of 5×5, and the third convolutional path is a dilated reparameter block with an expansion rate of 3 and a kernel size of 7×7.

[0017] The output features of the three convolutional paths are fused, and then added to the original input features through residual connections before output.

[0018] Preferably, the construction steps of the reparameterized detail enhancement detection head specifically include:

[0019] For the multi-scale feature maps from the neck network unit, channel adaptation is performed by 1×1 convolution and group normalization, respectively;

[0020] The channel-adapted feature map is input into a two-layer cascaded shared detail enhancement convolutional module, and features are extracted through convolutional kernels with shared weights.

[0021] The output of the shared detail enhancement convolution module is fed into the bounding box regression branch and the object classification branch, respectively.

[0022] Preferably, the shared detail enhancement convolutional module includes parallel standard convolutional paths, center difference convolutional paths, corner difference convolutional paths, horizontal difference convolutional paths, and vertical difference convolutional paths; during the model inference stage, the parallel convolutional paths of the shared detail enhancement convolutional module can be reparameterized into single-path standard convolutions.

[0023] Preferably, the backbone network unit includes two convolutional modules, a first C3k2 module, a third convolutional module, a second C3k2 module, a fourth convolutional module, a first parallel texture fusion module, a fifth convolutional module, a second parallel texture fusion module, a spatial pyramid pooling module, and a C2PSA attention module connected in sequence; wherein, the output of the second C3k2 module, the output of the first parallel texture fusion module, and the output of the C2PSA attention module are respectively connected to the neck network unit as three feature outputs of the backbone network unit.

[0024] Preferably, the neck network unit includes multiple upsampling modules, a splicing module, a downsampling convolution module, and a hole block weight parameter module;

[0025] The output of the C2PSA attention module of the backbone network unit is connected to the first upsampling module, and then connected to the output of the first parallel texture fusion module of the backbone network unit through the first splicing module, and input to the first hole block weight parameter module.

[0026] The output of the first void block weight parameter module is connected to the second upsampling module, and then connected to the output of the second C3k2 module of the backbone network unit through the second splicing module, and input to the second void block weight parameter module;

[0027] The output of the second hole block re-parameter module is connected to the first downsampling convolution module, and then connected to the output of the first hole block re-parameter module through the third splicing module, and input to the third hole block re-parameter module.

[0028] The output of the third dilated block reparameter module is connected to the second downsampling convolution module, and then connected to the output of the C2PSA attention module of the backbone network unit through the fourth splicing module, and input to the fourth dilated block reparameter module.

[0029] The outputs of the second void block reparameter module, the third void block reparameter module, and the fourth void block reparameter module are respectively connected to three different reparameterized detail enhancement detection heads.

[0030] Preferably, when training the hot-pressed light guide plate defect detection model, the hyperparameters of the model training are set as follows: 300 training rounds, batch size of 32, stochastic gradient descent as the optimizer, initial learning rate of 0.01, cosine annealing learning rate of 0.1, weight decay coefficient of 0.005, and dynamic momentum parameter of 0.937.

[0031] Compared with the prior art, the present invention has the following advantages:

[0032] This invention provides a method for detecting surface defects in hot-pressed light guide plates based on hybrid feature extraction. It offers an end-to-end industrial vision inspection solution that balances high detection accuracy and high inference efficiency. By deeply customizing the backbone network unit, neck network unit, and detection head unit based on the YOLOv11 architecture, it organically integrates hybrid modeling of local texture and global semantics, a dilated convolution mechanism with a large receptive field, and edge detail enhancement technology based on differential operators to collaboratively optimize the target detection model. This overcomes the technical bottlenecks of traditional detection models, such as the difficulty in feature extraction against the strong texture background of hot-pressed light guide plates and the easy omission of small and low-contrast defects. Furthermore, it achieves lightweight detection model through shared convolution and reparameterization strategies, thereby significantly reducing model computation cost and parameter quantity while ensuring accurate identification of various complex defects in hot-pressed light guide plates.

[0033] The parallel texture fusion module introduced in the backbone network unit of this invention extracts local texture detail features and captures global long-distance dependency information through parallel convolution branches and Transformer branches. This hybrid feature extraction mechanism enables the detection model to identify both small local defects and understand the global background distribution, thereby effectively distinguishing between normal dot textures and abnormal defects on the surface of the hot-pressed light guide plate and significantly reducing the false detection rate caused by complex background textures.

