Road intelligent disease detection method and device based on dynamic multi-scale convolution

Abstract

The invention discloses an intelligent road disease detection method and equipment based on dynamic multi-scale convolution. The method comprises the following steps: performing target category labeling on a training sample set to form a corresponding relationship between sample data and labels; establishing a detection framework including a backbone network, a neck network and a detection network, the backbone network extracting features through convolution layer and cross-stage local layer stacking, introducing spatial pyramid pooling to enhance multi-scale expression, and enhancing fine-grained disease area attention by enhancing spatial attention; the neck network is integrated with dynamic multi-scale convolution and is matched with an upper sampling layer and a cross-stage local layer to fuse high and low layer features; the detection network is provided with a plurality of parallel detection heads, and disease detection is carried out corresponding to different scale feature maps. According to the method, rapid detection and accurate positioning of disease targets of various sizes can be realized, and the detection capability of sparse and small-scale disease targets is remarkably improved through collaborative optimization of a dynamic multi-scale convolution kernel and an enhanced space attention mechanism.

Description

Technical Field

[0001] The present invention belongs to the technical field of natural image processing and relates to a technical solution for intelligent road defect detection based on dynamic multi-scale convolution. Background Art

[0002] Road Defect Detection (RDD) is a key task in smart transportation and intelligent road maintenance systems. Its primary goal is to automatically identify and classify pavement defects (such as cracks, potholes, subsidence, and repair marks) to support road maintenance decisions. Efficient and accurate defect detection is a crucial tool for ensuring road safety and extending road service life. Currently, road defect detection methods can be broadly categorized into two main categories: traditional image processing methods and deep learning-based methods. Traditional methods primarily rely on handcrafted features such as image grayscale, edge information, texture statistics, and geometric shapes, identifying defects using thresholds, filters, or machine learning-based classifiers. However, these methods are often sensitive to changes in lighting, noise, and background, and lack robustness.

[0003] With the rapid development of deep learning technology, particularly the widespread application of convolutional neural networks (CNNs) in image recognition, deep learning-based defect detection methods have significantly improved. These methods can automatically extract multi-level image features, significantly enhancing the ability to detect defects in complex backgrounds. For example, some researchers have employed object detection frameworks such as Faster-RCNN, YOLO, and SSD to train models on pavement image datasets, achieving automatic recognition and localization of common defect types. Furthermore, some studies have introduced strategies such as attention mechanisms and multi-scale feature fusion to further improve the recognition of irregular defects and small defects.

[0004] To address the limitations of traditional CNNs in processing complex damage scenarios, such as insufficient sensitivity to structural details and difficulty adapting to damage variations of varying shapes and scales, researchers have explored introducing new techniques such as deformable convolutions, attention mechanisms, and dynamic convolutions. For example, Zhang et al. proposed a lightweight pavement damage detection method that addresses the numerous model parameters, high computational complexity, and poor real-time performance of traditional deep models when deployed on embedded devices or mobile platforms. By pruning the network structure and optimizing the feature extraction module, this method significantly reduces the model's computational resource consumption while maintaining detection accuracy, providing a feasible solution for edge deployment on small and medium-sized devices in intelligent road inspection systems. Xu et al. compared the performance of Faster R-CNN and Mask R-CNN in pavement crack detection from the perspective of model detection performance. Mask R-CNN possesses stronger instance segmentation capabilities, enabling fine-grained segmentation of crack boundaries, making it more suitable for subsequent damage analysis and maintenance strategy development. Deng et al. proposed a two-stage crack detection method combining YOLOv5 and Res-UNet. This method first uses YOLOv5 to quickly and coarsely locate crack areas, then uses a UNet network optimized with a residual connection structure to perform pixel-level segmentation of the located areas. Experiments have shown that this method balances detection efficiency with segmentation accuracy, making it particularly suitable for typical road damage scenarios such as cracks with complex morphology and blurred edges.

[0005] While the aforementioned methods have achieved good results in typical urban road scenarios, they still face numerous challenges in specialized scenarios such as highways. Defects on highways are sparsely distributed, and defect targets are often small, resulting in the loss of characteristic information about small targets and making accurate identification difficult. This limits the effectiveness of existing methods in deploying real-world road inspection equipment. Therefore, there is an urgent need to design a defect detection method with stronger feature extraction and small target modeling capabilities to improve performance in complex highway environments. Summary of the Invention

[0006] Aiming at the problem that existing road defect detection algorithms are difficult to accurately identify sparse and small-sized defects in highway scenes, the purpose of the present invention is to provide a new road intelligent defect detection technology solution based on dynamic multi-scale convolution.

