Lightweight feature generation and global perception enhanced road crack detection method

Abstract

The application discloses a light-weight feature generation and global perception enhanced road crack detection method, and specifically implements the following steps: step 1, road crack image dataset acquisition and preprocessing; step 2, crack feature extraction and modeling based on a fusion feature enhancement strategy; step 3, detecting the input road crack image, obtaining the target position information of the road crack and the corresponding crack category information, and outputting the crack detection result. The application solves the problem that the existing road crack detection method in actual application introduces a complex network structure to improve detection accuracy, which leads to large model parameter size, high calculation cost, difficulty in deployment in actual inspection equipment, and easy weakening of the overall structure perception ability of the slender and continuous crack in the model lightening process.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and detection technology, specifically relating to a road crack detection method with lightweight feature generation and global perception enhancement. Background Technology

[0002] In recent years, with the continuous expansion of urban transportation infrastructure construction, roads and bridges have played a crucial role in the transportation system. During their long-term service, roads are continuously subjected to the effects of the natural environment and cyclical vehicle loads, inevitably leading to fatigue and deterioration of pavement materials, resulting in defects such as cracks. Cracks are a significant indicator of pavement structural performance degradation, typically characterized by their elongated shape, strong continuity, and large dimensional variations. If not detected and repaired in a timely manner, cracks will continue to expand under the influence of multiple factors such as rainwater infiltration and vehicle traffic, significantly weakening the load-bearing capacity of the road structure, threatening traffic safety, and shortening the road's service life. Therefore, effective detection of road cracks is a crucial technical aspect of highway maintenance and safety management.

[0003] Currently, road crack detection mainly employs manual inspection or automated methods. Manual inspection typically relies on inspectors visually examining the road surface by walking or driving at low speeds. This method suffers from low efficiency, high labor intensity, and reliance on human experience, making it unsuitable for large-scale road inspections. To improve efficiency, some studies have adopted automated detection methods based on traditional image processing techniques, extracting crack features through threshold segmentation, edge detection, and morphological operations. However, these methods have limited adaptability to changes in crack morphology, and their stability and applicability are somewhat limited when dealing with cracks with large scale spans and strong structural continuity.

[0004] With the development of computer vision and artificial intelligence technologies, deep learning-based road crack detection methods have been increasingly researched and applied. These methods automatically learn crack feature representations from sample data by constructing multi-layer neural network models, offering advantages in detection accuracy compared to traditional methods. Based on different detection processes, existing deep learning-based crack detection methods can be mainly divided into two-stage detection methods and single-stage detection methods.

[0005] Two-stage detection methods typically achieve crack detection through candidate region generation and refined classification or segmentation, achieving high localization accuracy. Related research has improved the detection capability for fine and complex-shaped cracks by introducing multi-scale feature fusion or refining the network structure. However, these methods have complex overall network structures, cumbersome computational processes, strong reliance on computational resources, and low inference efficiency, making it difficult to meet the real-time and lightweight deployment requirements of practical road inspection scenarios.

[0006] Single-stage detection methods achieve rapid localization and identification of crack targets through end-to-end regression mechanisms, exhibiting significant advantages in detection efficiency and gradually becoming an important research direction for real-time road crack detection. In recent years, related research has explored lightweight network structure design and feature fusion strategies to improve crack detection performance. However, road cracks typically exhibit characteristics such as large scale span, elongated and irregular morphology, and strong overall structural continuity. Some existing single-stage detection methods still have shortcomings in feature representation and overall structural perception capabilities, making it difficult to stably characterize the overall morphology of cracks while ensuring model lightweighting and detection efficiency.

[0007] Furthermore, in practical engineering applications, road crack detection models typically need to be deployed on vehicle-mounted equipment or other platforms with limited computing resources, placing high demands on the scale of model parameters and computational complexity. Some existing methods introduce complex network structures to improve detection accuracy, leading to a significant increase in the number of model parameters and computational overhead, which is detrimental to practical deployment. Conversely, lightweighting the model can weaken its ability to express key crack features, thus affecting the perception of the overall crack structure. Therefore, how to improve the effective expression of crack features and enhance the perception of the overall crack structure while meeting the requirements of lightweighting and detection efficiency has become a pressing technical problem in existing road crack detection technologies. Summary of the Invention

[0008] The purpose of this invention is to provide a lightweight feature generation and global perception enhancement method for road crack detection, which solves the problems of existing road crack detection methods in practical applications, such as the introduction of complex network structures to improve detection accuracy, resulting in large model parameter scale, high computational cost, difficulty in deployment in actual inspection equipment, and the weakening of the perception ability of the overall structure of slender and continuous cracks during the process of model lightweighting.

[0009] The technical solution adopted in this invention is a road crack detection method with lightweight feature generation and global perception enhancement, which is implemented according to the following steps:

[0010] Step 1: Acquisition and preprocessing of road crack image dataset; Step 2: Crack feature extraction and modeling based on fusion feature enhancement strategy; Step 3: Detect the input road crack image, obtain the target location information of the road crack and its corresponding crack category information, and output the crack detection results.

[0011] The invention is further characterized in that, Step 1 is implemented in the following steps: Step 1.1: Obtain road surface image data from multiple publicly available road crack datasets, and fuse the image data to construct a comprehensive road crack image dataset. The road crack image dataset covers three typical road crack types: transverse cracks, longitudinal cracks, and branch cracks. Step 1.2: The road crack images are manually annotated. The crack categories are distinguished according to the main extension direction and overall morphological characteristics of the cracks, and the corresponding crack target location information and category information are generated. The annotation results are saved in a unified data format to ensure the consistency of subsequent model training and performance evaluation. Step 1.3: The labeled dataset is divided into training, validation, and test sets in an 8:1:1 ratio for training, parameter tuning, and performance evaluation of the road crack detection model. By constructing a comprehensive dataset covering multiple crack morphologies, a data foundation is provided for the subsequent detection model to achieve stable detection under different crack structure conditions.

[0012] Step 2 is based on the YOLOv11 detection network. While maintaining the advantages of detection efficiency and lightweight design, the feature extraction and modeling process of the YOLOv11 detection network is improved. A road crack detection model based on lightweight feature generation and global perception enhancement, GSC-YOLO, is proposed. The GSC-YOLO network model includes a backbone network, a neck network and a head detection network connected in sequence.

[0013] In step 2, the backbone network is used to extract features from the input image. The input image sequentially passes through the first Ghost convolution, the second Ghost convolution, the first C3k2 module, the third Ghost convolution, the second C3k2 module, the fourth Ghost convolution, the third C3k2 module, the fifth Ghost convolution, the fourth C3k2 module, the SPPF_UniRepLK module, and the C2PSA module. The backbone network outputs three different levels of features to the neck network. Specifically, the output of the second C3k2 module is used as the first feature output, the output of the third C3k2 module is used as the second feature output, and the output of the C2PSA module is used as the third feature output.

[0014] In step 2, the neck network receives features from the backbone network and performs multi-scale fusion. Specifically, the third feature output of the backbone network is upsampled first and then concatenated with the second feature output of the backbone network, before being input to the first C3k2 module of the neck network. The output of the first C3k2 module of the neck network is upsampled second and then concatenated with the first feature output of the backbone network, before being input to the second C3k2 module of the neck network. The output of the second C3k2 module of the neck network is split into two paths: one path is input to the first CBAM attention module, and the other path is downsampled by the first Ghost convolution and then concatenated with the output of the first C3k2 module of the neck network, before being input to the third C3k2 module of the neck network. The output of the third C3k2 module of the neck network is also split into two paths: one path is input to the second CBAM attention module, and the other path is downsampled by the second Ghost convolution and then concatenated with the third feature output of the backbone network, before being input to the fourth C3k2 module of the neck network. Finally, the output of the fourth C3k2 module of the neck network is input to the third CBAM attention module. The head detection network in step 2 contains three parallel detection layers, which respectively receive the refined features output by the first CBAM attention module, the second CBAM attention module and the third CBAM attention module in the neck network, and are used to output target detection results at different scales.

[0015] Step 2 specifically includes the following improvement strategies: (1) Lightweight feature generation strategy; (2) Global awareness enhancement strategy; (3) Attention feature enhancement strategies.

[0016] The lightweight feature generation strategy is as follows: The Ghost convolution module effectively reduces computational complexity through a two-stage feature generation process. First, a small number of standard convolutions are used to generate intrinsic feature maps. Then, a computationally inexpensive linear transformation operation is applied to these intrinsic feature maps to generate extended feature maps. Finally, the intrinsic and extended feature maps are concatenated to obtain an output feature map comparable to that of standard convolutions. Specifically, the input feature map first undergoes standard convolution to generate intrinsic feature maps. These maps are then split into two paths: one path preserves them directly as an identity mapping, while the other path feeds into a parallel linear operation layer to generate extended feature maps. Finally, the two are concatenated to obtain the final output feature map.