[0034] The invention designs a dilated block reparameterization module in the neck network unit. By introducing dilated convolutions with different dilation rates, it can significantly expand the receptive field without reducing the feature map resolution and capture multi-scale contextual information. Combined with the reparameterization mechanism, it can improve the detection accuracy of extreme-scale defects such as large-area shadows and micro-scratches without increasing inference time.

[0035] The reparameterized detail enhancement detection head used in this invention integrates multiple differential convolution operators such as central difference convolution and corner difference convolution, which can keenly capture subtle changes in pixel gradients in the image, thereby significantly enhancing the ability to extract the edge contours of low-contrast defects. Multiple scale detection heads share the same set of convolution kernel weights, greatly reducing the number of parameters in the model. Attached Figure Description

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

[0037] Figure 1 This is a schematic diagram of the architecture of the hot-pressed light guide plate defect detection model (PDR-YOLOv11) built based on the YOLOv11 network architecture in this invention.

[0038] Figure 2 This is a schematic diagram of the architecture of the Parallel Texture Blending Module (PTFM) in this invention;

[0039] Figure 3 This is a schematic diagram of the architecture of the void block weight parameter module (DBR-C3k2) in this invention;

[0040] Figure 4 This is a schematic diagram of the architecture of the Reparameterized Detail Enhancement Detection Head (Rep-DEHead) in this invention;

[0041] Figure 5 This is a schematic diagram of the architecture of the shared detail enhancement convolutional module (DEConv) in this invention;

[0042] Figure 6 This is a schematic diagram of the surface defect types of the hot-pressed light guide plate in this invention, where a shows a line scratch defect, b shows a missing engraving defect, c shows a black shadow defect, d shows a white stain defect, e shows a dark spot defect, f shows a clump of stain defect, and g shows a white shadow defect.

[0043] Figure 7 This is a comparison chart of the test results of the hot-pressed light guide plate defect detection model in this invention with several mainstream target detection models on the self-built LGPs defect dataset;

[0044] Figure 8 This is a comparison chart of the visualization results output of the hot-pressed light guide plate defect detection model and the YOLOv11 detection model in this invention;

[0045] Figure 9 This is a comparison chart of the test results of the hot-pressed light guide plate defect detection model in this invention with several mainstream target detection models on the railway defect detection dataset. Detailed Implementation

[0046] The technical solution of the present invention will be further described in detail below through specific embodiments and in conjunction with the accompanying drawings. It should be understood that the implementation of the present invention is not limited to the following embodiments, and any modifications and / or alterations made to the present invention will fall within the protection scope of the present invention.

[0047] In this invention, unless otherwise specified, all parts and percentages are by weight, and the equipment and raw materials used are commercially available or commonly used in the art. Unless otherwise specified, the methods in the following embodiments are conventional methods in the art. Unless otherwise specified, the components or equipment in the following embodiments are general standard parts or components known to those skilled in the art, and their structures and principles can be learned by those skilled in the art through technical manuals or conventional experimental methods.

[0048] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings. In this detailed description, numerous specific details are set forth to facilitate explanation and provide a thorough understanding of the embodiments of the present invention. However, one or more embodiments may be practiced by those skilled in the art without these specific details.

[0049] An embodiment of the present invention discloses a method for detecting surface defects of a hot-pressed light guide plate based on hybrid feature extraction. The method includes the following steps:

[0050] A defect detection model for hot-pressed light guide plates is constructed based on the YOLOv11 network architecture. The hot-pressed light guide plate defect detection model includes a backbone network unit, a neck network unit, and a detection head unit. The backbone network unit embeds a parallel texture fusion module for collaboratively modeling local texture features and global context features. The neck network unit embeds a dilated block reparameterization module for expanding the receptive field through dilated convolution and reparameterization mechanisms. The detection head unit includes multiple reparameterized detail enhancement detection heads that employ shared convolution and differential convolution.

[0051] Surface images of the hot-pressed light guide plate were acquired and a defect dataset was constructed. The defect dataset was then used to train a defect detection model for the hot-pressed light guide plate.

[0052] The image of the hot-pressed light guide plate to be detected is input into the trained hot-pressed light guide plate defect detection model, and the defect detection result is output.

[0053] YOLOv11 is the eleventh generation model in the YOLO series. Through improved architecture design and training strategies, it achieves a good balance between detection performance and computational efficiency. However, it still has the following shortcomings in industrial scenarios such as hot-pressed light guide plate defect detection: insufficient sensitivity to small targets and defects with blurred boundaries, limited robustness to changes in lighting and complex textures, and room for improvement in feature fusion and real-time performance.

