A method and system for detecting road surface diseases through cross-level multi-scale feature fusion

Abstract

The application provides a kind of cross-layer multi-scale feature fusion pavement disease detection method and system, belong to computer vision and image processing technical field.The method includes: the input pavement disease image is extracted by the multi-scale feature extraction backbone network based on Res2Net architecture, and the multi-scale feature map containing different semantic levels is outputted;Through multi-level feature fusion neck network, the multi-scale feature map is cross-layer bidirectional feature fusion, the multi-scale information aggregation of deep semantic features and shallow detail features is carried out, and the multi-level feature map after fusion enhancement is outputted;The multi-level feature map is sent into detection head, and the positive and negative sample division threshold is adaptively adjusted by combining dynamic label allocation strategy, and the pavement disease target is detected and positioned.The application effectively solves the problem of detection performance decline caused by similar disease scale span, edge feature degradation and background noise interference, and improves the accuracy and robustness of pavement disease detection.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing technology, and more particularly to a method and system for detecting road surface defects through cross-level and multi-scale feature fusion. Background Technology

[0002] Road surface defect detection and localization technology is a core component of intelligent operation and maintenance of road infrastructure, and its detection accuracy directly affects the timeliness and economy of maintenance decisions. In recent years, deep learning-based target detection algorithms, with their adaptive feature learning and ability to adapt to complex environments, have gradually replaced traditional image processing methods as the mainstream technology. Based on differences in the detection process, existing algorithms can be divided into two main categories: single-stage detection (such as the YOLO series, SSD, etc.) and two-stage detection (such as Faster R-CNN, Mask R-CNN, etc.). Both use convolutional neural networks to extract multi-level features to achieve defect identification. However, the unique characteristics of road surface scenarios mean that algorithms must simultaneously address multiple challenges, including multi-scale targets, high interference noise, and low-quality images.

[0003] Current mainstream methods suffer from three significant limitations: First, similar defects (such as cracks) exhibit large pixel-scale variations due to differences in imaging distance, making it difficult for single-level features to simultaneously address the detection needs of both fine cracks and wide, network-like cracks, leading to missed detections or misjudgments. Second, interfering substances such as water stains and oil stains share similar edge texture features with real defects, and existing algorithms lack effective discrimination mechanisms during feature extraction, easily resulting in false positives. Third, motion blur and low-resolution imaging from inspection equipment exacerbate the degradation of defect edge features, particularly damaging the morphological integrity of millimeter-level cracks, and existing methods lack robustness to such degraded features. These shortcomings collectively restrict the detection accuracy and generalization performance in practical engineering scenarios. Summary of the Invention

[0004] This invention addresses the technical problems in existing pavement distress detection technologies, such as inaccurate detection due to large scale differences in similar distresses, performance degradation caused by noise similar to distress edge features (e.g., water stains and shadows), and degradation of distress edge features due to motion blur of inspection equipment and limited image resolution. It provides a pavement distress detection method and system based on cross-level, multi-scale feature fusion. This invention primarily utilizes a multi-scale feature extraction backbone network based on the Res2Net architecture for hierarchical feature extraction, performs cross-level bidirectional feature fusion through a multi-level feature fusion neck network, and adaptively adjusts the positive and negative sample division threshold using a dynamic label allocation strategy. This enhances the model's ability to extract features from distresses at different scales, improves information exchange between deep and shallow features, filters surface similarity noise, and improves the detection accuracy of edge-degraded targets, ultimately improving the accuracy and robustness of pavement distress detection.

[0005] The technical means employed in this invention are as follows: A method for detecting pavement defects through cross-level, multi-scale feature fusion includes: S1. The input road surface distress image is processed by a multi-scale feature extraction backbone network based on the Res2Net architecture to extract hierarchical features and output a multi-scale feature map containing different semantic levels. S2. The multi-level feature fusion neck network is used to perform cross-level bidirectional feature fusion on the output multi-scale feature map, and the deep semantic features and shallow detail features are aggregated at multiple scales to output the fused and enhanced multi-level feature map. S3. The output multi-level feature map is sent to the detection head, and the positive and negative sample division threshold is adaptively adjusted in combination with the dynamic label allocation strategy to detect and locate road surface defects and output the detection results.