[0007] The present invention provides a road intelligent disease detection method based on dynamic multi-scale convolution, which includes the following steps: The training sample set is labeled with target categories to form a corresponding relationship between sample data and labels; Establish a detection framework consisting of a backbone network, a neck network, and a detection network. The backbone network extracts features by stacking convolutional layers and cross-stage local layers, introduces a spatial pyramid pooling module to enhance multi-scale expression, and strengthens the focus on fine-grained disease areas by enhancing the spatial attention mechanism; The neck network integrates a dynamic multi-scale convolution module, extracts features through parallel multi-scale convolution kernels, and fuses high- and low-level features with upsampling layers and cross-stage local layers; The detection network is provided with a plurality of parallel detection heads, each corresponding to a feature map of different scales for disease detection; The established detection framework is trained using a training sample set, and intelligent road disease detection is achieved based on the trained detection framework.

[0008] Furthermore, the implementation of the enhanced spatial attention mechanism includes: Perform a convolution operation on the input feature map to generate the first convolution feature; Perform global average pooling and global maximum pooling on the input feature map respectively, and concatenate the results in the channel dimension to form a fusion feature; Processing the fused features through an activation function generates a first activation feature; The first convolution feature and the first activation feature are added element by element and then the second activation feature is generated by the activation function; The second activation feature is multiplied element-wise with the input feature map to output the enhanced feature.

[0009] Furthermore, the implementation of the dynamic multi-scale convolution module includes: Use multiple parallel convolution branches to extract features with convolution kernels of different sizes; Concatenate the output features of each branch in the channel dimension; Apply channel attention mechanism to the splicing features for weighted adjustment; The channel dimension is compressed through a 1×1 convolutional layer and the final fusion features are output.

[0010] Moreover, the convolution kernel sizes of the parallel convolution branches include 1×1, 3×3 and 5×5.

[0011] Furthermore, the neck network performs the following operations: Input the backbone network output features into the dynamic multi-scale convolution module; The processed features are fused with the shallow features of the backbone network through upsampling layers and cross-stage local layers in multiple stages; Generate three sets of feature maps of different scales and input them into the three detection heads of the detection network respectively.

[0012] Moreover, the structure of the backbone network includes sequentially connected convolutional layers, a composite structure consisting of several groups of convolutional layers and cross-stage local layers, a spatial pyramid pooling layer and an enhanced spatial attention layer.

[0013] Furthermore, data enhancement processing is performed on the training sample set, where the data enhancement processing includes at least one of rotation, scaling, cropping, color perturbation, and blurring.

[0014] On the other hand, the present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the intelligent road defect detection method based on dynamic multi-scale convolution as described above is implemented.

[0015] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned intelligent road disease detection method based on dynamic multi-scale convolution.

[0016] On the other hand, the present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the above-mentioned intelligent road disease detection method based on dynamic multi-scale convolution.

[0017] The present invention uses an intelligent road defect detection method based on dynamic multi-scale convolution, and the way it is used is different from other invention patents in that the present invention designs a dynamic multi-scale convolution module and an enhanced spatial attention mechanism module. The dynamic multi-scale convolution module realizes the adaptive expansion of the receptive field by introducing convolution kernels of variable sizes in the convolution process, so that the network can flexibly adjust the processing strategy according to the spatial structure of the input features, and enhance the target detection capability that is sensitive to scale changes; the enhanced spatial attention mechanism strengthens the model's response to the diseased area in the image through weighted learning of the spatial dimension, suppresses irrelevant background interference, and thus improves the significance and accuracy of small target detection, so that the test cases of the present invention are superior to the existing most advanced road defect target detection methods in both subjective and objective results.

[0018] The proposed solution is superior to the road damage target detection method in both qualitative and quantitative evaluations, and can accurately detect different road damages in highway scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The figure is a schematic diagram of the intelligent road defect detection process based on dynamic multi-scale convolution in an embodiment of the present invention.

[0020] Figure 2 This is a network framework diagram of intelligent road defect detection based on dynamic multi-scale convolution in an embodiment of the present invention.