[0054] This invention improves upon the YOLOv11 network architecture, constructing the PDR-YOLOv11 defect detection model for hot-pressed light guide plates to enhance small defect detection capabilities, improve multi-scale feature fusion effects, and increase computational efficiency. Specific improvements are as follows: A parallel texture fusion module (PTFM) is introduced into the backbone network unit to enhance the extraction of multi-scale textures and edge details; based on the DWR module, the DRB structure from UniRepLKNet is introduced, and the C3k2 module is optimized through fusion, thereby embedding a dilated block reparameterized module (DBR-C3k2) into the neck network unit. Through dilated convolution and reparameterization design, the receptive field and expressive power are enhanced, improving the performance of small target defect detection under complex textures; the reparameterized detail enhancement detection head (Rep-DEHead) can effectively fuse multi-scale semantic information and detail features, significantly improving defect detection accuracy while maintaining a low parameter count.

[0055] like Figure 1As shown, the backbone network unit of the PDR-YOLOv11 hot-pressed light guide plate defect detection model includes two convolutional modules, a first C3k2 module, a third convolutional module, a second C3k2 module, a fourth convolutional module, a first parallel texture fusion module, a fifth convolutional module, a second parallel texture fusion module, a spatial pyramid pooling module, and a C2PSA attention module connected in sequence. The outputs of the second C3k2 module, the first parallel texture fusion module, and the C2PSA attention module serve as the three feature outputs of the backbone network unit, which are then connected to the neck network unit.

[0056] The neck network unit comprises multiple upsampling modules, a stitching module, a downsampling convolutional module, and a dilated block reparameter module. The output of the C2PSA attention module of the backbone network unit is connected to the first upsampling module, then connected to the output of the first parallel texture fusion module of the backbone network unit via the first stitching module, and input to the first dilated block reparameter module. The output of the first dilated block reparameter module is connected to the second upsampling module, then connected to the output of the second C3k2 module of the backbone network unit via the second stitching module, and input to the second dilated block reparameter module. The output of the second dilated block reparameter module is connected to the first downsampling convolutional module, then connected to the output of the first dilated block reparameter module via the third stitching module, and input to the third dilated block reparameter module. The output of the third dilated block reparameter module is connected to the second downsampling convolutional module, then connected to the output of the C2PSA attention module of the backbone network unit via the fourth stitching module, and input to the fourth dilated block reparameter module.

[0057] The detection head unit includes three reparameterized detail enhancement detection heads. The outputs of the second, third, and fourth hole block reparameterized modules in the neck network unit are respectively connected to the three different reparameterized detail enhancement detection heads.

[0058] The following sections provide detailed explanations of the Parallel Texture Blending Module (PTFM), the Hole Block Reparameterization Module (DBR-C3k2), and the Reparameterized Detail Enhancement Detection Head (Rep-DEHead).

[0059] PTFM achieves collaborative modeling of local and global features by fusing CNNs and Transformers through a parallel architecture. It employs a channel-level hybrid strategy, introducing Transformers only in certain channels to enhance global modeling capabilities while controlling computational overhead. PTFM divides the input feature map into two paths: one convolutional path extracts local texture, while the other captures global context through multi-head self-attention (MHSA), ultimately fusing them to generate discriminative features. Figure 2As shown, the specific steps for building PTFM include:

[0060] The input feature map is divided into a first branch and a second branch in the channel dimension. The first branch is used to extract local texture features through convolution operations, and the second branch is used to capture global long-range dependency information through the MHSA-CGLU submodule.

[0061] The output features of the first branch and the output features of the second branch are concatenated and fused, and then channel integration is performed through a convolutional layer before output.

[0062] The MHSA-CGLU submodule comprises a first two-dimensional normalization layer (LayerNorm2d), a multi-head self-attention (MHSA) layer, a first droppath layer, a second two-dimensional normalization layer, a convolutional gated linear unit (CGLU) layer, and a second droppath layer, connected sequentially. This module aims to integrate global modeling capabilities with local nonlinear representation capabilities, thereby enhancing the model's ability to extract and generalize complex features.

[0063] First, the input feature map is normalized using LayerNorm2d. This operation standardizes the mean and variance of each spatial location along the channel dimension, effectively mitigating the internal covariate shift problem and accelerating model convergence. The calculation formula is as follows:

[0064]

[0065]

[0066]

[0067] Where x is the input feature map, Let represent the feature value at position (i,j) within the channel, H and W be the height and width of the feature map, μ and σ represent the mean and standard deviation of the current channel feature map, respectively, and ϵ be a small constant to prevent division by zero. These are the normalized eigenvalues.