[0006] Further, step S1 includes: S11, through three consecutive 3 3. Convolution is used to extract features from the input road surface distress image and downsample the image resolution to 1 / 4 of the original resolution to output the initial feature map; S12. Configure the channel width of each Res2Net module and calculate the channel width of each Res2Net module. The formula is as follows:

[0007] in, Indicates the number of input channels. and The base width and number of groups for each group are set to 4 and 8 respectively; S13. The initial feature map is sequentially input into multiple Res2Net modules with configured channel widths for hierarchical feature extraction, outputting multi-scale feature maps of four semantic levels: P2, P3, P4, and P5. Each Res2Net module uniformly maps the input feature map along the channel dimension. Each feature mapping subset is used with... express, Each Having the same spatial size and the same number of channels as the original number of channels The feature subsets are processed using a hierarchical residual connection method. 3. Convolution processing ensures that each feature subset receives feature information from the preceding feature subset, and the output after residual connection... satisfy:

[0008] in, Indicates 3 3 convolution operations, each feature subset is processed The output results have a larger receptive field. The overall module output contains different numbers and combinations of receptive field sizes, and the output contains multi-scale feature maps with different receptive field combinations. S14. The output P5 level highest-level features are aggregated through multi-scale pooling using the spatial pyramid pooling module to output the aggregated high-level semantic feature map.

[0009] Furthermore, step S13 includes the following five stages: Stage 0: The initial feature map is processed by 3×3 depthwise convolution and 3×3 max pooling, and the output is a Stem-level feature map with a resolution of H / 4×W / 4 and 64 channels. Stage 1: By stacking 3 Bottle2neck modules [1×1,128; 3×3,32×4; 1×1,256], feature extraction is performed on the feature map output from Stage 0, and the output is a P2 level feature map with a resolution of H / 4×W / 4 and 256 channels. Stage 2: Features are extracted from the feature map output from Stage 1 by stacking 4 Bottle2neck modules [1×1,256; 3×3,64×4; 1×1,512], and the output is a P3 level feature map with a resolution of H / 8×W / 8 and 512 channels. Stage 3: Features are extracted from the feature map output from Stage 2 by stacking 6 Bottle2neck modules [1×1,512; 3×3,128×4; 1×1,1024], and the output is a P4 level feature map with a resolution of H / 16×W / 16 and 1024 channels. Stage 4: By stacking 3 Bottle2neck modules [1×1,1024; 3×3,256×4; 1×1,2048], feature extraction is performed on the feature map output from Stage 3, and the output is a P5 level feature map with a resolution of H / 32×W / 32 and 2048 channels.

[0010] Further, in step S2, the multi-level feature fusion neck network includes a top-down feature fusion path and a bottom-up feature fusion path, wherein: The top-down feature fusion path includes: S2a1. The enhanced feature alignment module transforms the dimensions of the output P5 level features and inputs them as high-level features into the cross-level fusion module. S2a2, The P5 level features after dimensionality transformation are input together with the P3 and P4 level features output by the backbone network into the cross-level fusion module for multi-scale feature fusion, and the fused P4 level features are output; wherein, the cross-level fusion module first performs a unified feature alignment operation on the features of the three different levels respectively:

[0011] in, Indicates the first The output features of each hierarchical feature after feature alignment operation; Indicates containing 1 1. Feature alignment operation of projection convolution and resampling operators to ensure that all output features They have the same dimensions and size; Indicates the input number of the first... Each level of feature; This represents a hierarchical index, with values... (lower floors) (Middle level) (High-level); S2a3: Concatenate the aligned features and reassemble them using a 1×1 convolution:

[0012] in, This represents the final output feature of the cross-level fusion module; This represents a 1×1 convolution operation; Indicates feature concatenation operation; This represents the aligned low-level features; This represents the aligned mid-level features; This represents the aligned high-level features; S2a4. After adjusting the channels of the fused P4-level features through a 1×1 convolution, the features are enhanced by the reparameterized feature enhancement module, and the enhanced P4-level feature map is output. S2a5: The enhanced P4 level features are transformed in dimension by the enhanced feature alignment module and input into the next cross-level fusion module, where they are fused with the P1 / P2 and P3 level features output by the backbone network, and the fused P3 level features are output. S2a6. After the fused P3-level features are processed by 1×1 convolution, the P3-level feature map is output through the reparameterized feature enhancement module and used as the input of the small target detection head.

[0013] The bottom-up feature fusion path includes: S2b1: The output P3 level features are downsampled by a convolution with a stride of 2 through the enhanced feature alignment module, reducing the scale to P4 size. S2b2: The downsampled P3-level features are concatenated and fused with the output P4-level features, and then processed by the reparameterized feature enhancement module to output the fused and enhanced P4-level feature map as the input of the mid-target detection head; S2b3: The output P4 level features are downsampled by convolution with a stride of 2 through the enhanced feature alignment module, reducing the scale to P5 size; S2b4. The downsampled P4-level features are concatenated and fused with the output P5-level features, and then processed by the reparameterized feature enhancement module to output the fused and enhanced P5-level feature map as the input of the large target detection head.