[0021] Figure 3 This is an architecture diagram of the dynamic multi-scale convolution module and enhanced spatial attention module of an embodiment of the present invention.

[0022] Figure 4 The following diagrams show experimental results on a highway road dataset for an embodiment of the present invention. Part (a) shows the original input image; part (b) shows the detection result using YOLOv11; and part (c) shows the detection result of the present invention. DETAILED DESCRIPTION

[0023] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0024] This paper proposes a technical solution for intelligent road defect detection based on dynamic multi-scale convolution to overcome the shortcomings of existing technologies. Because defects in highway scenarios are sparse and small in size, traditional fixed convolution kernels struggle to fully capture the local details of defects, leading to false detections and missed detections. To address the issues of insufficient feature extraction and limited small-target perception, a dynamic multi-scale convolution module is proposed. This module adaptively adjusts the weights and structure of the convolution kernel based on the feature content of the input image, thereby enhancing responsiveness to critical areas and improving the accuracy and robustness of the model for detecting small-scale defects.

[0025] On this basis, the present invention proposes an intelligent road disease detection network based on dynamic multi-scale convolution, comprising: A backbone network for intelligent road defect detection based on dynamic multi-scale convolution was established. This network uses a stack of convolutional layers and cross-stage local layers to gradually extract target features from the image. A spatial pyramid pooling module was then introduced to enhance the multi-scale representation of features. An enhanced spatial attention mechanism was also introduced to strengthen the network's ability to focus on fine-grained defect areas at the spatial level. The neck network of the intelligent road defect detection network based on dynamic multi-scale convolution is established. The neck network mainly integrates dynamic multi-scale convolution modules to achieve adaptive feature extraction of defect targets of different scales. Combined with upsampling layers and cross-stage local layer fusion, it effectively aggregates high-level semantic information and low-level spatial details to improve detection accuracy. A detection network for intelligent road defect detection based on dynamic multi-scale convolution is established. The detection network uses three parallel detection heads, corresponding to feature maps of different scales, to achieve rapid detection of defect targets of various scales.

[0026] Reference Figure 1-Figure 3 The embodiment of the present invention provides an intelligent road defect detection method based on dynamic multi-scale convolution, including the following specific steps: In step a, the original images in the training sample set are first labeled with target categories manually or automatically to form a one-to-one correspondence dataset between images and target category labels, ensuring that all road disease targets in each image have accurate bounding boxes and category labels.

[0027] To improve the model's ability to identify various diseased targets (especially small and sparse targets), various data augmentation processes are performed on the original annotated data. Specifically, these augmentation methods can include operations such as image rotation, scaling, cropping, color perturbation, and blurring. By interfering with image content and geometry, these methods simulate the complex variations found in real-world environments, thereby improving the network's generalization capabilities.

[0028] Step b: Establish a backbone network (i.e., a road intelligent disease detection network) based on dynamic multi-scale convolution. The backbone network uses a stacking method of convolutional layers and cross-stage local layers to gradually extract the target features of the image. Then, a spatial pyramid pooling module is introduced to enhance the multi-scale expression of features. At the same time, an enhanced spatial attention mechanism is introduced to strengthen the network's ability to focus on fine-grained disease areas at the spatial level.

[0029] See also Figure 2 The backbone network of the intelligent road defect detection network based on dynamic multi-scale convolution set in the embodiment of the present invention includes a convolution layer and a cross-stage local layer for extracting image target features, a spatial pyramid pooling layer, and an enhanced spatial attention layer.

[0030] In the preferred embodiment, the image is first input into the first convolutional layer. , perform preliminary feature extraction on the image and input the extracted features into the second convolutional layer Further extract low-level texture and edge features, and then pass the cross-stage local layer Introducing cross-channel information interaction in shallow features to improve expression capabilities, + Construct the first convolutional layer and the cross-stage local layer combination; then transmit it to the second convolutional layer and the cross-stage local layer combination in sequence ( + ), which is used to integrate the feature representations between different channels and continue to input into the combination of the third convolutional layer and the cross-stage local layer ( + ), used to further improve the semantic information fusion capability, input to the combination of the fourth convolutional layer and the cross-stage local layer ( + ), which is used to enhance the multi-scale semantic expression of deep features. Finally, it is input to the spatial pyramid pooling layer and the enhanced spatial attention layer. The spatial pyramid pooling layer uses pooling operations at multiple scales and splices the results to expand the receptive field, enhancing the network's adaptability to diseased targets of different sizes. The enhanced spatial attention layer is used to generate an attention weight map based on the spatial dimension, enhancing the network's response to the diseased area and suppressing background interference. The process of the backbone network can be expressed as: in represents the input image, Indicates the convolutional layers, Indicates the cross-stage local layers, SPP represents the spatial pyramid pooling module, ESA represents the enhanced spatial attention module, Represents the global feature map output by the backbone network.