[0017] A comparison of the computational cost and parameter count of the two convolutional structures is provided: The computational cost of a standard convolution is: (1) in, The computational cost is the same as that of a regular convolution. The kernel size is [size]. Input the number of channels. The number of convolution kernels and the number of output channels, and To determine the width and height of the output feature map, each convolution kernel generates a feature map when the input image undergoes a convolution operation. These feature maps together form the final multi-channel feature map. Therefore, the number of convolution kernels is equal to the number of channels in the output feature map. The computational complexity of the Ghost convolutional module is: (2) (3) in, The computational cost of Ghost convolution. The number of channels in the intrinsic feature map generated by a standard convolution. To increase the multiplier, The size of the convolution kernel for linear operations; The computational cost ratio of regular convolution to Ghost convolution The ratio of parameters They are respectively: (4) (5) As can be seen from the above expression, under the condition of equivalent parameter configuration, the Ghost convolution module can effectively reduce the amount of computation and parameters required for convolution operations. Therefore, through this strategy, the network parameter size and computational overhead can be effectively reduced without significantly reducing the feature expression capability, thereby improving the model running efficiency and providing computational margin for subsequent global perception and attention feature enhancement. The specific strategies for enhancing global awareness are as follows: This paper improves the spatial pyramid pooling module in the detection network by introducing the large kernel convolution concept and proposing a feature fusion structure based on global modeling using a large convolution kernel, namely the SPPF_UniRepLK module. UniRepLKNet adopts a hierarchical design, with input features sequentially passing through downsampling and stages 1, 2, 3, and 4. Stages 1, 2, and 4 are mainly used for conventional feature extraction and downsampling operations at different scales on the feature maps, constructing a multi-scale feature pyramid. Stage 3, as the core feature extraction stage, includes alternating large and small kernel modules. The structure of the large kernel module is as follows: after feature input, the main branch sequentially passes through a dilated reparameterization module and a batch normalization layer; then the features are split into two paths to generate attention weights, one path is processed by average pooling, ReLU activation function, and Sigmoid function, and the other path is element-wise multiplied with the main branch features; the processed features then pass through a GELU activation function and a batch normalization layer; finally, the output of the large kernel module is obtained through residual connections.

[0018] In this invention, the implementation and specific connection relationship of the SPPF_UniRepLK module are as follows: the input features it receives come from the output of the fourth C3k2 module in the aforementioned backbone network; after the input features enter the SPPF_UniRepLK module, they first pass through a regular convolutional layer, and then are respectively input into three sequentially connected two-dimensional max pooling layers; next, the output of the convolutional layer and the output of the three two-dimensional max pooling layers are concatenated; the concatenated multi-scale fusion features are sent to the UniRepLK module; the final output features serve as the final output of the entire SPPF_UniRepLK module and are directly passed to the C2PSA module of the backbone network for subsequent processing.

[0019] The specific strategies for enhancing attention features are as follows: A CBAM attention module is embedded before the detection layer to adaptively enhance crack-related features. The CBAM module consists of a cascaded channel attention module and a spatial attention module. The channel attention module characterizes the importance of different feature channels in the crack detection task, while the spatial attention module characterizes the saliency distribution of crack features in spatial locations. The input feature map received by the channel attention module comes from the output of the C3k2 module in the neck network (i.e., the second, third, or fourth C3k2 module of the neck network). The channel attention module first processes the input feature map... The data is processed through max pooling and average pooling layers to extract salient response information and global context information for each channel. Then, the two feature vectors are fed into a multilayer perceptron with shared weights for nonlinear mapping. Next, the two feature vectors output by the multilayer perceptron are added element-wise, and the output feature map of the channel attention module is generated using a sigmoid activation function. That is, the channel attention weights; finally, these weights are compared with the original input feature map. Channel-by-channel multiplication enables adaptive recalibration of key feature channels, resulting in an enhanced feature map. .

[0020] After completing the channel dimension feature enhancement, the implementation path of the spatial attention module is as follows: Figure 7 As shown. The spatial attention module outputs feature maps as the channel attention module. As input, salient region information and overall response intensity are extracted by performing max pooling and average pooling operations along the channel dimension, respectively, and then concatenated along the channel dimension. Subsequently, spatial features are aggregated using convolutional layers, and the spatial attention module output feature map is generated by passing through a sigmoid activation function. That is, the spatial attention weight map; finally, the weight map is... Feature maps input to the spatial attention module Pixel-by-pixel multiplication is used to highlight the spatial location of the crack region, and the final refined features are then fed into the head detection network.

[0021] Step 3 is implemented in the following steps: The GSC-YOLO road crack detection model, constructed based on the above feature extraction and modeling process, performs forward inference on the input road crack image and outputs the bounding box coordinates of the crack target and the corresponding crack category confidence. The crack categories include transverse cracks, longitudinal cracks and branch cracks.

[0022] The beneficial effects of this invention are that it provides a lightweight feature generation and global perception enhancement method for road crack detection. This method employs a lightweight feature generation strategy, which effectively reduces the model parameter size and computational complexity while maintaining crack feature extraction capabilities, significantly improving the model's deployability in computationally limited environments. Through a global perception enhancement strategy, it enhances the model's ability to perceive the overall structural information of cracks, thereby improving the detection stability of slender and continuous crack targets. By fusing an attention feature enhancement strategy, it achieves responsiveness to key crack features and suppresses interference from irrelevant features, thus enhancing the robustness and reliability of detection. This invention can stably and accurately identify transverse, longitudinal, and branched crack targets, and is suitable for vehicle-mounted inspection equipment, UAV monitoring, and edge computing platforms. It can be widely applied in various road inspection and maintenance scenarios, demonstrating good engineering application prospects and promotion potential. Attached Figure Description

[0023] Figure 1 This is a sample image of a typical road crack; Figure 2 This is a schematic diagram of the overall network structure of the GSC-YOLO road crack detection method with lightweight feature generation and global perception enhancement proposed in this invention; Figure 3 This is the network structure diagram of the lightweight feature generation Ghost convolutional module; Figure 4 This is a network structure diagram of the UniRepLK enhancement unit; Figure 5 This is a diagram of the CBAM attention module network structure; Figure 6 This is a network structure diagram of the channel attention module; Figure 7 This is a network structure diagram of the spatial attention module; Figure 8 This is a graph showing the trend of model detection performance as the number of training rounds increases. Detailed Implementation

[0024] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0025] The lightweight feature generation and global perception enhancement road crack detection method of this invention is implemented according to the following steps: Step 1: Acquisition and preprocessing of road crack image dataset; Step 1 is implemented in the following steps: Step 1.1: Obtain road surface image data from multiple publicly available road crack datasets, and fuse the image data to construct a comprehensive road crack image dataset with 4000 images. The road crack image dataset covers three typical road crack types: transverse cracks, longitudinal cracks, and branching cracks. Typical samples are illustrated below. Figure 1 As shown.

[0026] Step 1.2: The road crack images are manually annotated. The crack categories are distinguished according to the main extension direction and overall morphological characteristics of the cracks, and the corresponding crack target location information and category information are generated. The annotation results are saved in a unified data format to ensure the consistency of subsequent model training and performance evaluation. Step 1.3: The labeled dataset is divided into training, validation, and test sets in an 8:1:1 ratio for training, parameter tuning, and performance evaluation of the road crack detection model. By constructing a comprehensive dataset covering multiple crack morphologies, a data foundation is provided for the subsequent detection model to achieve stable detection under different crack structure conditions.

[0027] Step 2: Crack feature extraction and modeling based on fusion feature enhancement strategy; In step 2, To address the common characteristics of road cracks, such as large scale span, elongated shape, and strong overall structural continuity, and the shortcomings of existing lightweight road crack detection models in terms of feature representation and overall structural perception, this invention, based on the YOLOv11 detection network, improves the feature extraction and modeling process of the YOLOv11 detection network while maintaining its detection efficiency and lightweight advantages. This results in a lightweight road crack detection method with enhanced feature generation and global perception, named GSC-YOLO. Its network structure diagram is shown below. Figure 2 As shown in the figure, the positions marked with asterisks indicate specific technical improvements made compared to the basic detection network.

[0028] Specifically, the GSC-YOLO network model includes a backbone network, a neck network, and a head detection network connected in sequence.

[0029] The backbone network is used to extract features from the input image. The input image sequentially passes through a first Ghost convolution, a second Ghost convolution, a first C3k2 module, a third Ghost convolution, a second C3k2 module, a fourth Ghost convolution, a third C3k2 module, a fifth Ghost convolution, a fourth C3k2 module, an SPPF_UniRepLK module, and a C2PSA module. The backbone network outputs three different levels of features to the neck network. Specifically, the output of the second C3k2 module is used as the first feature output, the output of the third C3k2 module is used as the second feature output, and the output of the C2PSA module is used as the third feature output.

[0030] The neck network receives features from the backbone network and performs multi-scale fusion. Specifically, the third feature output of the backbone network is upsampled first and then concatenated with the second feature output of the backbone network, and input to the first C3k2 module of the neck network; the output of the first C3k2 module of the neck network is upsampled second and then concatenated with the first feature output of the backbone network, and input to the second C3k2 module of the neck network; the output of the second C3k2 module of the neck network is split into two paths: one path is input to the first CBAM attention module, and the other path is downsampled by the first Ghost convolution and then concatenated with the output of the first C3k2 module of the neck network, and input to the third C3k2 module of the neck network; the output of the third C3k2 module of the neck network is also split into two paths: one path is input to the second CBAM attention module, and the other path is downsampled by the second Ghost convolution and then concatenated with the third feature output of the backbone network, and input to the fourth C3k2 module of the neck network; finally, the output of the fourth C3k2 module of the neck network is input to the third CBAM attention module.

