Model and method for detecting pavement defects

Abstract

The invention discloses a model and a method for pavement defect detection, and belongs to the technical field of target detection, the model is mainly composed of three parts, namely a backbone network, a detection neck and a detection head, the backbone network is an improved CSPDarkNet53, the backbone network comprises four stages, and the main function is feature extraction; the detection neck is an improved path aggregation feature pyramid network (PAFPN) and comprises three feature fusion processes from top to bottom, from bottom to top and from top to bottom; according to the method, the problem of inaccurate target detection in a complex scene in the prior art is solved, and the defects on the road surface can be effectively detected; particularly, the problems that in the prior art, small target detection (small cracks on the road surface) performance is insufficient, complex background interference is caused, and irregularly-shaped cracks are not well processed are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of target detection, and in particular to a model and method for road surface defect detection. Background Art

[0002] Existing pavement defect detection methods usually use traditional target detection algorithms, such as YOLO, SSD, etc. Although these methods have real-time performance and good detection capabilities, they show obvious limitations when dealing with small target detection in complex scenes, complex background interference, and irregular cracks. Traditional methods have deficiencies in the accuracy of small target detection, the ability to suppress complex backgrounds, and the capture of irregular crack shapes, making it difficult to meet the high-precision requirements of actual engineering applications.

[0003] Therefore, there is an urgent need for an improved pavement defect detection method that can achieve high-precision detection of pavement defects in complex scenarios, especially for the detection needs of small targets, complex backgrounds and irregular crack shapes. Summary of the invention

[0004] The technical solution of the present invention is as follows: a model for road surface defect detection is proposed, including a backbone network, a detection neck and a detection head. The backbone network is an improved CSPDarkNet53 (a name of a backbone network), which includes four stages and its main function is feature extraction; the detection neck is an improved path aggregation feature pyramid network (PAFPN), which includes three feature fusion processes from top to bottom, from bottom to top and from top to bottom; the detection head mainly completes classification and regression tasks; wherein a star operation module (SOM) is introduced in CSPDarkNet53, and SOM replaces the ConvModule of the third and fourth stages in CSPDarkNet53; a multi-dimensional auxiliary fusion (MAF) module is introduced in the original PAFPN, and before the feature fusion is performed from the bottom to the top, the features are first weighted by the MAF module, and the output features of the four stages of the backbone network are preliminarily fused, and a wavelet transform convolution (WTC) module is introduced in the detection head, and WTC is used to replace the first two ConvModules in the detection head to reduce the complexity of the model, and then the features are classified and regressed.

[0005] A method for road surface defect detection. After receiving the input road surface defect image, the model first enters the backbone network, which is the upgraded CSPDarkNet53. Specifically, in the third and fourth stages of the backbone network, we replace the traditional ConvModule module with the Star Operation Module (SOM) to enhance the feature extraction capability of the backbone network. Eight consecutive SOMs are applied in the third stage, and five consecutive SOMs are applied in the fourth stage. In the SOM, the deep convolution layer and the normalization layer perform preliminary processing on the features. The spatial and channel dimensions of the features are adjusted through deep convolution (DW-Conv) and batch normalization (BN) to capture more contextual information. After being activated by Relu6 (a variant of the rectified linear unit (Relu)), the features enhance the feature selection capability by element-by-element multiplication, providing high-quality input for subsequent feature aggregation. The key of SOM is to map features to high-dimensional space, which enhances the ability of feature extraction. Then the output features are input into the improved Path Aggregation Feature Pyramid Network (PAFPN) for feature fusion processing. This part mainly performs feature fusion on the output of the backbone network. Compared with the original PAFPN, the improved PAFPN introduces a multi-dimensional auxiliary fusion (MAF) module. MAF processes the output from the backbone network. Here, the output of the backbone is changed, and the output features of the first stage of CSPDarkNet53 are added, which contains more detailed information. MAF allows the detailed information contained in the high-dimensional features to be better integrated with the semantic information in the low-dimensional features, and weights the fused features, which significantly improves the detection effect for small targets. Subsequently, the output features of PAFPN are input into the detection head. The detection head mainly performs classification and regression tasks. We replace the two-layer ConvModule originally contained in the detection head with a Wavelet Transform Convolution (WTC) module. In this part, the input features are transformed by wavelet transform to extract high-frequency and low-frequency components, and then dimensionality reduction operations are performed on them respectively. Small convolution kernels are used to operate on low-dimensional features to obtain a large receptive field and significantly reduce the number of model parameters. After WTC, classification and regression tasks are performed to obtain the output results of the model.