[0068] The normalized feature map is fed into MHSA to extract long-range dependency information from the image. This module uses eight attention heads, each independently learning feature relationships in different regions, thereby improving the diversity of feature representation and global modeling capabilities. The core computation process of MHSA includes linear mapping of Query, Key, and Value, attention score calculation, and output concatenation. The calculation formula is as follows:

[0069]

[0070]

[0071]

[0072]

[0073] The results from multiple attention heads are concatenated and then linearly projected to obtain the final output, which is calculated as follows:

[0074]

[0075] Where X represents the input feature map. , , The weight matrix represents the linear mapping. , , It is a bias term. This represents the dimension of the Key, where h represents the number of attention heads. The projection matrix for multi-head attention output.

[0076] To enhance the model's generalization ability and robustness, MHSA output randomly discards a portion of the network path using DropPath to simulate different network depths and prevent overfitting. The effect of DropPath can be formalized as shown in the following formula:

[0077]

[0078] Where x is the DropPath input value, usually representing a feature or intermediate output, and p represents the probability of dropping part of the path.

[0079] The first-stage output is obtained through residual linking. As shown in the following calculation formula:

[0080]

[0081] Subsequently, the fused features are again standardized using LayerNorm2d and fed into CGLU to enhance the non-linear expressive power of local features. CGLU includes parallel main convolutional branches and gated branches. The gated branches use the sigmoid function to generate a weight map and multiply it element-wise with the output of the main convolutional branch to achieve dynamic feature selection. Specifically, it includes:

[0082] The main branch performs a 3×3 convolution to obtain candidate features x. (1) The gated branches use Sigmoid activation to generate a weight graph. The two are multiplied element-wise to achieve dynamic feature selection, as shown in the following formula:

[0083]

[0084] Final output Compared with the results of the previous stage Residual fusion is performed, and model robustness is improved through DropPath, as shown in the following calculation formula:

[0085]

[0086] The MHSA-CGLU module uses MHSA to extract global context and CGLU to enhance nonlinear feature representation. The two work together to improve the model's adaptability and robustness in complex scenarios while maintaining feature representation capabilities.

[0087] DBR-C3k2 employs a staged, multi-scale contextual modeling strategy to replace single-path convolutional feature extraction, effectively enhancing receptive field diversity and semantic expressive power. For example... Figure 3 As shown, the specific construction steps of DBR-C3k2 include:

[0088] In the regional residual modeling stage, the original input features are downsampled and basic semantics are extracted through 3×3 convolution, batch normalization and SiLU activation function. The intermediate feature map obtained by this process has compact representation ability and simulates the structural features of different regions from the spatial dimension, providing more discriminative input for subsequent semantic extraction.

[0089] A semantic-level residual enhancement stage is constructed, which further processes the output features of the region-level residual modeling stage through three parallel convolutional paths. The first convolutional path is a standard convolution with an dilation rate of 1, the second convolutional path is a dilated reparameter block with an dilation rate of 2 and a kernel size of 5×5, and the third convolutional path is a dilated reparameter block with an dilation rate of 3 and a kernel size of 7×7. During training, multi-branch convolutions enhance feature learning, and during inference, they are fused into a single large convolutional kernel, which significantly reduces the amount of computation while maintaining accuracy.

[0090] The output features of the three convolutional paths are fused, and then added to the original input features through residual connections before output.

[0091] DBR-C3k2 first extracts basic features from the input image, denoted as... ,in Indicates input data, The standard convolution operation is represented by the following mathematical expression:

[0092]

[0093] Subsequently, dilated convolution is introduced to further process the features, where The weights of the dilated convolution are... The feature map after dilated convolution is shown in the following calculation formula:

[0094]

[0095] Furthermore, the features are reparameterized, denoted as Reparam, which represents the transformation of features through dilatational reparameterization, as shown in the following formula:

[0096]

[0097] Finally, multiple features are fused, and the output is denoted as Y. The dilated convolution features and reparameterized features are integrated through additive fusion, as shown in the following calculation formula:

[0098]

[0099] This two-stage strategy shifts the function of deep convolution from global modeling to structural filtering of refined features, thereby standardizing feature learning and enhancing multi-scale contextual representation. After channel concatenation, normalization, and 1×1 convolution compression, multi-scale features maintain feature consistency and robustness through residual connections, making it suitable for sparse and fine-grained light guide plate defect detection.