[0031] The head detection network comprises three parallel detection layers, which respectively receive refined features output by the first CBAM attention module, the second CBAM attention module, and the third CBAM attention module in the neck network, and are used to output target detection results at different scales.

[0032] This invention introduces multiple feature enhancement strategies in the feature extraction and feature fusion stages, improving the ability to express the overall structural features and key local features of road cracks while reducing the model parameter scale and computational complexity. Specifically, these improvements include the following strategies: (1) Lightweight feature generation strategy To address the problem that conventional standard convolutions have a large number of parameters and high computational complexity during feature extraction, which is not conducive to the deployment of road inspection equipment and edge platforms, this invention introduces a lightweight feature generation strategy in the backbone feature extraction stage of the detection network to replace some standard convolutional structures.

[0033] Specifically, the Ghost convolutional module replaces the conventional standard convolutional layers. The Ghost convolutional module effectively reduces computational complexity through a two-stage feature generation process. First, a small number of standard convolutions are used to generate intrinsic feature maps; then, a computationally inexpensive linear transformation operation is applied to the intrinsic feature maps to generate extended feature maps. Finally, the intrinsic feature maps and extended feature maps are concatenated to obtain an output feature map comparable to that of standard convolutions. Its module structure diagram is shown below. Figure 3 As shown. Specifically, the input feature map is first processed by standard convolution to generate a feature map. Then, this feature map is divided into two paths: one path is directly preserved as an identity mapping, and the other path is input to a parallel linear operation layer to generate an extended feature map. Finally, the two are concatenated to obtain the final output feature map.

[0034] To illustrate the advantages of the Ghost convolution module over the conventional standard convolution structure in terms of computational complexity, the following comparison of the computational cost and parameter count of the two convolution structures is provided: The computational cost of a standard convolution is: (1) in, The computational cost is the same as that of a regular convolution. The kernel size is [size]. Input the number of channels. The number of convolution kernels and the number of output channels, and To determine the width and height of the output feature map, each convolution kernel generates a feature map when the input image undergoes a convolution operation. These feature maps together form the final multi-channel feature map. Therefore, the number of convolution kernels is equal to the number of channels in the output feature map. The computational complexity of the Ghost convolutional module is: (2) (3) in, The computational cost of Ghost convolution. The number of channels in the intrinsic feature map generated by a standard convolution. To increase the multiplier, The size of the convolution kernel for linear operations; The computational cost ratio of regular convolution to Ghost convolution The ratio of parameters They are respectively: (4) (5) As can be seen from the above expression, under the condition of equivalent parameter configuration, the Ghost convolution module can effectively reduce the amount of computation and parameters required for convolution operations. Therefore, through this strategy, the network parameter size and computational overhead can be effectively reduced without significantly reducing the feature expression capability, thereby improving the model running efficiency and providing computational margin for subsequent global perception and attention feature enhancement. (2) Global Awareness Enhancement Strategy To address the problem that traditional spatial pyramid pooling structures, which primarily rely on max pooling, are prone to losing detailed crack information during downsampling, thus limiting the ability to detect small and continuous crack structures, this invention improves the spatial pyramid pooling module in the detection network. Inspired by the UniRepLKNet network, it introduces the concept of large-kernel convolution and proposes a feature fusion structure based on global modeling with large convolutional kernels, namely the SPPF_UniRepLK module. The network structure diagram of UniRepLKNet is shown below. Figure 4 As shown.

[0035] like Figure 4 As shown, UniRepLKNet employs a hierarchical design, with input features sequentially passing through downsampling and stages 1, 2, 3, and 4. Stages 1, 2, and 4 primarily perform conventional feature extraction and downsampling operations at different scales on the feature maps, constructing a multi-scale feature pyramid. Stage 3, as the core feature extraction stage, contains alternating large and small kernel modules. The structure of the large kernel module is as follows: after feature input, the main branch sequentially passes through a dilated reparameterization module and a batch normalization layer; subsequently, the features are split into two paths to generate attention weights, one path undergoes average pooling, ReLU activation, and a Sigmoid function, while the other path is element-wise multiplied with the main branch features; the processed features then pass through a GELU activation function and a batch normalization layer; finally, the output of the large kernel module is obtained through residual connections.

[0036] In this invention, the implementation and specific connection relationship of the SPPF_UniRepLK module are as follows: the input features it receives come from the output of the fourth C3k2 module in the aforementioned backbone network; after the input features enter the SPPF_UniRepLK module, they first pass through a regular convolutional layer, and then are respectively input into three sequentially connected two-dimensional max pooling layers; next, the output of the convolutional layer and the output of the three two-dimensional max pooling layers are concatenated; the concatenated multi-scale fusion features are sent to the UniRepLK module; the final output features serve as the final output of the entire SPPF_UniRepLK module and are directly passed to the C2PSA module of the backbone network for subsequent processing.

[0037] The SPPF_UniRepLK module expands the network's effective receptive field by introducing non-dilated, large-size convolutional kernels, thereby enhancing its ability to model the overall extension structure and long-range dependency features of road cracks. Simultaneously, through the synergistic effect of convolutional kernels of different scales, it achieves a joint representation of global semantic features and local detail features, effectively mitigating the loss of crack detail information caused by max pooling operations.

[0038] In addition, the module incorporates a feature reparameterization mechanism in its structural design, enabling large convolutional kernels to have strong feature representation capabilities during the training phase without significantly increasing computational complexity during the inference phase. This allows the module to enhance global perception capabilities while maintaining model lightweightness and detection efficiency.

[0039] (3) Attention feature enhancement strategies To address the issue that crack areas in real-world road scenarios are easily affected by factors such as road surface texture and lighting changes, leading to insufficient response of detection models to key crack regions, this invention introduces an attention feature enhancement strategy in the feature extraction and fusion stage of the detection network. Specifically, a CBAM attention module is embedded before the detection layer to adaptively enhance crack-related features and suppress redundant features unrelated to crack detection, thereby improving the model's ability to discriminate road crack targets and its robustness in detection. Its module structure diagram is shown below. Figure 5 As shown.

[0040] The CBAM module consists of a cascaded channel attention module and a spatial attention module. The channel attention module characterizes the importance of different feature channels in the crack detection task, while the spatial attention module characterizes the saliency distribution of crack features in spatial locations. The structure diagrams of the channel attention module and the spatial attention module are shown below. Figure 6 and Figure 7 As shown. Specifically, the input feature map received by the channel attention module comes from the output of the C3k2 module in the neck network (i.e., the second, third, or fourth C3k2 module of the neck network). Figure 6 As shown, the channel attention module first processes the input feature map The data is processed through max pooling and average pooling layers to extract salient response information and global context information for each channel. Then, the two feature vectors are fed into a multilayer perceptron with shared weights for nonlinear mapping. Next, the two feature vectors output by the multilayer perceptron are added element-wise, and the output feature map of the channel attention module is generated using a sigmoid activation function. That is, the channel attention weights; finally, these weights are compared with the original input feature map. Channel-by-channel multiplication enables adaptive recalibration of key feature channels, resulting in an enhanced feature map. .

[0041] After completing the channel dimension feature enhancement, the implementation path of the spatial attention module is as follows: Figure 7 As shown. The spatial attention module outputs feature maps as the channel attention module. As input, salient region information and overall response intensity are extracted by performing max pooling and average pooling operations along the channel dimension, respectively, and then concatenated along the channel dimension. Subsequently, spatial features are aggregated using convolutional layers, and the spatial attention module output feature map is generated by passing through a sigmoid activation function. That is, the spatial attention weight map; finally, the weight map is... Feature maps input to the spatial attention module Pixel-by-pixel multiplication is used to highlight the spatial location of the crack region, and the final refined features are then fed into the head detection network.

[0042] By employing the aforementioned attention feature enhancement strategy, the detection network can more accurately focus on key channels and spatial regions related to cracks, effectively suppressing interference from background textures and non-crack structures, thereby improving the detection stability and robustness for both small and complex crack morphologies.

[0043] Step 3: Detect the input road crack image, obtain the target location information of the road crack and its corresponding crack category information, and output the crack detection results.

[0044] Step 3 is implemented in the following steps: The road crack detection model constructed based on the above feature extraction and modeling process performs forward reasoning on the input road crack image and outputs the location information of the crack target and the corresponding crack category confidence. The crack categories include transverse cracks, longitudinal cracks and branch cracks.

[0045] Example 1 The lightweight feature generation and global perception enhancement road crack detection method of this invention is implemented according to the following steps: Step 1: Acquisition and preprocessing of road crack image dataset; Step 2: Crack feature extraction and modeling based on fusion feature enhancement strategy; Step 3: Detect the input road crack image, obtain the target location information of the road crack and its corresponding crack category information, and output the crack detection results.

[0046] Example 2 The lightweight feature generation and global perception enhancement road crack detection method of this invention is implemented according to the following steps: Step 1: Road Crack Image Dataset Acquisition and Preprocessing. Road surface image data from multiple publicly available road crack datasets are acquired and fused to construct a comprehensive road crack image dataset of 4000 images. This dataset covers three typical road crack types: transverse cracks, longitudinal cracks, and branching cracks. Typical samples are illustrated below. Figure 1 As shown.