[0006] Furthermore, the star operation module processes specifically as follows: CSPDarkNet53 processes the output features of the second stage, whose dimensions are H×W×C, where H, W, and C represent the length, width, and number of channels of the features, respectively. The H×W×C features are downsampled to H / 2×W / 2×C dimensions.

[0007] Then the features enter the deep convolution layer, that is, each channel has an independent convolution kernel for preliminary feature extraction, and the output dimension remains unchanged, that is, H / 2×W / 2×C.

[0008] After the first deep convolutional layer, batch normalization is performed to keep the feature dimension unchanged.

[0009] The output feature x is then convolved point-wise (1×1 convolution) simultaneously to obtain x1 and x2, which have the same dimension and the number of channels becomes three times the original, which is H / 2×W / 2×3C.

[0010] Apply the Relu6 activation function on x1 to introduce nonlinearity, and the dimension remains the original H / 2×W / 2×3C.

[0011] Then, the activated x1 is multiplied element-wise with x2, so that the activation value of x1 controls the feature of x2 and obtains feature x.

[0012] Next, x is input into a layer of point-by-point convolution (1×1 convolution), the number of channels of x is restored to C, and the output feature dimension is H / 2×W / 2×C. The output feature continues to input a layer of deep convolution, and the dimension remains unchanged, still H / 2×W / 2×C. This layer of deep convolution compresses and refines the features after fusion to further enhance important spatial information, remove redundant features, and maintain low computational complexity.

[0013] Finally, a layer of 1×1 point-by-point convolution is used to change the number of channels of the output feature, and the dimension is changed to H / 2×W / 2×2C. At this point, a complete star operation module operation is completed.

[0014] Furthermore, the third stage applies eight consecutive star operation modules, and the fourth stage applies five consecutive star operation modules.

[0015] Next, the output features are input into the Path Aggregation Feature Pyramid Network (PAFPN) for feature fusion processing, in which a multi-dimensional auxiliary fusion module (MAF) is introduced.

[0016] Furthermore, the multi-dimensional auxiliary fusion module processes specifically as follows: The output features (160×160×64) of the first stage of the backbone network are convolved to (80×80×64) through the MAF module. The output features (80×80×128) of the second stage are concatenated with the features after the convolution operation in the first stage in the channel dimension to become (80×80×192) after the channel attention mechanism. The number of channels is adjusted to 128 after a 1×1 convolution layer horizontally, and the features become (40×40×64) after a convolution layer and a 1×1 convolution layer vertically. The output features (40×40×256) of the third stage are convolved with the channel attention mechanism. It is concatenated with the features fused in the second stage to become (40×40×320), and passes through a 1×1 convolution layer horizontally to adjust the number of channels to 256. It passes through a convolution layer and a 1×1 convolution layer vertically to become (20×20×64). The output features of the fourth stage (20×20×512) maintain the original resolution. After the channel attention mechanism, it is concatenated with the features fused in the three stages in the channel dimension to generate a comprehensive multi-scale feature. The concatenated dimension is (20×20×576), and it passes through a 1×1 convolution layer horizontally to adjust the number of channels to 512.

[0017] Furthermore, in the detection head, the wavelet transform convolution module (WTC) extracts the high-frequency and low-frequency components of the features through wavelet transform and performs dimensionality reduction operations on them respectively.