[0031] See also Figure 3 , this embodiment preferably proposes that the implementation method of enhancing the spatial attention module is as follows: First, the feature map output by the previous layer of the backbone network is input into the first convolutional layer for convolution processing to obtain the first convolution processing result; then, the input feature map is input into the global average pooling layer and the global maximum pooling layer respectively to obtain the global average pooling feature map and the global maximum pooling feature map respectively; then, the global average pooling feature map and the global maximum pooling feature map are spliced ​​in the channel dimension to form a fused feature map; then, the fused feature map is input into the activation function for nonlinear transformation processing to obtain the first activated feature map; the first convolution processing result is added to the first activated feature map element by element to obtain the sum feature map; further, the sum feature map is input into the activation function for nonlinear transformation to obtain the second activated feature map; finally, the second activated feature map is multiplied element by element with the feature map of the initial input to obtain the enhanced feature map As module output. The process can be expressed as:

[0032] in and Represent the global average pooling layer and the global maximum pooling layer respectively, Represents the connection layer, Represents the activation function.

[0033] Step c: Establish the neck network of the intelligent road defect detection network based on dynamic multi-scale convolution. The neck network mainly integrates dynamic multi-scale convolution modules to achieve adaptive feature extraction of defect targets of different scales. Combined with the upsampling layer and cross-stage local layer basis fusion, it effectively aggregates high-level semantic information and low-level spatial details to improve detection accuracy.

[0034] See also Figure 2 In the embodiment of the present invention, the neck network is composed of a dynamic multi-scale convolution module, an upsampling layer, a cross-stage local layer, and a convolution layer in a preset order; wherein, the dynamic multi-scale convolution module extracts multi-scale context information in parallel by introducing convolution kernels of different scales, thereby enhancing the model's ability to express disease targets with significant size differences; the upsampling layer is used to restore the spatial resolution of the feature map, and the cross-stage local layer is used to fuse feature maps of different scales to maintain the continuity of spatial structural information; the multi-scale features are finally sent to the subsequent detection network for target detection. The output features of the backbone network are input into the dynamic multi-scale convolution module; the processed features are multi-level fused with the shallow features of the backbone network through the upsampling layer and the cross-stage local layer; several groups of feature maps of different scales are generated and respectively input into the corresponding detection heads of the detection network.

[0035] The specific process preferably adopted in the embodiment is as follows: the output features of the backbone network (that is, the enhanced spatial attention output ESA) are input to the dynamic multi-scale convolution , and then with the fifth convolutional layer of the backbone network The output is combined and input to the first upsampling layer , and then input into the first connection layer in sequence again , the fifth cross-stage local layer , the second upsampling layer and the second cross-stage local layer of the backbone network Combine the inputs to the second connection layer , the sixth cross-stage local layer , and finally input to the large target detector in the detection network. Similarly, the 5th cross-stage local layer With the 6th convolutional layer Combine the inputs to the 3rd connection layer , the 7th cross-stage local layer , the obtained features are input to the target detector in the detection network. Input to the 7th convolutional layer , the 4th connection layer And combined with the enhanced spatial attention output ESA in the backbone network to input into the 8th cross-stage local layer , the obtained features are input to the small target detector.

[0036] See also Figure 3 , this embodiment preferably proposes that the dynamic multi-scale convolution module is implemented as follows: The dynamic multi-scale convolution module includes multiple parallel convolution branches. Each branch uses convolution kernels of different sizes to extract features from the input feature map F. Convolution kernels of three sizes, 1×1, 3×3, and 5×5, are preferably used to extract context information under different receptive fields. The feature maps output by each branch are spliced ​​in the channel dimension to form a fusion feature. .