[0047] Step 2: Crack Feature Extraction and Modeling Based on Fusion Feature Enhancement Strategy. Based on the road crack image data obtained in Step 1, the image data is input into the road crack detection network. A fusion feature enhancement strategy is introduced during the feature extraction and modeling process to model the crack features. This fusion feature enhancement strategy includes a lightweight feature generation strategy, a global perception enhancement strategy, and an attention feature enhancement strategy. Through the synergistic effect of multiple feature enhancement mechanisms, the ability to express the overall structural features and key local features of road cracks is improved while reducing the number of model parameters and computational complexity.

[0048] Step 3: Crack Target Detection and Result Output. Based on the road crack detection model constructed in Step 2, the input road crack image is detected to obtain the target location information of the road cracks and their corresponding crack category information, and the crack detection results are output. The crack categories include transverse cracks, longitudinal cracks, and branch cracks.

[0049] Example 3 The lightweight feature generation and global perception enhancement road crack detection method of this invention is implemented according to the following steps: Step 1: Acquisition and preprocessing of road crack image dataset; Step 1 is implemented in the following steps: Step 1.1: Obtain road surface image data from multiple publicly available road crack datasets, and fuse the image data to construct a comprehensive road crack image dataset with 4000 images. The road crack image dataset covers three typical road crack types: transverse cracks, longitudinal cracks, and branching cracks. Typical samples are illustrated below. Figure 1 As shown.

[0050] Step 1.2: The road crack images are manually annotated. The crack categories are distinguished according to the main extension direction and overall morphological characteristics of the cracks, and the corresponding crack target location information and category information are generated. The annotation results are saved in a unified data format to ensure the consistency of subsequent model training and performance evaluation. Step 1.3: The labeled dataset is divided into training, validation, and test sets in an 8:1:1 ratio for training, parameter tuning, and performance evaluation of the road crack detection model. By constructing a comprehensive dataset covering multiple crack morphologies, a data foundation is provided for the subsequent detection model to achieve stable detection under different crack structure conditions.

[0051] Step 2: Crack feature extraction and modeling based on fusion feature enhancement strategy; Step 3: Detect the input road crack image, obtain the target location information of the road crack and its corresponding crack category information, and output the crack detection results.

[0052] Example 4 The lightweight feature generation and global perception enhancement road crack detection method of this invention is implemented according to the following steps: Step 1: Acquisition and preprocessing of road crack image dataset; Step 1 is implemented in the following steps: Step 1.1: Obtain road surface image data from multiple publicly available road crack datasets, and fuse the image data to construct a comprehensive road crack image dataset with 4000 images. The road crack image dataset covers three typical road crack types: transverse cracks, longitudinal cracks, and branching cracks. Typical samples are illustrated below. Figure 1 As shown.

[0053] Step 1.2: The road crack images are manually annotated. The crack categories are distinguished according to the main extension direction and overall morphological characteristics of the cracks, and the corresponding crack target location information and category information are generated. The annotation results are saved in a unified data format to ensure the consistency of subsequent model training and performance evaluation. Step 1.3: The labeled dataset is divided into training, validation, and test sets in an 8:1:1 ratio for training, parameter tuning, and performance evaluation of the road crack detection model. By constructing a comprehensive dataset covering multiple crack morphologies, a data foundation is provided for the subsequent detection model to achieve stable detection under different crack structure conditions.

[0054] Step 2: Crack feature extraction and modeling based on fusion feature enhancement strategy; In step 2, To address the common characteristics of road cracks, such as large scale span, elongated shape, and strong overall structural continuity, and the shortcomings of existing lightweight road crack detection models in terms of feature representation and overall structural perception, this invention, based on the YOLOv11 detection network, improves the feature extraction and modeling process of the YOLOv11 detection network while maintaining its detection efficiency and lightweight advantages. This results in a lightweight road crack detection method with enhanced feature generation and global perception, named GSC-YOLO. Its network structure diagram is shown below. Figure 2 As shown in the figure, the positions marked with asterisks indicate specific technical improvements made compared to the basic detection network.

[0055] Specifically, the GSC-YOLO network model includes a backbone network, a neck network, and a head detection network connected in sequence.

[0056] The backbone network is used to extract features from the input image. The input image sequentially passes through a first Ghost convolution, a second Ghost convolution, a first C3k2 module, a third Ghost convolution, a second C3k2 module, a fourth Ghost convolution, a third C3k2 module, a fifth Ghost convolution, a fourth C3k2 module, an SPPF_UniRepLK module, and a C2PSA module. The backbone network outputs three different levels of features to the neck network. Specifically, the output of the second C3k2 module is used as the first feature output, the output of the third C3k2 module is used as the second feature output, and the output of the C2PSA module is used as the third feature output.

[0057] The neck network receives features from the backbone network and performs multi-scale fusion. Specifically, the third feature output of the backbone network is upsampled first and then concatenated with the second feature output of the backbone network, and input to the first C3k2 module of the neck network; the output of the first C3k2 module of the neck network is upsampled second and then concatenated with the first feature output of the backbone network, and input to the second C3k2 module of the neck network; the output of the second C3k2 module of the neck network is split into two paths: one path is input to the first CBAM attention module, and the other path is downsampled by the first Ghost convolution and then concatenated with the output of the first C3k2 module of the neck network, and input to the third C3k2 module of the neck network; the output of the third C3k2 module of the neck network is also split into two paths: one path is input to the second CBAM attention module, and the other path is downsampled by the second Ghost convolution and then concatenated with the third feature output of the backbone network, and input to the fourth C3k2 module of the neck network; finally, the output of the fourth C3k2 module of the neck network is input to the third CBAM attention module.

[0058] The head detection network comprises three parallel detection layers, which respectively receive refined features output by the first CBAM attention module, the second CBAM attention module, and the third CBAM attention module in the neck network, and are used to output target detection results at different scales.

[0059] Step 3: Detect the input road crack image, obtain the target location information of the road crack and its corresponding crack category information, and output the crack detection results.

[0060] Example 5 The lightweight feature generation and global perception enhancement road crack detection method of this invention is implemented according to the following steps: Step 1: Acquisition and preprocessing of road crack image dataset; Step 1 is implemented in the following steps: Step 1.1: Obtain road surface image data from multiple publicly available road crack datasets, and fuse the image data to construct a comprehensive road crack image dataset with 4000 images. The road crack image dataset covers three typical road crack types: transverse cracks, longitudinal cracks, and branching cracks. Typical samples are illustrated below. Figure 1 As shown.

[0061] Step 1.2: The road crack images are manually annotated. The crack categories are distinguished according to the main extension direction and overall morphological characteristics of the cracks, and the corresponding crack target location information and category information are generated. The annotation results are saved in a unified data format to ensure the consistency of subsequent model training and performance evaluation. Step 1.3: The labeled dataset is divided into training, validation, and test sets in an 8:1:1 ratio for training, parameter tuning, and performance evaluation of the road crack detection model. By constructing a comprehensive dataset covering multiple crack morphologies, a data foundation is provided for the subsequent detection model to achieve stable detection under different crack structure conditions.

[0062] Step 2: Crack feature extraction and modeling based on fusion feature enhancement strategy; In step 2, To address the common characteristics of road cracks, such as large scale span, elongated shape, and strong overall structural continuity, and the shortcomings of existing lightweight road crack detection models in terms of feature representation and overall structural perception, this invention, based on the YOLOv11 detection network, improves the feature extraction and modeling process of the YOLOv11 detection network while maintaining its detection efficiency and lightweight advantages. This results in a lightweight road crack detection method with enhanced feature generation and global perception, named GSC-YOLO. Its network structure diagram is shown below. Figure 2 As shown in the figure, the positions marked with asterisks indicate specific technical improvements made compared to the basic detection network.

[0063] Specifically, the GSC-YOLO network model includes a backbone network, a neck network, and a head detection network connected in sequence.

[0064] The backbone network is used to extract features from the input image. The input image sequentially passes through a first Ghost convolution, a second Ghost convolution, a first C3k2 module, a third Ghost convolution, a second C3k2 module, a fourth Ghost convolution, a third C3k2 module, a fifth Ghost convolution, a fourth C3k2 module, an SPPF_UniRepLK module, and a C2PSA module. The backbone network outputs three different levels of features to the neck network. Specifically, the output of the second C3k2 module is used as the first feature output, the output of the third C3k2 module is used as the second feature output, and the output of the C2PSA module is used as the third feature output.

[0065] The neck network receives features from the backbone network and performs multi-scale fusion. Specifically, the third feature output of the backbone network is upsampled first and then concatenated with the second feature output of the backbone network, and input to the first C3k2 module of the neck network; the output of the first C3k2 module of the neck network is upsampled second and then concatenated with the first feature output of the backbone network, and input to the second C3k2 module of the neck network; the output of the second C3k2 module of the neck network is split into two paths: one path is input to the first CBAM attention module, and the other path is downsampled by the first Ghost convolution and then concatenated with the output of the first C3k2 module of the neck network, and input to the third C3k2 module of the neck network; the output of the third C3k2 module of the neck network is also split into two paths: one path is input to the second CBAM attention module, and the other path is downsampled by the second Ghost convolution and then concatenated with the third feature output of the backbone network, and input to the fourth C3k2 module of the neck network; finally, the output of the fourth C3k2 module of the neck network is input to the third CBAM attention module.