[0018] The wavelet transform convolution module is configured as follows: The input feature X is first transformed by wavelet transform (WT) and decomposed into four components: low-frequency component X LL (1) , horizontal high frequency component X LH (1) , vertical high frequency component X HL (1) and the diagonal high frequency component X HH (1) , the feature size is reduced to half of the original, the resolution of each component is H / 2×W / 2×C, and the low-frequency component X LL (1) The main structure and global information of the image are retained, and the high-frequency component X LH (1) , X HL (1) , X HH (1) Capture edge and texture detail information, and obtain the low-frequency component X in the first decomposition LL (1) Perform another wavelet decomposition on it to obtain four new components: X LL (2) , X LH (2), X HL (2) , X HH (2) , the size of the newly decomposed feature is (H / 4)×(W / 4)×C, and the convolution operation is performed on each component feature obtained by the first and second wavelet decompositions, respectively. LL (1) , X LH (1) , X HL (1) , X HH (1) , and X LL (2) , X LH (2) , X HL (2) , X HH (2) Perform convolution in sequence to obtain the convolution result Y LL (1) , Y LH (1) , Y HL (1) , Y HH (1) , and Y LL (2) , Y LH (2) , Y HL (2) , Y HH (2) , the convolutional features are fused through inverse wavelet transform (IWT), and the first-level convolution result Y LL (1) , Y LH (1) , Y HL (1) , Y HH (1) , perform inverse wavelet transform and get feature Z (1) , for the secondary convolution result Y LL (2) , Y LH (2) , Y HL (2) , Y HH (2) Perform inverse wavelet transform to obtain feature Z (2) By inverse transforming and fusing the low-frequency and high-frequency components, the generated features contain both the global information of the low-frequency components and the detailed information of the high-frequency components. The feature size after inverse wavelet transform is restored to be consistent with the input. The obtained feature Z (1) and Z (2)The final output feature Z is formed by adding pixels one by one. (0) , the size of the output feature is consistent with the input, that is, H×W×C.

[0019] The beneficial effects of the present invention are: The present invention overcomes the problem of inaccurate target detection in complex scenes in the prior art. The present invention can effectively detect defects on the road surface, especially solving the problems of insufficient performance of small target detection (small cracks on the road surface), complex background interference, and poor handling of irregular-shaped cracks in the prior art. The present invention not only significantly improves the detection accuracy of small targets in the defect detection process of the model, but also enhances its robustness in complex road scenes, and significantly improves the overall performance and reliability of the model when dealing with complex backgrounds and various types of defects. This improvement can ensure accurate detection of defects under various complex road conditions and improve the efficiency and quality of road maintenance work. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is the overall structure diagram of the model.

[0021] Figure 2 This is the structure diagram of the star operation (SOM) module.

[0022] Figure 3 Figure 3 is the spatial attention structure diagram (SA) of the multi-dimensional auxiliary fusion module.

[0023] Figure 4 This is the flow chart of the wavelet transform convolution (WTC) module.

[0024] In the figure: CSPlayer cross-stage partial stacking layer, ConvModule convolution module, SOM star operation module, SPPF spatial pyramid fast pooling layer, SA spatial attention mechanism, Concat splicing layer, Conv1*11×1 convolution layer, Upsample upsampling layer, WTC wavelet transform convolution, Backbone backbone network, Neck detection, Head detection head, Input input, Stage stage, offset offset, hard sigmoid activation function. DETAILED DESCRIPTION

[0025] The purpose of the present invention is to overcome the problem of inaccurate target detection in complex scenes in the prior art, and propose a method for road surface defect detection, which can effectively detect defects on the road surface, especially solve the problems of insufficient performance of small target detection (smaller cracks in the road surface), complex background interference and poor handling of irregular shaped cracks in the prior art. By integrating the Star Operation Module (SOM), Multidimensional Auxiliary Fusion (MAF) module and Wavelet Transform Convolution (WTC) module, the present invention not only significantly improves the detection accuracy of small targets in the defect detection process of the model, but also enhances its robustness in complex road scenes, and significantly improves the overall performance and reliability of the model when dealing with complex backgrounds and various types of defects. This improvement can ensure accurate detection of defects under various complex road conditions and improve the efficiency and quality of road maintenance work.

[0026] The technical solution adopted by the present invention to achieve the above-mentioned purpose is: In a first aspect, a method for road surface defect detection is provided.