[0037] Furthermore, the module introduces a channel attention mechanism after feature concatenation to weight the importance of different channels, thereby dynamically emphasizing key scale features and suppressing redundant or invalid information. To reduce computational complexity, the feature map after the channel attention mechanism is input into a 1×1 convolutional layer for channel compression, and the final fused feature map is output. This fused feature not only maintains multi-scale semantic information but also has channel selectivity enhancement capabilities, significantly improving the model's recognition ability and detection accuracy for diseased targets of different sizes. The process can be expressed as:

[0038] in express The convolution operation, represents the splicing operation of the channel dimension, Represents the corresponding learnable channel attention weight. In the preferred solution adopted in the embodiment, k =1,2,3.

[0039] In step d, a detection network based on dynamic multi-scale convolution is established for intelligent road defect detection. The detection network uses three parallel detection heads, each corresponding to a feature map of different scales, to achieve rapid detection of defect targets of various scales.

[0040] See also Figure 2 The detection network in this embodiment of the present invention consists of three parallel detection heads, each of which processes feature maps from the neck network at different scales. Each detection head includes several convolutional layers and an output layer, which regresses the bounding box position of the diseased object and the confidence score of the corresponding category.

[0041] Steps b, c, and d construct a complete intelligent road defect detection network based on dynamic multi-scale convolution. This network is then trained using the training sample set obtained in step a. In practice, real-time images are fed into the trained network framework to generate real-time detection results.

[0042] In specific implementation, the method proposed in the technical solution of the present invention can be automatically run by those skilled in the art using computer software technology. System devices that implement the method, such as computer-readable storage media that store the corresponding computer program of the technical solution of the present invention and computer equipment that runs the corresponding computer program, should also be within the scope of protection of the present invention.

[0043] The following embodiment describes an electronic device for establishing an intelligent road defect detection method based on dynamic multi-scale convolution provided by the present invention. The electronic device for establishing an intelligent road defect detection method based on dynamic multi-scale convolution described below and the intelligent road defect detection method based on dynamic multi-scale convolution described above can be referenced to each other.

[0044] The electronic device may include a processor, a communications interface, memory, and a communications bus. The processor, communications interface, and memory communicate with each other via the communications bus. The processor may invoke logic instructions in the memory to execute the intelligent road defect detection method based on dynamic multi-scale convolution, which primarily includes the software processing portion of the aforementioned steps.

[0045] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0046] In some possible embodiments, a road intelligent disease detection system based on dynamic multi-scale convolution is provided, comprising the following modules: The first module is used to label the target category of the training sample set to form a corresponding relationship between sample data and labels; The second module is used to establish a detection framework consisting of a backbone network, a neck network, and a detection network. The backbone network extracts features by stacking convolutional layers and cross-stage local layers, introduces a spatial pyramid pooling module to enhance multi-scale expression, and strengthens attention to fine-grained disease areas by enhancing the spatial attention mechanism. The neck network integrates a dynamic multi-scale convolution module, extracts features through parallel multi-scale convolution kernels, and fuses high- and low-level features with upsampling layers and cross-stage local layers. The detection network is equipped with several parallel detection heads, each of which performs disease detection on feature maps of different scales. The third module is used to train the established detection framework using a training sample set, and realize intelligent road disease detection based on the trained detection framework.

[0047] In some possible embodiments, a non-transitory computer-readable storage medium is provided, including a readable storage medium having a computer program stored thereon. When the computer program is executed, the above-mentioned intelligent road defect detection method based on dynamic multi-scale convolution is implemented.

[0048] In some possible embodiments, a computer program product is provided, including a computer program, which, when executed by a processor, implements the above-mentioned intelligent road disease detection method based on dynamic multi-scale convolution.

[0049] Reference Figure 4 , the application test experiment of the embodiment of the present invention is as follows: The present invention is compared with other road disease detection methods, and experimental data are provided to demonstrate the effectiveness of the present invention. The evaluation indicators of the comparative experimental results are shown in Table 1 below. Figure 4 As shown in FIG, part (a) is the original image; part (b) is the detection result image using YOLOv11; and part (c) is the detection result image of the present invention.

[0050] Table 1

[0051] It can be seen from the experimental data in the above table that the method of the present invention achieves higher accuracy than the comparative method, that is, it is superior to the comparative method.