[0066] The head detection network comprises three parallel detection layers, which respectively receive refined features output by the first CBAM attention module, the second CBAM attention module, and the third CBAM attention module in the neck network, and are used to output target detection results at different scales.

[0067] This invention introduces multiple feature enhancement strategies in the feature extraction and feature fusion stages, improving the ability to express the overall structural features and key local features of road cracks while reducing the model parameter scale and computational complexity. Specifically, these improvements include the following strategies: (1) Lightweight feature generation strategy To address the problem that conventional standard convolutions have a large number of parameters and high computational complexity during feature extraction, which is not conducive to the deployment of road inspection equipment and edge platforms, this invention introduces a lightweight feature generation strategy in the backbone feature extraction stage of the detection network to replace some standard convolutional structures.

[0068] Specifically, the Ghost convolutional module replaces the conventional standard convolutional layers. The Ghost convolutional module effectively reduces computational complexity through a two-stage feature generation process. First, a small number of standard convolutions are used to generate intrinsic feature maps; then, a computationally inexpensive linear transformation operation is applied to the intrinsic feature maps to generate extended feature maps. Finally, the intrinsic feature maps and extended feature maps are concatenated to obtain an output feature map comparable to that of standard convolutions. Its module structure diagram is shown below. Figure 3As shown. Specifically, the input feature map is first processed by standard convolution to generate a feature map. Then, this feature map is divided into two paths: one path is directly preserved as an identity mapping, and the other path is input to a parallel linear operation layer to generate an extended feature map. Finally, the two are concatenated to obtain the final output feature map.

[0069] To illustrate the advantages of the Ghost convolution module over the conventional standard convolution structure in terms of computational complexity, the following comparison of the computational cost and parameter count of the two convolution structures is provided: The computational cost of a standard convolution is: (1) in, The computational cost is the same as that of a regular convolution. The kernel size is [size]. Input the number of channels. The number of convolution kernels and the number of output channels, and To determine the width and height of the output feature map, each convolution kernel generates a feature map when the input image undergoes a convolution operation. These feature maps together form the final multi-channel feature map. Therefore, the number of convolution kernels is equal to the number of channels in the output feature map. The computational cost of the Ghost convolutional module is: (2) (3) in, The computational cost of Ghost convolution. The number of channels in the intrinsic feature map generated by a standard convolution. To increase the multiplier, The size of the convolution kernel for linear operations; The computational cost ratio of regular convolution to Ghost convolution The ratio of parameters They are respectively: (4) (5) As can be seen from the above expression, under the condition of equivalent parameter configuration, the Ghost convolution module can effectively reduce the amount of computation and parameters required for convolution operations. Therefore, through this strategy, the network parameter size and computational overhead can be effectively reduced without significantly reducing the feature expression capability, thereby improving the model running efficiency and providing computational margin for subsequent global perception and attention feature enhancement. (2) Global Awareness Enhancement Strategy To address the problem that traditional spatial pyramid pooling structures, which primarily rely on max pooling, are prone to losing detailed crack information during downsampling, thus limiting the ability to detect small and continuous crack structures, this invention improves the spatial pyramid pooling module in the detection network. Inspired by the UniRepLKNet network, it introduces the concept of large-kernel convolution and proposes a feature fusion structure based on global modeling with large convolutional kernels, namely the SPPF_UniRepLK module. The network structure diagram of UniRepLKNet is shown below. Figure 4 As shown.

[0070] like Figure 4 As shown, UniRepLKNet employs a hierarchical design, with input features sequentially passing through downsampling and stages 1, 2, 3, and 4. Stages 1, 2, and 4 primarily perform conventional feature extraction and downsampling operations at different scales on the feature maps, constructing a multi-scale feature pyramid. Stage 3, as the core feature extraction stage, contains alternating large and small kernel modules. The structure of the large kernel module is as follows: after feature input, the main branch sequentially passes through a dilated reparameterization module and a batch normalization layer; subsequently, the features are split into two paths to generate attention weights, one path undergoes average pooling, ReLU activation, and a Sigmoid function, while the other path is element-wise multiplied with the main branch features; the processed features then pass through a GELU activation function and a batch normalization layer; finally, the output of the large kernel module is obtained through residual connections.

[0071] In this invention, the implementation and specific connection relationship of the SPPF_UniRepLK module are as follows: the input features it receives come from the output of the fourth C3k2 module in the aforementioned backbone network; after the input features enter the SPPF_UniRepLK module, they first pass through a regular convolutional layer, and then are respectively input into three sequentially connected two-dimensional max pooling layers; next, the output of the convolutional layer and the output of the three two-dimensional max pooling layers are concatenated; the concatenated multi-scale fusion features are sent to the UniRepLK module; the final output features serve as the final output of the entire SPPF_UniRepLK module and are directly passed to the C2PSA module of the backbone network for subsequent processing.

[0072] The SPPF_UniRepLK module introduces non-dilated, large-size convolutional kernels to expand the network's effective receptive field while maintaining feature map resolution, thereby enhancing its ability to model the overall extension structure and long-range dependency features of road cracks. Simultaneously, through the synergistic effect of convolutional kernels of different scales, it achieves a joint expression of global semantic features and local detail features, effectively mitigating the loss of crack detail information caused by max pooling operations.

[0073] Furthermore, the module incorporates a feature reparameterization mechanism in its structural design, enabling large convolutional kernels to possess strong feature representation capabilities during the training phase without significantly increasing computational complexity during the inference phase. This allows for both enhanced global perception capabilities and lightweight model while maintaining detection efficiency.

[0074] (3) Attention feature enhancement strategies To address the issue that crack areas in real-world road scenarios are easily affected by factors such as road surface texture and lighting changes, leading to insufficient response of detection models to key crack regions, this invention introduces an attention feature enhancement strategy in the feature extraction and fusion stage of the detection network. Specifically, a CBAM attention module is embedded before the detection layer to adaptively enhance crack-related features and suppress redundant features unrelated to crack detection, thereby improving the model's ability to discriminate road crack targets and its robustness in detection. Its module structure diagram is shown below. Figure 5 As shown.

[0075] The CBAM module consists of a cascaded channel attention module and a spatial attention module. The channel attention module characterizes the importance of different feature channels in the crack detection task, while the spatial attention module characterizes the salience distribution of crack features in spatial location. The structural diagrams of the channel attention module and the spatial attention module are shown below. Figure 6 and Figure 7 As shown. Specifically, the input feature map received by the channel attention module comes from the output of the C3k2 module in the neck network (i.e., the second, third, or fourth C3k2 module of the neck network). Figure 6 As shown, the channel attention module first processes the input feature map The data is processed through max pooling and average pooling layers to extract salient response information and global context information for each channel. Then, the two feature vectors are fed into a multilayer perceptron with shared weights for nonlinear mapping. Next, the two feature vectors output by the multilayer perceptron are added element-wise, and the output feature map of the channel attention module is generated using a sigmoid activation function. That is, the channel attention weights; finally, these weights are compared with the original input feature map. Channel-by-channel multiplication enables adaptive recalibration of key feature channels, resulting in an enhanced feature map. .

[0076] After completing the channel dimension feature enhancement, the implementation path of the spatial attention module is as follows: Figure 7 As shown. The spatial attention module outputs feature maps as the channel attention module. As input, salient region information and overall response intensity are extracted by performing max pooling and average pooling operations along the channel dimension, respectively, and then concatenated along the channel dimension. Subsequently, spatial features are aggregated using convolutional layers, and the spatial attention module output feature map is generated by passing through a sigmoid activation function. That is, the spatial attention weight map; finally, the weight map is... Feature maps input to the spatial attention module Pixel-by-pixel multiplication is used to highlight the spatial location of the crack region, and the final refined features are then fed into the head detection network.

[0077] By employing the aforementioned attention feature enhancement strategy, the detection network can more accurately focus on key channels and spatial regions related to cracks, effectively suppressing interference from background textures and non-crack structures, thereby improving the detection stability and robustness for both small and complex crack morphologies.

[0078] Step 3: Detect the input road crack image, obtain the target location information of the road crack and its corresponding crack category information, and output the crack detection results.

[0079] Example 6 The lightweight feature generation and global perception enhancement road crack detection method of this invention is implemented according to the following steps: Step 1: Acquisition and preprocessing of road crack image dataset; Step 1 is implemented in the following steps: Step 1.1: Obtain road surface image data from multiple publicly available road crack datasets, and fuse the image data to construct a comprehensive road crack image dataset with 4000 images. The road crack image dataset covers three typical road crack types: transverse cracks, longitudinal cracks, and branching cracks. Typical samples are illustrated below. Figure 1 As shown.

[0080] Step 1.2: The road crack images are manually annotated. The crack categories are distinguished according to the main extension direction and overall morphological characteristics of the cracks, and the corresponding crack target location information and category information are generated. The annotation results are saved in a unified data format to ensure the consistency of subsequent model training and performance evaluation. Step 1.3: The labeled dataset is divided into training, validation, and test sets in an 8:1:1 ratio for training, parameter tuning, and performance evaluation of the road crack detection model. By constructing a comprehensive dataset covering multiple crack morphologies, a data foundation is provided for the subsequent detection model to achieve stable detection under different crack structure conditions.