[0027] When the network receives the road defect image, it first enters the CSPDarkNet53 backbone network for feature extraction, and a star operation module (SOM) is configured in CSPDarkNet53. Specifically, it replaces the ConvModule modules in the third and fourth stages of the network to enhance the feature extraction capability of the backbone network. The SOM is located in the deep feature extraction stage of the backbone network. Eight consecutive SOMs are applied in the third stage, and five consecutive SOMs are applied in the fourth stage. After experiments, the detection accuracy of the model can be highest when the SOM depth of the two stages is 8 and 5. SOM can map data to a high-dimensional, nonlinear feature space, in which the hidden patterns and complex relationships of the data become more obvious and easier to analyze. Especially when the task is complex, the deep network can extract and combine more abstract and advanced features layer by layer, so that road defect images in complex scenes can be better processed. The output features are then input into the Path Aggregation Feature Pyramid Network (PAFPN) for feature fusion processing. This part mainly performs feature fusion on the output of the backbone network. The Multi-dimensional Auxiliary Fusion (MAF) module is introduced in PAFPN. MAF processes the output from the backbone network. Here, the output of the backbone is changed by adding the output features of the first stage of CSPDarkNet53, which contains more detailed information. MAF allows the detailed information contained in the high-dimensional features to be better integrated with the semantic information in the low-dimensional features, and performs weighted processing on the fused features, which significantly improves the detection effect for small targets. Subsequently, the output features of PAFPN are input into the detection head (YOLOv8-head). The detection head mainly performs classification and regression tasks. However, before classification and regression, the two-layer ConvModule originally contained in YOLOv8 is replaced with a Wavelet Transform Convolution (WTC) module. In this part, the input features are transformed by wavelet transform to extract high-frequency and low-frequency components, and dimensionality reduction operations are performed on them respectively. Small convolution kernels are used to operate on low-dimensional features to obtain a large receptive field, which significantly reduces the number of model parameters. After WTC, classification and regression tasks are performed to obtain the output results of the model.

[0028] First, the output tensor of the second stage of CSPDarkNet53 is processed. The output tensor dimension of the second stage is H×W×C, where H, W, and C represent the length, width, and number of channels of the feature, respectively. The H×W×C tensor is downsampled and the tensor dimension becomes H / 2×W / 2×C.

[0029] The tensor is input into a layer of deep convolution, that is, each channel has an independent convolution kernel for preliminary feature extraction. The deep convolution operation reduces the amount of computation and maintains the extraction of spatial features. The output dimension remains unchanged, that is, H / 2×W / 2×C.

[0030] Batch normalization is enabled after the first layer of depthwise convolution, which helps stabilize the training process, speed up convergence, and prevent the problem of vanishing or exploding gradients. The tensor dimensions remain unchanged.

[0031] Then the output tensor x is convolved point by point (1×1 convolution) at the same time to obtain x1 and x2, which have the same dimensions, but the number of channels is tripled to H / 2×W / 2×3C. The change in the number of channels here can be regarded as a linear combination of each input channel, which is used to fuse information from different channels. By increasing the feature dimension, the model can learn richer expressions and thus capture more complex patterns and features. This is feature amplification and enhancement, expressing the features in a higher-dimensional space. Apply the Relu6 activation function on x1 to introduce nonlinearity, and the dimension is still the original H / 2×W / 2×3C.

[0032] Then, the activated x1 is multiplied element-by-element with x2, so that the activation value of x1 controls the features of x2, and the tensor x is obtained. The dimension remains unchanged in this step. This element-by-element multiplication produces more details and interactions of high-order features in high-dimensional space, helping the model focus on more useful information and ignore irrelevant information.

[0033] Next, x is input into a layer of point-by-point convolution (1×1 convolution), the number of channels of x is restored to C, and the output tensor dimension is H / 2×W / 2×C. Because after the high-dimensional element-by-element multiplication, the feature space contains more complex relationships and high-order information, this point-by-point convolution layer integrates these high-order features and extracts the most important parts, ensuring that subsequent features can retain important representation information while reducing the dimension.

[0034] The output tensor continues to input a layer of deep convolution, and the dimension remains unchanged, still H / 2×W / 2×C. This layer of deep convolution compresses and refines the features after fusion to further enhance important spatial information and remove redundant features while maintaining low computational complexity.

[0035] Finally, a layer of 1×1 convolution is used to change the number of channels of the output tensor, and the dimension is changed to H / 2×W / 2×2C. At this point, a complete star operation module operation is completed.

[0036] The original output of the Backbone of the Yolov8 model is three tensors, which come from the second, third and fourth stages. The backbone has been modified so that the tensors of the first stage are also input into the neck. This adjustment provides more detailed information support for the subsequent multi-dimensional auxiliary fusion module (MAF), which helps to achieve better results in small target detection and complex background suppression. After the modification, the output tensors of the Backbone are expanded from the original three to four. These tensors correspond to the features extracted at different stages. The specific dimensional changes include: The first stage output tensor: Dimensions: (160×160×64) Feature content: The tensor in the first stage contains high-resolution features, which retains a lot of detail information in the image. Adding the output of the first stage to the Neck layer can combine more low-level detail features in subsequent feature fusion, which helps improve the effect of small target detection, especially when detecting subtle defects or small cracks.