[0052] In summary, the technical solution for intelligent road disease detection based on dynamic multi-scale convolution proposed in the present invention first constructs a unified detection framework including a backbone network, a neck network and a detection network, wherein the backbone network extracts global feature information from the image by stacking convolution layers, cross-stage local layers, spatial pyramid pooling layers and enhanced spatial attention modules; the neck network integrates dynamic multi-scale convolution modules, upsampling layers and cross-stage local layers, further fusing feature representations at different levels to achieve fine-grained modeling of multi-scale disease targets; the detection network adopts three parallel detection heads, corresponding to feature maps of different scales, to adapt to the detection requirements of disease targets of various sizes. Through the design of dynamic multi-scale convolution kernels and the introduction of enhanced spatial attention mechanism, the present invention can effectively improve the detection performance of small-sized and sparse targets in road disease detection tasks. The network proposed in the present invention is superior to other existing road disease detection methods in both qualitative analysis and quantitative index evaluation.

[0053] Compared to existing methods, this invention offers advantages and positive effects: It proposes an intelligent road defect detection method suitable for highway scenarios. This method utilizes a dynamic multi-scale convolution module to adaptively extract features of defects at different scales. It also incorporates an enhanced spatial attention module to focus on key regional features, maintaining high detection accuracy even in complex environments. The proposed method has been validated using both subjective and objective evaluation metrics on a highway road defect dataset.

[0054] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Persons skilled in the art may make various modifications, additions, or substitutions to the described specific embodiments without departing from the spirit of the present invention or exceeding the scope of the appended claims.

Claims

1. A road intelligent disease detection method based on dynamic multi-scale convolution, characterized by: The following processes are included: The training sample set is labeled with target categories to form a corresponding relationship between sample data and labels; Establish a detection framework consisting of a backbone network, a neck network, and a detection network. The backbone network extracts features by stacking convolutional layers and cross-stage local layers, introduces a spatial pyramid pooling module to enhance multi-scale expression, and strengthens the focus on fine-grained disease areas by enhancing the spatial attention mechanism; The neck network integrates a dynamic multi-scale convolution module, extracts features through parallel multi-scale convolution kernels, and fuses high- and low-level features with upsampling layers and cross-stage local layers; The detection network is provided with a plurality of parallel detection heads, each corresponding to a feature map of different scales for disease detection; The established detection framework is trained using a training sample set, and intelligent road disease detection is achieved based on the trained detection framework.

2. The method according to claim 1, characterized in that The implementation of the enhanced spatial attention mechanism includes: Perform a convolution operation on the input feature map to generate the first convolution feature; Perform global average pooling and global maximum pooling on the input feature map respectively, and concatenate the results in the channel dimension to form a fusion feature; Processing the fused features through an activation function generates a first activation feature; The first convolution feature and the first activation feature are added element by element and then the second activation feature is generated by the activation function; The second activation feature is multiplied element-wise with the input feature map to output the enhanced feature.

3. The method according to claim 1, characterized in that The implementation of the dynamic multi-scale convolution module includes: Use multiple parallel convolution branches to extract features with convolution kernels of different sizes; Concatenate the output features of each branch in the channel dimension; Apply channel attention mechanism to the splicing features for weighted adjustment; The channel dimension is compressed through a 1×1 convolutional layer and the final fusion features are output.

4. The method according to claim 3, characterized in that The convolution kernel sizes of the parallel convolution branches include 1×1, 3×3, and 5×5.

5. The method according to claim 1, wherein The neck network performs the following operations: Input the backbone network output features into the dynamic multi-scale convolution module; The processed features are fused with the shallow features of the backbone network through upsampling layers and cross-stage local layers in multiple stages; Generate three sets of feature maps of different scales and input them into the three detection heads of the detection network respectively.

6. The method according to claim 1, characterized in that The structure of the backbone network includes sequentially connected convolutional layers, a composite structure consisting of several groups of convolutional layers and cross-stage local layers, a spatial pyramid pooling layer and an enhanced spatial attention layer.

7. The method according to claim 1, characterized in that Data enhancement processing is performed on the training sample set, where the data enhancement processing includes at least one of rotation, scaling, cropping, color perturbation, and blurring.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the intelligent road defect detection method based on dynamic multi-scale convolution as described in any one of claims 1 to 7 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for intelligent road defect detection based on dynamic multi-scale convolution as claimed in any one of claims 1 to 7 is implemented.

10. A computer program product comprising a computer program, characterized in that: When the computer program is executed by a processor, the method for intelligent road defect detection based on dynamic multi-scale convolution as claimed in any one of claims 1 to 7 is implemented.

Images