[0081] Step 2: Crack feature extraction and modeling based on fusion feature enhancement strategy; In step 2, To address the common characteristics of road cracks, such as large scale span, elongated shape, and strong overall structural continuity, and the shortcomings of existing lightweight road crack detection models in terms of feature representation and overall structural perception, this invention, based on the YOLOv11 detection network, improves the feature extraction and modeling process of the YOLOv11 detection network while maintaining its detection efficiency and lightweight advantages. This results in a lightweight road crack detection method with enhanced feature generation and global perception, named GSC-YOLO. Its network structure diagram is shown below. Figure 2 As shown in the figure, the positions marked with asterisks indicate specific technical improvements made compared to the basic detection network.

[0082] Specifically, the GSC-YOLO network model includes a backbone network, a neck network, and a head detection network connected in sequence.

[0083] The backbone network is used to extract features from the input image. The input image sequentially passes through a first Ghost convolution, a second Ghost convolution, a first C3k2 module, a third Ghost convolution, a second C3k2 module, a fourth Ghost convolution, a third C3k2 module, a fifth Ghost convolution, a fourth C3k2 module, an SPPF_UniRepLK module, and a C2PSA module. The backbone network outputs three different levels of features to the neck network. Specifically, the output of the second C3k2 module is used as the first feature output, the output of the third C3k2 module is used as the second feature output, and the output of the C2PSA module is used as the third feature output.

[0084] The neck network receives features from the backbone network and performs multi-scale fusion. Specifically, the third feature output of the backbone network is upsampled first and then concatenated with the second feature output of the backbone network, and input to the first C3k2 module of the neck network; the output of the first C3k2 module of the neck network is upsampled second and then concatenated with the first feature output of the backbone network, and input to the second C3k2 module of the neck network; the output of the second C3k2 module of the neck network is split into two paths: one path is input to the first CBAM attention module, and the other path is downsampled by the first Ghost convolution and then concatenated with the output of the first C3k2 module of the neck network, and input to the third C3k2 module of the neck network; the output of the third C3k2 module of the neck network is also split into two paths: one path is input to the second CBAM attention module, and the other path is downsampled by the second Ghost convolution and then concatenated with the third feature output of the backbone network, and input to the fourth C3k2 module of the neck network; finally, the output of the fourth C3k2 module of the neck network is input to the third CBAM attention module.

[0085] The head detection network comprises three parallel detection layers, which respectively receive refined features output by the first CBAM attention module, the second CBAM attention module, and the third CBAM attention module in the neck network, and are used to output target detection results at different scales.

[0086] This invention introduces multiple feature enhancement strategies in the feature extraction and feature fusion stages, improving the ability to express the overall structural features and key local features of road cracks while reducing the model parameter scale and computational complexity. Specifically, these improvements include the following strategies: (1) Lightweight feature generation strategy To address the problem that conventional standard convolutions have a large number of parameters and high computational complexity during feature extraction, which is not conducive to the deployment of road inspection equipment and edge platforms, this invention introduces a lightweight feature generation strategy in the backbone feature extraction stage of the detection network to replace some standard convolutional structures.

[0087] Specifically, the Ghost convolutional module replaces the conventional standard convolutional layers. The Ghost convolutional module effectively reduces computational complexity through a two-stage feature generation process. First, a small number of standard convolutions are used to generate intrinsic feature maps; then, a computationally inexpensive linear transformation operation is applied to the intrinsic feature maps to generate extended feature maps. Finally, the intrinsic feature maps and extended feature maps are concatenated to obtain an output feature map comparable to that of standard convolutions. Its module structure diagram is shown below. Figure 3As shown. Specifically, the input feature map is first processed by standard convolution to generate a feature map. Then, this feature map is divided into two paths: one path is directly preserved as an identity mapping, and the other path is input to a parallel linear operation layer to generate an extended feature map. Finally, the two are concatenated to obtain the final output feature map.

[0088] To illustrate the advantages of the Ghost convolution module over the conventional standard convolution structure in terms of computational complexity, the following comparison of the computational cost and parameter count of the two convolution structures is provided: The computational cost of a standard convolution is: (1) in, The computational cost is the same as that of a regular convolution. The kernel size is [size]. Input the number of channels. The number of convolution kernels and the number of output channels, and To determine the width and height of the output feature map, each convolution kernel generates a feature map when the input image undergoes a convolution operation. These feature maps together form the final multi-channel feature map. Therefore, the number of convolution kernels is equal to the number of channels in the output feature map. The computational complexity of the Ghost convolutional module is: (2) (3) in, The computational cost of Ghost convolution. The number of channels in the intrinsic feature map generated by a standard convolution. To increase the multiplier, The size of the convolution kernel for linear operations; The computational cost ratio of regular convolution to Ghost convolution The ratio of parameters They are respectively: (4) (5) As can be seen from the above expression, under the condition of equivalent parameter configuration, the Ghost convolution module can effectively reduce the amount of computation and parameters required for convolution operations. Therefore, through this strategy, the network parameter size and computational overhead can be effectively reduced without significantly reducing the feature expression capability, thereby improving the model running efficiency and providing computational margin for subsequent global perception and attention feature enhancement. (2) Global Awareness Enhancement Strategy To address the problem that traditional spatial pyramid pooling structures, which primarily rely on max pooling, are prone to losing detailed crack information during downsampling, thus limiting the ability to detect small and continuous crack structures, this invention improves the spatial pyramid pooling module in the detection network. Inspired by the UniRepLKNet network, it introduces the concept of large-kernel convolution and proposes a feature fusion structure based on global modeling with large convolutional kernels, namely the SPPF_UniRepLK module. The network structure diagram of UniRepLKNet is shown below. Figure 4 As shown.

[0089] like Figure 4 As shown, UniRepLKNet employs a hierarchical design, with input features sequentially passing through downsampling and stages 1, 2, 3, and 4. Stages 1, 2, and 4 primarily perform conventional feature extraction and downsampling operations at different scales on the feature maps, constructing a multi-scale feature pyramid. Stage 3, as the core feature extraction stage, contains alternating large and small kernel modules. The structure of the large kernel module is as follows: after feature input, the main branch sequentially passes through a dilated reparameterization module and a batch normalization layer; subsequently, the features are split into two paths to generate attention weights, one path undergoes average pooling, ReLU activation, and a Sigmoid function, while the other path is element-wise multiplied with the main branch features; the processed features then pass through a GELU activation function and a batch normalization layer; finally, the output of the large kernel module is obtained through residual connections.

[0090] In this invention, the implementation and specific connection relationship of the SPPF_UniRepLK module are as follows: the input features it receives come from the output of the fourth C3k2 module in the aforementioned backbone network; after the input features enter the SPPF_UniRepLK module, they first pass through a regular convolutional layer, and then are respectively input into three sequentially connected two-dimensional max pooling layers; next, the output of the convolutional layer and the output of the three two-dimensional max pooling layers are concatenated; the concatenated multi-scale fusion features are sent to the UniRepLK module; the final output features serve as the final output of the entire SPPF_UniRepLK module and are directly passed to the C2PSA module of the backbone network for subsequent processing.

[0091] The SPPF_UniRepLK module introduces non-dilated, large-size convolutional kernels to expand the network's effective receptive field while maintaining feature map resolution, thereby enhancing its ability to model the overall extension structure and long-range dependency features of road cracks. Simultaneously, through the synergistic effect of convolutional kernels of different scales, it achieves a joint expression of global semantic features and local detail features, effectively mitigating the loss of crack detail information caused by max pooling operations.

[0092] Furthermore, the module incorporates a feature reparameterization mechanism in its structural design, enabling large convolutional kernels to possess strong feature representation capabilities during the training phase without significantly increasing computational complexity during the inference phase. This allows for both enhanced global perception capabilities and lightweight model while maintaining detection efficiency.

[0093] (3) Attention feature enhancement strategies To address the issue that crack areas in real-world road scenarios are easily affected by factors such as road surface texture and lighting changes, leading to insufficient response of detection models to key crack regions, this invention introduces an attention feature enhancement strategy in the feature extraction and fusion stage of the detection network. Specifically, a CBAM attention module is embedded before the detection layer to adaptively enhance crack-related features and suppress redundant features unrelated to crack detection, thereby improving the model's ability to discriminate road crack targets and its robustness in detection. Its module structure diagram is shown below. Figure 5 As shown.

[0094] The CBAM module consists of a cascaded channel attention module and a spatial attention module. The channel attention module characterizes the importance of different feature channels in the crack detection task, while the spatial attention module characterizes the salience distribution of crack features in spatial location. The structural diagrams of the channel attention module and the spatial attention module are shown below. Figure 6 and Figure 7 As shown. Specifically, the input feature map received by the channel attention module comes from the output of the C3k2 module in the neck network (i.e., the second, third, or fourth C3k2 module of the neck network). Figure 6 As shown, the channel attention module first processes the input feature map The data is processed through max pooling and average pooling layers to extract salient response information and global context information for each channel. Then, the two feature vectors are fed into a multilayer perceptron with shared weights for nonlinear mapping. Next, the two feature vectors output by the multilayer perceptron are added element-wise, and the output feature map of the channel attention module is generated using a sigmoid activation function. That is, the channel attention weights; finally, these weights are compared with the original input feature map. Channel-by-channel multiplication enables adaptive recalibration of key feature channels, resulting in an enhanced feature map. .