[0037] The second stage output tensor: Dimensions: (80×80×128) Feature content: After downsampling once, this tensor contains medium-resolution features and extracts some details and shape information of the image. After inputting this feature into the Neck layer and combining it with the high-resolution features of the first stage, the model's performance in detecting larger objects can be further enhanced, and the interference of complex backgrounds can be effectively suppressed.

[0038] The third stage output tensor: Dimensions: (40×40×256) Feature content: The features at this stage contain strong semantic information and can identify structural features such as the direction and edge of the crack. In complex backgrounds, these semantic features provide strong support to ensure that the model can accurately identify irregularly shaped defects.

[0039] The fourth stage outputs the tensor: Dimensions: (20×20×512) Feature content: This feature describes the global information of the image at the lowest resolution and has a higher level of abstract semantics. It provides the model with an expression of global features, which helps ensure the consistency of detection results. At the Neck layer, the multi-dimensional auxiliary fusion module (MAF) combines the features of the above four stages and completes feature fusion through a series of processing steps, including the following dimensional change process: Feature splicing and fusion: The first-stage features (160×160×64) are convolved to (80×80×64) through the MAF module. The second-stage features (80×80×128) are concatenated with the features after the first-stage convolution operation in the channel dimension after the channel attention mechanism to become (80×80×192). The number of channels is adjusted to 128 after a 1×1 convolution layer horizontally, and the features become (40×40×64) after a convolution layer and a 1×1 convolution layer vertically. The third-stage features (40×40×256) are convolved to (80×80×192) through the channel attention mechanism. The force mechanism is concatenated with the features after the second stage fusion to become (40×40×320), and the number of channels is adjusted to 256 after a 1×1 convolution layer horizontally. The features become (20×20×64) after a convolution layer and a 1×1 convolution layer vertically. The fourth stage features (20×20×512) keep the original resolution, and all features are concatenated in the channel dimension to generate a comprehensive multi-scale feature. The concatenated dimension is (20×20×576), and the number of channels is adjusted to 512 after a 1×1 convolution layer horizontally.

[0040] After the output of stages 2, 3, and 4, a channel attention mechanism is added to dynamically assign the weight of each channel. The weight of low-level features is higher to highlight small targets and detailed information; the weight of high-level features is relatively low to capture global information.

[0041] By introducing the first-stage output tensor and combining it with a multi-dimensional auxiliary fusion module (MAF), the present invention achieves higher flexibility and robustness in feature fusion. The multi-scale feature processing, convolution operation, weighted processing and attention mechanism of the MAF module significantly improve the detection accuracy and reliability of the model, giving it a clear advantage in complex road defect detection tasks, especially in small target detection and suppression of complex background interference.

[0042] The input feature map X is first transformed by wavelet transform (WT) and decomposed into four components: low-frequency component X LL (1) , horizontal high frequency component X LH (1) , vertical high frequency component X HL (1) and the diagonal high frequency component X HH (1) , the feature map size is reduced to half of the original size, and the resolution of each component is H / 2×W / 2×C. LL (1) The main structure and global information of the image are retained, which helps to capture macro features and high-frequency components X LH (1) , X HL (1) , XHH (1) Capturing detailed information such as edges and textures enables the model to perform better in detecting small features such as edges and cracks.

[0043] Next, the low-frequency component X obtained from the first decomposition LL (1) Perform another wavelet decomposition on it to obtain four new components: X LL (2) , X LH (2) , X HL (2) , X HH (2) , the size of the newly decomposed feature map is (H / 4)×(W / 4)×C, which further reduces the spatial resolution. The secondary decomposition provides the feature map with lower resolution and higher abstract information, enabling the model to effectively deal with global complex features and background interference. In the multi-scale irregular crack scene, such decomposition helps the model better identify targets of different scales.