[0100] like Figure 4 As shown, the specific steps for building Rep-DEHead include:

[0101] For multi-scale feature maps from the neck network unit, channel adaptation is performed by 1×1 convolution and group normalization, respectively.

[0102] The channel-adapted feature map is input into a two-layer cascaded shared detail enhancement convolutional module, and features are extracted through convolutional kernels with shared weights.

[0103] The output of the shared detail enhancement convolution module is fed into the bounding box regression branch (output channel is 4×reg_max) and the object classification branch (output channel is the number of categories).

[0104] The bounding box regression section introduces the Distributed Focal Loss (DFL) function to enhance the accuracy modeling and numerical stability of location information, especially showing better regression capabilities in small target scenarios.

[0105] The entire forward reasoning process can be represented by the following calculation formula:

[0106]

[0107] in, Represents input features; This represents a detail-enhancing convolution operation that enhances the perception of edges, details, and noise through parallel standard convolution and differential convolution. This indicates a regression branch convolution, responsible for predicting the offset of the bounding box, with an output channel count of 4×reg_max; This represents a learnable scaling factor module used to automatically adjust the scale of the outputs from different layers (related to the feature map downsampling factor). This indicates a classification branch convolution, with the number of output channels equal to the number of categories (used for category prediction). This means concatenating the regression output and the classification output along the channel dimension to form the final output; This represents the feature tensor (containing bounding boxes and class predictions) that is ultimately output by the detection head.

[0108] During the inference phase, the regression output is decoded by combining the predicted anchor boxes and feature layer stride information, and then concatenated with the classification results to finally generate a standard-format detection box output. Because the Shared Detail Enhancement Convolutional Module (DEConv) employs a structured reparameterization design, it can be transformed into a standard convolutional form during inference without adding any extra parameters or computational burden, ensuring inference efficiency while enhancing the model's expressive power. For example... Figure 5 As shown, DEConv includes parallel standard convolutional paths (VC), central difference convolutional paths (CDC), corner difference convolutional paths (ADC), horizontal difference convolutional paths (HDC), and vertical difference convolutional paths (VDC).

[0109] To verify the performance of the PDR-YOLOv11 hot-pressed light guide plate defect detection model in defect detection tasks, this invention trained and ablated the model on a self-built LGPs (hot-pressed light guide plate) defect dataset, and compared it with several mainstream target detection models.

[0110] The self-built LGPs defect dataset was constructed as follows: images of the hot-pressed light guide plate were acquired using a linear scan camera; effective defect image fragments were extracted using a fixed-size 448×448 window by manually cropping defect areas; and the dataset was divided into the following sections: Figure 6 The dataset shows scratch, missing, dark shadow, white stain, dark spot, scar, and white shadow defects. It contains 10,075 images and 11,880 labels. The specific number of images for each defect type is detailed in Table 1.

[0111] Table 1. Distribution of the number of defect types in the LGPs defect dataset.

[0112]

[0113] Based on the sample size and training requirements, the samples of each type of defect were divided into training set, validation set and test set in a ratio of 6:2:2. The specific distribution of each type of defect in different datasets is shown in Table 2.

[0114] Table 2. Distribution of training, validation, and test sets for the LGPs defect dataset.

[0115]

[0116] When training the defect detection model for hot-pressed light guide plates, the hyperparameters for model training were set as follows: 300 training epochs, batch size of 32, stochastic gradient descent as the optimizer, initial learning rate of 0.01, cosine annealing learning rate of 0.1, weight decay coefficient of 0.005, and dynamic momentum parameter of 0.937. Details are shown in Table 3.

[0117] Table 3. Hyperparameters for training the defect detection model of the hot-pressed light guide plate

[0118]

[0119] Multiple metrics are used to evaluate model performance, including average accuracy (AP) for each class, mean average precision (mAP), number of parameters, gross computational cost (GFLOPs), precision, recall, F1 score, and frames per second (FPS). Among them, mAP comprehensively reflects the accuracy of defect detection across multiple classes, F1 score balances precision and recall, FPS assesses the model's real-time performance, and GFLOPs measure computational resource consumption.

[0120] To analyze the effects of PTFM, DBR-C3k2, and Rep-DEHead on the performance of the PDR-YOLOv11 defect detection model for hot-pressed light guide plates, ablation experiments were conducted, and the experimental results are shown in Table 4.