[0095] After completing the channel dimension feature enhancement, the implementation path of the spatial attention module is as follows: Figure 7 As shown. The spatial attention module outputs feature maps as the channel attention module. As input, salient region information and overall response intensity are extracted by performing max pooling and average pooling operations along the channel dimension, respectively, and then concatenated along the channel dimension. Subsequently, spatial features are aggregated using convolutional layers, and the spatial attention module output feature map is generated by passing through a sigmoid activation function. That is, the spatial attention weight map; finally, the weight map is... Feature maps input to the spatial attention module Pixel-by-pixel multiplication is used to highlight the spatial location of the crack region, and the final refined features are then fed into the head detection network.

[0096] By employing the aforementioned attention feature enhancement strategy, the detection network can more accurately focus on key channels and spatial regions related to cracks, effectively suppressing interference from background textures and non-crack structures, thereby improving the detection stability and robustness for both small and complex crack morphologies.

[0097] Step 3: Detect the input road crack image, obtain the target location information of the road crack and its corresponding crack category information, and output the crack detection results.

[0098] Step 3 is implemented in the following steps: The road crack detection model constructed based on the above feature extraction and modeling process performs forward reasoning on the input road crack image and outputs the location information of the crack target and the corresponding crack category. The crack category includes transverse cracks, longitudinal cracks and branch cracks.

[0099] During model training, by reasonably setting the number of training rounds, learning rate, and parameter update strategy, the detection model can complete effective feature learning and parameter convergence within a limited number of training rounds, avoiding underfitting due to insufficient training, while suppressing performance fluctuations and resource waste caused by excessively long training, thereby ensuring the stability of the model training process and detection efficiency.

[0100] Under the same training and validation set conditions, the detection performance was evaluated for different numbers of training epochs. Accuracy was selected. Recall rate Mean precision of all categories ( and Commonly used evaluation metrics in object detection, such as [list of metrics], are used to statistically analyze the changes in the model's detection performance under different training epochs, while also recording the corresponding training time costs. The trends are shown below. Figure 8 As shown.

[0101] Depend on Figure 8It can be seen that as the number of training rounds increases, the various detection performance indicators of the model show a rapid improvement trend in the early stage of training, indicating that the detection model can effectively learn the discrimination features of road crack targets within a relatively small number of training rounds. When the number of training rounds reaches a certain range, the various detection performance indicators gradually stabilize, with small overall fluctuations, indicating that the model has entered a stable convergence state. At this point, further increasing the number of training rounds has limited effect on improving detection accuracy. At the same time, the model training time increases approximately linearly with the number of training rounds. When the number of training rounds is set too large, it will significantly increase the training time cost and reduce the model training efficiency without a significant improvement in detection performance. Considering both detection accuracy and training efficiency, the number of model training rounds is set to 50 rounds, enabling the model to obtain better detection performance under limited training resources.

[0102] Based on the above training strategy, in order to verify the comprehensive advantages of the lightweight feature generation and global perception enhancement road crack detection method (GSC-YOLO) proposed in this invention in terms of detection performance and model efficiency, the detection model is compared and verified with a variety of existing target detection models on the same road crack dataset. The selected comparison models include the mainstream YOLO series detection networks. During the comparative validation process, all detection models used the same dataset, and hyperparameters such as the number of training epochs and learning rate were kept consistent to ensure the fairness and consistency of the comparison results. Precision was used. Recall rate , and The detection performance and computational overhead of each model were comprehensively evaluated based on various metrics, including detection performance indicators, model parameter count, floating-point operation volume, model size, and inference time. The comparison results are shown in Table 1.

[0103] Table 1 Comparison of experimental metrics between the GSC-YOLO algorithm proposed in this invention and various basic object detection models.

[0104] As shown in Table 1, the road crack detection model achieves a relatively balanced overall performance between detection accuracy and computational efficiency. In terms of overall detection performance, the model... and In terms of key indicators, it is at a superior level compared to the comparative model. Meanwhile, its accuracy... With recall rate All indicators remained at a high level, effectively suppressing false detections while ensuring a high detection rate for crack targets, demonstrating good detection stability and reliability.

[0105] In terms of model complexity and computational efficiency, the proposed detection model has significantly lower parameter size and floating-point operation requirements than detection networks with larger parameter sizes, exhibiting better lightweight characteristics and adaptability to edge deployment. Compared to some models that have a slight advantage in detection accuracy but higher computational overhead, the proposed method can effectively reduce computational resource consumption while ensuring detection performance, making it more suitable for resource-constrained application scenarios such as actual road inspection.

[0106] The aforementioned comprehensive performance advantages are mainly attributed to the lightweight feature generation strategy, the global perception enhancement strategy based on large convolutional kernels, and the attention feature enhancement strategy introduced in the feature extraction and fusion stages of the detection model. Through the synergistic effect of multiple feature enhancement mechanisms, the detection model can more accurately model target morphologies such as slender cracks and branch cracks under complex background conditions, thereby improving detection accuracy while maintaining high computational efficiency.

[0107] Furthermore, to verify the advancement of the proposed road crack detection method compared to existing improved crack detection algorithms, it was compared with several improved YOLO models proposed for road defect detection and slender target recognition tasks. These comparison models included G-YOLO, CW-YOLO, CLFI-YOLO, and WSS-YOLO. All models were trained and evaluated using the same dataset and consistent training strategy. The comparison results are shown in Table 2.

[0108] Table 2. Comparison of experimental performance indicators between the GSC-YOLO algorithm proposed in this invention and various improved crack detection algorithms.

[0109] As shown in Table 2, the road crack detection model exhibits significant comprehensive performance advantages across multiple key detection indicators, enabling stable and reliable crack detection. Although it slightly increases in parameter size and computational overhead compared to the lightweight model, the detection model achieves a better balance between improved detection performance and computational resource consumption, making it more suitable for practical road inspection applications requiring high detection accuracy and robustness.

[0110] The above comparative verification results further demonstrate that the road crack detection method with lightweight feature generation and global perception enhancement significantly improves the detection accuracy and stability of complex crack morphologies while ensuring model lightweighting and detection efficiency, and has good engineering application value.

[0111] To verify the technical effectiveness of each improved module in this invention, under the same training strategy and testing conditions, the YOLOv11 detection network was used as the baseline model. Multiple character model variants were constructed by gradually introducing the lightweight feature generation Ghost convolution module, the global perception enhancement SPPF_UniRepLK module, and the CBAM attention feature enhancement module. The detection performance and computational efficiency under different module combinations were compared and verified. The verification results are shown in Table 3.

[0112] Table 3 Ablation analysis of the impact of the improved strategy introduced in this invention on the performance of the proposed GSC-YOLO algorithm.

[0113] The results show that, after introducing the lightweight feature generation module, the model parameter size and computational complexity are effectively reduced and the inference efficiency is significantly improved while the detection accuracy remains basically stable. After introducing the global perception enhancement module, the model's ability to model slender cracks and continuous crack structures is enhanced, and the detection performance at high IOU thresholds is improved. After introducing the attention feature enhancement module, the model's response to key crack regions is enhanced, effectively suppressing background interference and improving detection accuracy.

[0114] When the above modules are introduced and integrated to build a complete detection model, the model achieves high accuracy. Recall rate and The method achieves optimal overall performance across key indicators while maintaining low computational overhead and fast inference speed, demonstrating that the method described in this invention achieves a good balance between detection accuracy and operational efficiency through multi-module collaborative design, thus verifying the effectiveness and rationality of the overall technical solution.

[0115] To verify the robustness and generalization ability of the road crack detection method described in this invention under unknown scenarios, the detection model parameters trained on the original training dataset were transferred to the DeepCrack public road crack dataset for testing. Road crack targets were detected without any model fine-tuning or parameter updates. The test dataset contains complex road backgrounds and various crack morphologies, and is used for model performance evaluation after unified data preprocessing and augmentation operations.

[0116] Table 4. Test metrics of the method proposed in this invention on the DeepCrack dataset.

[0117] The test results are shown in Table 4. The data in the table shows that the method of this invention can accurately identify road crack targets and stably output the location and category information of the cracks, maintaining a high level of overall detection accuracy. Meanwhile, the single-frame inference time is only about 3.7ms (about 270 FPS), indicating low computational overhead during inference, which meets the requirements of real-time detection. This demonstrates that the road crack detection method described in this invention has good adaptability and stability under different data distributions and complex visual environments, making it suitable for widespread application in multi-scenario road inspection tasks.

Claims

1. A lightweight feature generation and global perception enhancement method for road crack detection, characterized in that, The specific steps are as follows: Step 1: Acquisition and preprocessing of road crack image dataset; Step 2: Crack feature extraction and modeling based on fusion feature enhancement strategy; Step 3: Detect the input road crack image, obtain the target location information of the road crack and its corresponding crack category information, and output the crack detection results.