[0044] Then, convolution operations are performed on each component feature map obtained from the first and second wavelet decompositions. LL (1) , X LH (1) , X HL (1) , X HH (1) , and X LL (2) , X LH (2) , X HL (2) , X HH (2) Perform convolution in sequence to obtain the convolution result Y LL (1) , Y LH (1) , Y HL (1) , Y HH (1) , and Y LL (2) , Y LH (2) , Y HL (2) , Y HH (2)In this step, by performing convolution on the low-frequency and high-frequency components respectively, it is possible to enhance the capture of detail features while maintaining the global structural information. After reducing the spatial resolution, the convolution operation using a small convolution kernel can not only cover the necessary receptive field, but also reduce the computational overhead, thereby improving the efficiency of the model.

[0045] The convolution feature map is fused through inverse wavelet transform (IWT). First, the first-level convolution result Y LL (1) , Y LH (1) , Y HL (1) , Y HH (1) , perform inverse wavelet transform and obtain the feature map Z (1) Similarly, for the secondary convolution result Y LL (2) , Y LH (2) , Y HL (2) , Y HH (2) Perform inverse wavelet transform to obtain the feature map Z (2) ,By inversely transforming and fusing the low-frequency and high-frequency components, the generated feature map contains both the global information of the low-frequency components and the detail information of the high-frequency components. The size of the feature map after the inverse wavelet transform is restored to be consistent with the input.

[0046] The obtained feature map Z (1) and Z (2) The final output feature map Z is formed by adding pixels one by one. (0) The size of the output feature map is consistent with the input, that is, (H×W×C). The pixel-by-pixel addition fusion method allows information of different scales to be combined on the same level, thus forming a comprehensive feature map containing multi-scale and multi-frequency, which not only retains the multi-scale features, but also has rich detail information.

[0047] The above is only a specific implementation of the present invention, but the protection scope of the present invention is not limited thereto. Any modification, equivalent replacement and improvement made by any person skilled in the art within the technical scope disclosed by the present invention and within the spirit and principle of the present invention shall be covered within the protection scope of the present invention. At the same time, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. A model for road surface defect detection, comprising: The invention comprises a backbone network, a detection neck and a detection head in four stages, wherein the backbone network is an improved CSPDarkNet53, whose main function is feature extraction, and a star operation module is introduced in the third and fourth stages to replace the original convolution module; the detection neck is an improved path aggregation feature pyramid network, which comprises three feature fusion processes of top-down, bottom-up and top-down, and a multi-dimensional auxiliary fusion module is introduced in the path aggregation feature pyramid network. Before the feature fusion is performed from the bottom to the top, the features are weighted by the multi-dimensional auxiliary fusion module, and the output features of the four stages of the backbone network are preliminarily fused; the detection head mainly completes the classification and regression tasks, and a wavelet transform convolution module is introduced in the detection head. The wavelet transform convolution module is used to replace the first two convolution modules in the detection head to reduce the complexity of the model, and then the features are classified and regressed.

2. A model for road surface defect detection according to claim 1, characterized in that: The third stage applies eight consecutive star operation modules, and the fourth stage applies five consecutive star operation modules.

3. A method for road surface defect detection, characterized in that: The following steps are involved: The S1 model receives the input road defect image and processes it in the star operation module of the backbone network. The deep convolution layer and normalization layer perform preliminary processing on the features, adjust the spatial and channel dimensions of the features through deep convolution and batch normalization, and use the fully connected layer to expand the dimension of the features to capture context information. S2 inputs the features processed in step S1 into the path aggregation feature pyramid network for feature fusion, multi-dimensional auxiliary fusion module processing, and performs weighted processing on the fused features; S3 inputs the features processed by step S2 into the detection head, processes them using the wavelet transform convolution module, performs wavelet transform on the features to extract high-frequency and low-frequency components, performs dimensionality reduction operations on them respectively, performs classification and regression tasks, and obtains the output results of the model.