[0121] Table 4. Ablation Experiment Results of Defect Detection Model for Hot-pressed Light Guide Plate

[0122]

[0123] Table 5 shows that the introduction of different improved modules all had a positive impact on the performance of PDR-YOLOv11. After adding PTFM, the model's mean accuracy (mAP) increased from 96.1% to 96.8%, the number of model parameters decreased from 2.58M to 2.41M, and the computational cost decreased from 6.3 GFLOPs to 5.9 GFLOPs. At the same time, the model size was reduced from 5.2MB to 5.0MB, indicating that this module not only improved detection performance but also had good lightweighting and compression capabilities. The model with DBR-C3k2 introduced performed excellently in terms of accuracy (96.8%) and recall (94.1%). Although its number of parameters and computational cost were slightly higher than PTFM, the model size was only 5.3MB, which remained within a controllable range and is suitable for detection tasks with higher accuracy requirements. The model with only Rep-DEHead achieved an mAP of 96.6%, reduced model parameters to 2.26M, and model size to 5.0MB, making it the smallest of the three single-module structures. Simultaneously, inference speed increased to 2143.7 FPS, suitable for high real-time scenarios. In module combination experiments, the combination of PTFM and DBR-C3k2 improved mAP to 96.9% while maintaining the model size at 5.0MB, demonstrating good control over parameter and computational loads and excellent overall efficiency. The combination of PTFM and Rep-DEHead performed best, not only improving mAP to 97.3% and achieving an F1 score of 95.39%, but also reducing parameter count to 2.14M and further decreasing the model size to 4.7MB, the smallest of all combinations, showcasing strong collaborative compression capabilities and potential for accuracy improvement. The combination of DBR-C3k2 and Rep-DEHead ensures high performance while keeping the model size under 5.0MB, achieving an mAP of 97.0%. With balanced performance across all metrics, it is suitable for practical applications that prioritize stable accuracy.

[0124] After incorporating the PTFM, DBR-C3k2, and Rep-DEHead modules, the model achieved optimal performance: mAP increased to 97.4%, F1 score to 95.44%, parameter count reduced to 2.11M, computational cost decreased to 5.9 GFLOPs, inference speed reached 1464.8 FPS, and model size was only 4.8MB. Compared to the original YOLOv11 model, PDR-YOLOv11 not only improved detection accuracy by 1.3% but also significantly reduced resource consumption, demonstrating excellent deployment adaptability and industrial practical value.

[0125] To further verify the detection performance and deployment efficiency of the PDR-YOLOv11 model for detecting defects on the surface of hot-pressed light guide plates, a comparative experiment was conducted. Several mainstream object detection models were selected for comparison, including Faster-RCNN, RT-DETR-ResNet50, YOLOv7-tiny, EC-PFN, YOLOv9-t, YOLOv10, YOLOv11, and YOLOv12. All models were trained and tested on a self-built LGPs defect dataset. The comparison of test results is shown in the figure below. Figure 7 As shown in the comparison chart of the visualization results, Figure 8 As shown in Table 5, the results of the comparative experiments are presented.

[0126] Table 5 Comparison Experimental Results of Defect Detection Models for Hot-pressed Light Guide Plates

[0127]

[0128] As shown in Table 6, PDR-YOLOv11 achieved the best performance on its self-built LGPs defect dataset, with an mAP of 97.4%. Compared with mainstream detection models Faster R-CNN, RT-DETR-ResNet50, YOLOv7-tiny, EC-PFN, YOLOv9-t, YOLOv10, YOLOv11, and YOLOv12, the mAP was improved by 1.7%, 3.1%, 2.3%, 1.4%, 1.7%, 1.4%, 1.3%, and 0.9%, respectively, fully demonstrating the model's superior detection and feature extraction capabilities in the hot-pressed light guide plate defect detection task.

[0129] While YOLOv7-tiny, YOLOv9-t, and YOLOv10 demonstrate fast detection speeds (1570.5, 810.5, and 1585.7 FPS respectively), their detection accuracy is relatively low (mAP 95.1%, 95.7%, and 96.0% respectively), limiting their application in high-precision industrial quality inspection scenarios. RT-DETR-ResNet50 exhibits strong global modeling capabilities in feature learning, but its inference speed is only 130.8 FPS, resulting in poor real-time performance. Faster-R-CNN, a classic two-stage detection framework, achieves an mAP of 95.7%, but its inference speed of only 90.1 FPS is insufficient for real-time detection requirements. EC-PFN, a lightweight feature pyramid network, improves the accuracy of small target detection through multi-scale context fusion, achieving an mAP of 96.0%. However, its parameter count is as high as 24.8M, resulting in significant model complexity. Although its inference speed reaches 1621.6 FPS, there is still room for improvement in balancing accuracy and complexity. YOLOv12 optimizes the structure of YOLOv11, increasing the mAP to 96.5%, showing some performance improvement. However, it is slightly lower than PDR-YOLOv11 in comprehensive metrics such as Recall and F1-Score.