2. The lightweight feature generation and global perception enhancement road crack detection method according to claim 1, characterized in that, Step 1 is implemented in the following steps: Step 1.1: Obtain road surface image data from multiple publicly available road crack datasets, and fuse the image data to construct a comprehensive road crack image dataset. The road crack image dataset covers three typical road crack types: transverse cracks, longitudinal cracks, and branch cracks. Step 1.2: The road crack images are manually annotated. The crack categories are distinguished according to the main extension direction and overall morphological characteristics of the cracks, and the corresponding crack target location information and category information are generated. The annotation results are saved in a unified data format to ensure the consistency of subsequent model training and performance evaluation. Step 1.3: The labeled dataset is divided into training, validation and test sets in an 8:1:1 ratio for training, parameter tuning and performance evaluation of the road crack detection model. By constructing a comprehensive dataset covering multiple types of crack morphology, a data foundation is provided for the subsequent detection model to achieve stable detection under different crack structure conditions.

3. The lightweight feature generation and global perception enhancement road crack detection method according to claim 2, characterized in that, In step 2, based on the YOLOv11 detection network, while maintaining the advantages of detection efficiency and lightweight design, the feature extraction and modeling process of the YOLOv11 detection network is improved, and a road crack detection model GSC-YOLO based on lightweight feature generation and global perception enhancement is proposed. The GSC-YOLO network model includes a backbone network, a neck network and a head detection network connected in sequence.

4. The lightweight feature generation and global perception enhancement road crack detection method according to claim 3, characterized in that, The backbone network in step 2 is used to extract features from the input image. The input image sequentially passes through the first Ghost convolution, the second Ghost convolution, the first C3k2 module, the third Ghost convolution, the second C3k2 module, the fourth Ghost convolution, the third C3k2 module, the fifth Ghost convolution, the fourth C3k2 module, the SPPF_UniRepLK module, and the C2PSA module. The backbone network outputs three different levels of features to the neck network. Specifically, the output of the second C3k2 module is used as the first feature output, the output of the third C3k2 module is used as the second feature output, and the output of the C2PSA module is used as the third feature output.

5. The lightweight feature generation and global perception enhancement road crack detection method according to claim 4, characterized in that, In step 2, the neck network receives features from the backbone network and performs multi-scale fusion. Specifically, the third feature output of the backbone network is upsampled first and then concatenated with the second feature output of the backbone network, and input to the first C3k2 module of the neck network. The output of the first C3k2 module of the neck network is upsampled second and then concatenated with the first feature output of the backbone network, and input to the second C3k2 module of the neck network. The output of the second C3k2 module of the neck network is split into two paths: one path is input to the first CBAM attention module, and the other path is downsampled by the first Ghost convolution and then concatenated with the output of the first C3k2 module of the neck network, and input to the third C3k2 module of the neck network. The output of the third C3k2 module of the neck network is also split into two paths: one path is input to the second CBAM attention module, and the other path is downsampled by the second Ghost convolution and then concatenated with the third feature output of the backbone network, and input to the fourth C3k2 module of the neck network. Finally, the output of the fourth C3k2 module of the neck network is input to the third CBAM attention module. The head detection network in step 2 contains three parallel detection layers, which respectively receive refined features output by the first CBAM attention module, the second CBAM attention module and the third CBAM attention module in the neck network, and are used to output target detection results at different scales.

6. The lightweight feature generation and global perception enhancement road crack detection method according to claim 5, characterized in that, Step 3 specifically includes the following improvement strategies: (1) Lightweight feature generation strategy; (2) Global awareness enhancement strategy; (3) Attention feature enhancement strategies.

7. The lightweight feature generation and global perception enhancement road crack detection method according to claim 6, characterized in that, The lightweight feature generation strategy is as follows: The Ghost convolution module effectively reduces computational complexity through a two-stage feature generation process. First, it uses a small number of standard convolutions to generate eigenmaps. Then, it applies a computationally inexpensive linear transformation to the eigenmaps to generate extended featuremaps. Finally, it concatenates the eigenmaps and extended featuremaps to obtain an output featuremap comparable to that of standard convolution. Specifically, the input featuremap first undergoes standard convolution to generate eigenmaps. These eigenmaps are then divided into two paths: one path preserves them directly as an identity mapping, while the other path inputs them to a parallel linear operation layer to generate extended featuremaps. Finally, the two are concatenated to obtain the final output featuremap. A comparison of the computational cost and parameter count of the two convolutional structures is provided: The computational cost of a standard convolution is: （1） in, The computational cost is the same as that of a regular convolution. The kernel size is [size]. Input the number of channels. The number of convolution kernels and the number of output channels, and To determine the width and height of the output feature map, each convolution kernel generates a feature map when the input image undergoes a convolution operation. These feature maps together form the final multi-channel feature map. Therefore, the number of convolution kernels is equal to the number of channels in the output feature map. The computational cost of the Ghost convolutional module is: （2） （3） in, The computational cost of Ghost convolution. The number of channels in the intrinsic feature map generated by a standard convolution. To increase the multiplier, The size of the convolution kernel for linear operations; The computational cost ratio of regular convolution to Ghost convolution The ratio of parameters They are respectively: （4） （5） As can be seen from the above expression, under the condition of equivalent parameter configuration, the Ghost convolution module can effectively reduce the amount of computation and parameters required for convolution operations. Therefore, through this strategy, the network parameter size and computational overhead can be effectively reduced without significantly reducing the feature representation capability, thereby improving the model running efficiency and providing computational margin for subsequent global perception and attention feature enhancement.

8. The lightweight feature generation and global perception enhancement road crack detection method according to claim 7, characterized in that, The global awareness enhancement strategy is as follows: This paper improves the spatial pyramid pooling module in the detection network by introducing the large kernel convolution concept and proposing a feature fusion structure based on global modeling of large convolution kernels, namely the SPPF_UniRepLK module. UniRepLKNet adopts a hierarchical design, with input features sequentially passing through downsampling and stages 1, 2, 3, and 4. Stages 1, 2, and 4 are mainly used to perform conventional feature extraction and downsampling operations at different scales on the feature maps to construct a multi-scale feature pyramid. Stage 3, as the core feature extraction stage, contains alternating large kernel modules and small kernel modules. The structure of the large kernel module is as follows: after feature input, the main branch sequentially passes through a dilated reparameterization module and a batch normalization layer; then the features are split into two paths to generate attention weights, one path is processed by average pooling, ReLU activation function, and Sigmoid function, and the other path is multiplied element-wise with the main branch features; the processed features then pass through a GELU activation function and a batch normalization layer; finally, the output of the large kernel module is obtained through residual connections. The implementation and specific connection relationship of the SPPF_UniRepLK module are as follows: the input features it receives come from the output of the fourth C3k2 module in the aforementioned backbone network; after the input features enter the SPPF_UniRepLK module, they first pass through a regular convolutional layer, and then are input into three sequentially connected two-dimensional max pooling layers; next, the output of the convolutional layer and the output of the three two-dimensional max pooling layers are concatenated; the concatenated multi-scale fused features are sent to the UniRepLK module; the final output features serve as the final output of the entire SPPF_UniRepLK module and are directly passed to the C2PSA module of the backbone network for subsequent processing.

9. The lightweight feature generation and global perception enhancement method for road crack detection according to claim 8, characterized in that, The attention feature enhancement strategy is as follows: A CBAM attention module is embedded before the detection layer to adaptively enhance crack-related features. The CBAM module consists of a cascaded channel attention module and a spatial attention module. The channel attention module characterizes the importance of different feature channels in the crack detection task, while the spatial attention module characterizes the saliency distribution of crack features in spatial locations. The input feature map received by the channel attention module comes from the output of the C3k2 module in the neck network. The channel attention module first processes the input feature map... The data is processed through max pooling and average pooling layers respectively to extract significant response information and global context information of the channels. Subsequently, the two obtained feature vectors are input into a multilayer perceptron with shared weights for nonlinear mapping; then, the two feature vectors output by the multilayer perceptron are summed element-wise, and the output feature map of the channel attention module is generated by passing a sigmoid activation function. That is, the channel attention weights; finally, these weights are compared with the original input feature map. Channel-by-channel multiplication enables adaptive recalibration of key feature channels, resulting in an enhanced feature map. ; After channel-dimensional feature enhancement, the spatial attention module outputs feature maps as the channel attention module. As input, salient region information and overall response intensity are extracted by performing max pooling and average pooling operations along the channel dimension, respectively, and then concatenated along the channel dimension. Subsequently, spatial features are aggregated using convolutional layers, and the spatial attention module output feature map is generated by passing through a sigmoid activation function. That is, the spatial attention weight map; finally, the weight map is... Feature maps input to the spatial attention module Pixel-by-pixel multiplication is used to highlight the spatial location of the crack region, and the final refined features are then fed into the head detection network.

10. The lightweight feature generation and global perception enhancement method for road crack detection according to claim 9, characterized in that, Step 3 is implemented in the following steps: The road crack detection model constructed based on the above feature extraction and modeling process performs forward reasoning on the input road crack image and outputs the location information of the crack target and the corresponding crack category. The crack category includes transverse cracks, longitudinal cracks and branch cracks.

Images