4. A method for detecting road surface defects according to claim 3, characterized in that: The star operation module is processed as follows: CSPDarkNet53 processes the output features of the second stage. The output feature dimension of the second stage is H×W×C, where H, W, and C represent the length, width, and number of channels of the feature, respectively. The H×W×C feature is downsampled and the dimension becomes H / 2×W / 2×C. Then the feature enters the deep convolution layer, that is, each channel has an independent convolution kernel for preliminary feature extraction, and the output dimension remains unchanged, that is, H / 2×W / 2×C. After the first deep convolution layer, batch normalization is performed, and the feature dimension remains unchanged. Then the output feature x is convolved point by point at the same time to obtain x1 and x2, which have the same dimension, and the number of channels becomes three times the original, which is H / 2×W / 2×3C. The Relu6 activation function is applied to x1 to introduce nonlinearity. The dimension is still the original H / 2×W / 2×3C. Then the activated x1 and x2 are multiplied element by element, so that x1 The activation value controls the feature of x2 to obtain feature x, and x is input into a layer of point-by-point convolution to restore the number of channels of x to C. The output feature dimension is H / 2×W / 2×C. The output feature continues to input a layer of deep convolution, and the dimension remains unchanged, still H / 2×W / 2×C. This layer of deep convolution compresses and refines the features after fusion to further enhance important spatial information and remove redundant features while maintaining low computational complexity. Finally, a layer of 1×1 point-by-point convolution is used to change the number of channels of the output feature, and the dimension changes to H / 2×W / 2×2C. At this point, a complete star operation module operation is completed.

5. A method for road surface defect detection according to claim 3, characterized in that: The processing of the multi-dimensional auxiliary fusion module is as follows: the first-stage output feature of the backbone network 160×160×64 is converted to 80×80×64 through the convolution operation of the multi-dimensional auxiliary fusion module. The second-stage output feature 80×80×128 is converted to 80×80×192 after the channel attention mechanism and the feature after the convolution operation in the first stage is concatenated in the channel dimension. The number of channels is adjusted to 128 after a 1×1 convolution layer horizontally, and the feature becomes 40×40×64 after a convolution layer and a 1×1 convolution layer vertically. The third-stage output feature 40×40×256 is converted to 40×40×256 through the channel attention mechanism. The channel attention mechanism is concatenated with the features after the second-stage fusion to become 40×40×320. It passes through a 1×1 convolution layer horizontally, and the number of channels is adjusted to 256. It passes through a convolution layer and a 1×1 convolution layer vertically, and the features become 20×20×64. The fourth stage outputs the feature 20×20×512 to maintain the original resolution. After the channel attention mechanism, it is concatenated with the features after the three-stage fusion in the channel dimension to generate a comprehensive multi-scale feature. The concatenated dimension is 20×20×576. It passes through a 1×1 convolution layer horizontally to adjust the number of channels to 512.

6. A method for road surface defect detection according to claim 3, characterized in that: The wavelet transform convolution module is specifically processed as follows: the input feature X is first transformed by wavelet transform and decomposed into four components: low frequency component X LL (1) , horizontal high frequency component X LH (1) , vertical high frequency component X HL (1) and the diagonal high frequency component X HH (1) , the feature size is reduced to half of the original, the resolution of each component is H / 2×W / 2×C, and the low-frequency component X LL (1) The main structure and global information of the image are retained, and the high-frequency component X LH (1) , X HL (1) , X HH (1) Capture edge and texture detail information, and obtain the low-frequency component X in the first decomposition LL (1) Perform another wavelet decomposition on it to obtain four new components: X LL (2) , X LH (2) , X HL (2) , X HH (2) , the size of the newly decomposed feature is (H / 4)×(W / 4)×C, and the convolution operation is performed on each component feature obtained by the first and second wavelet decompositions, respectively. LL (1) , X LH (1) , X HL (1) , X HH (1) , and X LL (2) , X LH (2) , X HL (2) , X HH (2) Perform convolution in sequence to obtain the convolution result Y LL (1) , Y LH (1) , Y HL (1) , Y HH (1) , and Y LL (2) , Y LH (2) , Y HL (2) , Y HH (2) , the convolutional features are fused through inverse wavelet transform, and the first-level convolution result Y LL (1) , Y LH (1) , Y HL (1) , Y HH (1) , perform inverse wavelet transform and get feature Z (1) , for the secondary convolution result Y LL (2) , Y LH (2) , Y HL (2) , Y HH (2) Perform inverse wavelet transform to obtain feature Z (2) By inverse transforming and fusing the low-frequency and high-frequency components, the generated features contain both the global information of the low-frequency components and the detailed information of the high-frequency components. The feature size after inverse wavelet transform is restored to be consistent with the input. The obtained feature Z (1) and Z (2) The final output feature Z is formed by adding pixels one by one. (0) , the size of the output feature is consistent with the input, that is, H×W×C.

Images