[0130] Furthermore, detection speed is a key metric for evaluating the practicality of a model in engineering. PDR-YOLOv11 achieves a real-time inference speed of 1464.8 FPS while maintaining high accuracy. Compared to Faster R-CNN, RT-DETR-ResNet50, and YOLOv9-t, the FPS is improved by 1374.7, 1334.0, and 654.3, respectively. Although slightly lower than YOLOv11 by 599.6 FPS, its comprehensive improvements in accuracy, robustness, and stability make PDR-YOLOv11 the best performer in balancing accuracy and speed, possessing high engineering deployment value and industrial application potential.

[0131] To further verify the generalization ability of the PDR-YOLOv11 model for detecting defects in hot-pressed light guide plates, a generalization experiment was conducted on a publicly available railway defect detection dataset, comparing it with several other mainstream target detection models. The comparison results are shown in the figure below. Figure 9 As shown in Table 6, the results of the generalization experiment are as follows.

[0132] Table 6. Generalization Experiment Results of the Defect Detection Model for Hot-Pressed Light Guide Plates

[0133]

[0134] As shown in Table 6, PDR-YOLOv11 performs best in all performance metrics on the rail defect detection dataset, with an mAP of 97.8%. Compared with Faster-RCNN, RT-DETR-ResNet50, YOLOv7-tiny, EC-PFN, YOLOv9-t, YOLOv10, YOLOv11 and YOLOv12, it improves by 5.0%, 1.5%, 3.9%, 3.1%, 8.1%, 0.5%, 0.4% and 3.7% respectively, fully verifying the significant advantage of this model in detection accuracy.

[0135] While YOLOv10, YOLOv11, and YOLOv12 already demonstrated excellent detection performance, PDR-YOLOv11 achieved further improvements in Precision (96.2%) and Recall (94.5%), with an F1 score of 95.34%, representing improvements of 1.2%, 2.7%, and 1.06% respectively compared to YOLOv11, and improvements of 5.2%, 6.1%, and 5.66% respectively compared to YOLOv12. This indicates that PDR-YOLOv11 possesses stronger robustness and adaptability in handling challenging detection tasks such as small defects, multi-scale targets, and complex textured backgrounds.

[0136] In terms of inference speed, PDR-YOLOv11 achieves a real-time detection performance of 1178.4 FPS. Compared with Faster-RCNN, RT-DETR-ResNet50, and YOLOv9-t, the inference speed is improved by 1073.8, 1061.5, and 405.2 FPS, respectively. Although slightly lower than YOLOv11 (reduced by 488.7 FPS) and YOLOv12 (reduced by 333.4 FPS), it surpasses them in key metrics such as detection accuracy and recall, demonstrating superior overall performance.

[0137] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A method for detecting surface defects in a hot-pressed light guide plate based on hybrid feature extraction, characterized in that: The method includes the following steps: A defect detection model for hot-pressed light guide plates is constructed based on the YOLOv11 network architecture. The hot-pressed light guide plate defect detection model includes a backbone network unit, a neck network unit, and a detection head unit. The backbone network unit embeds a parallel texture fusion module for collaboratively modeling local texture features and global context features. The neck network unit embeds a dilated block reparameterization module for expanding the receptive field through dilated convolution and reparameterization mechanisms. The detection head unit includes multiple reparameterized detail enhancement detection heads that employ shared convolution and differential convolution. A surface image of the hot-pressed light guide plate is acquired and a defect dataset is constructed. The defect dataset is then used to train a defect detection model for the hot-pressed light guide plate. The image of the hot-pressed light guide plate to be detected is input into the trained hot-pressed light guide plate defect detection model, and the defect detection result is output.

2. The method for detecting surface defects of a hot-pressed light guide plate based on hybrid feature extraction according to claim 1, characterized in that: The construction steps of the parallel texture fusion module specifically include: The input feature map is divided into a first branch and a second branch in the channel dimension. The first branch is used to extract local texture features through convolution operations, and the second branch is used to capture global long-range dependency information through the MHSA-CGLU submodule. The output features of the first branch and the output features of the second branch are concatenated and fused, and then channel integration is performed through a convolutional layer before output.

3. The method for detecting surface defects of a hot-pressed light guide plate based on hybrid feature extraction according to claim 2, characterized in that: The MHSA-CGLU submodule includes a first two-dimensional normalization layer, a multi-head self-attention mechanism layer, a first random path drop-off layer, a second two-dimensional normalization layer, a convolutional gated linear unit layer, and a second random path drop-off layer connected in sequence. The convolutional gated linear unit layer includes a parallel main convolutional branch and a gated branch. The gated branch uses the Sigmoid function to generate a weight map and multiplies it element-wise with the output of the main convolutional branch to achieve dynamic feature selection.

4. The method for detecting surface defects of a hot-pressed light guide plate based on hybrid feature extraction according to claim 1, characterized in that: The void block reparameter module adopts a phased multi-scale context modeling strategy. The construction steps of the void block reparameter module specifically include: In the stage of constructing regional residual modeling, the original input features are downsampled and basic semantics are extracted through 3×3 convolution, batch normalization and SiLU activation function; A semantic-level residual enhancement stage is constructed. The output features of the region-level residual modeling stage are further processed through three parallel convolutional paths. The first convolutional path is a standard convolution with an expansion rate of 1, the second convolutional path is a dilated reparameter block with an expansion rate of 2 and a kernel size of 5×5, and the third convolutional path is a dilated reparameter block with an expansion rate of 3 and a kernel size of 7×7. The output features of the three convolutional paths are fused, and then added to the original input features through residual connections before output.

5. The method for detecting surface defects of a hot-pressed light guide plate based on hybrid feature extraction according to claim 1, characterized in that: The construction steps of the reparameterized detail enhancement detection head specifically include: For the multi-scale feature maps from the neck network unit, channel adaptation is performed by 1×1 convolution and group normalization, respectively; The channel-adapted feature map is input into a two-layer cascaded shared detail enhancement convolutional module, and features are extracted through convolutional kernels with shared weights. The output of the shared detail enhancement convolution module is fed into the bounding box regression branch and the object classification branch, respectively.

6. The method for detecting surface defects of a hot-pressed light guide plate based on hybrid feature extraction according to claim 5, characterized in that: The shared detail enhancement convolution module includes parallel standard convolution paths, center difference convolution paths, corner difference convolution paths, horizontal difference convolution paths, and vertical difference convolution paths; during the model inference stage, the parallel convolution paths of the shared detail enhancement convolution module can be reparameterized into single-path standard convolutions.

7. The method for detecting surface defects of a hot-pressed light guide plate based on hybrid feature extraction according to claim 1, characterized in that: The backbone network unit includes two convolutional modules, a first C3k2 module, a third convolutional module, a second C3k2 module, a fourth convolutional module, a first parallel texture fusion module, a fifth convolutional module, a second parallel texture fusion module, a spatial pyramid pooling module, and a C2PSA attention module connected in sequence. The outputs of the second C3k2 module, the first parallel texture fusion module, and the C2PSA attention module are connected to the neck network unit as three feature outputs of the backbone network unit.

8. The method for detecting surface defects of a hot-pressed light guide plate based on hybrid feature extraction according to claim 7, characterized in that: The neck network unit includes multiple upsampling modules, splicing modules, downsampling convolution modules, and hole block reparameter modules; The output of the C2PSA attention module of the backbone network unit is connected to the first upsampling module, and then connected to the output of the first parallel texture fusion module of the backbone network unit through the first splicing module, and input to the first hole block weight parameter module. The output of the first void block weight parameter module is connected to the second upsampling module, and then connected to the output of the second C3k2 module of the backbone network unit through the second splicing module, and input to the second void block weight parameter module; The output of the second hole block re-parameter module is connected to the first downsampling convolution module, and then connected to the output of the first hole block re-parameter module through the third splicing module, and input to the third hole block re-parameter module. The output of the third dilated block reparameter module is connected to the second downsampling convolution module, and then connected to the output of the C2PSA attention module of the backbone network unit through the fourth splicing module, and input to the fourth dilated block reparameter module. The outputs of the second void block reparameter module, the third void block reparameter module, and the fourth void block reparameter module are respectively connected to three different reparameterized detail enhancement detection heads.

9. The method for detecting surface defects of a hot-pressed light guide plate based on hybrid feature extraction according to claim 1, characterized in that: When training the hot-pressed light guide plate defect detection model, the hyperparameters of the model training are set as follows: 300 training rounds, batch size of 32, stochastic gradient descent as the optimizer, initial learning rate of 0.01, cosine annealing learning rate of 0.1, weight decay coefficient of 0.005, and dynamic momentum parameter of 0.937.

Images