[0014] Furthermore, in the top-down feature fusion path and the bottom-up feature fusion path, the reparameterized feature enhancement module specifically performs feature enhancement including: During the training phase, a residual topology structure consisting of three parallel branches—a 3×3 main convolution branch, a 1×1 auxiliary convolution branch, and an identity branch—is used to extract multi-band features from the input features. During the inference deployment phase, leveraging the linear properties of convolutional layers, the parameters of the batch normalization layers in each branch are fused into the convolutional kernels of each branch. The batch normalization output is represented as:

[0015] in, This indicates the batch normalization operator; Indicates the input tensor; Indicates the mean of channel characteristics; Indicates the standard deviation of channel characteristics; Indicates the scaling parameter; Indicates the offset parameter; The equivalent convolutional kernel weights and biases after fusion are as follows:

[0016]

[0017] in, This represents the equivalent convolution kernel weight tensor after incorporating batch normalized parameters; Represents the original convolution kernel weight tensor; Indicates the first Scaling parameters for each output channel; Indicates the first The characteristic standard deviation of each output channel; This represents the local minimum constant, used to prevent the denominator from being zero; This represents the equivalent bias value after fusion; Indicates the first The characteristic mean of each output channel; Indicates the first Offset parameters for each output channel; The multi-branch structure is transformed into an equivalent single-path linear operator structure, where the equivalent convolution kernel weights and biases are the sum of the parameters after the branches are fused, as shown in the following formula:

[0018]

[0019] in, This represents the final equivalent convolution kernel weight tensor; This represents the weights of the 3×3 main branches after batch normalization. This represents the weight after merging the 1×1 auxiliary branches; This represents the weight after merging the identity branches; This represents the final equivalent bias vector; This represents the bias after merging the 3×3 main branches; This represents the bias after merging 1×1 auxiliary branches; This represents the bias after merging identity branches; Utilize the obtained equivalent single-path linear operator structure for input features Perform inference calculations and output the enhanced features:

[0020] in, This indicates a convolution operation.

[0021] Furthermore, in the top-down feature fusion path and the bottom-up feature fusion path, the dimensional transformation through the enhanced feature alignment module specifically includes: Local semantic information is extracted using a 1×1 standard convolution while preserving original spatial detail features, resulting in an output channel count of [number missing]. Basic feature map :

[0022] in, This represents a 1×1 standard convolution; Indicates the input tensor; Nonlinear evolution of feature representation is achieved through a cascaded sequence of convolutional operators. First, 3×3 convolutions are used to map high-dimensional information between channels and establish cross-channel semantic dependencies. Then, depthwise separable convolutions decouple the correlation between spatial and channel dimensions. Finally, the signal is processed by the GELU activation function, resulting in an output with [number of channels]. Texture feature map :

[0023] in, Represents the activation function of the Gaussian error linear unit; The weights represent the depthwise separable convolutions; Represents the weights of a 3×3 dense convolution; Indicates the convolution operation; Basic feature map and Channel splicing is performed to obtain the fused spliced ​​feature map. :

[0024] in, This indicates a channel-level concatenation operation; splicing feature map The reshaping and shuffle operations are performed to uniformly resample the feature space, outputting the final enhanced feature-aligned feature map. :

[0025] in, This indicates a Shuffle operation, which is an operation that swaps positions within a group; This indicates a reshaping operation, which converts the spliced ​​feature map into two sets of feature structures consisting of the outputs of two branches.

[0026] Further, step S3 includes: S31. For each ground truth bounding box, select the frontmost bounding box whose center point is closest to the ground truth bounding box in each feature layer. One candidate anchor box; S32. Classification score based on network prediction The intersection-union ratio of candidate anchor boxes and true labeled boxes. Weighted processing is performed, and the crossover ratio (CRO) of the merged data is calculated. The mean of the CRO is then calculated. satisfy:

[0027] in, Indicates the number of candidate anchor boxes; S33. Calculate the standard deviation of the cross-union ratio (CUNR), and use the sum of the mean and standard deviation as the dynamic CUNR threshold. :

[0028] in, Indicates the first The crossover ratio of the candidate anchor boxes; S34. Select anchor boxes with a fusion intersection-union ratio higher than the dynamic threshold and whose center point falls within the ground truth bounding box as positive samples; if an anchor box corresponds to multiple ground truth bounding boxes, select the ground truth bounding box with the largest fusion intersection-union ratio for matching; S35. Based on the filtered positive samples, the detection head is used to classify the road surface defects and perform bounding box regression to obtain the final predicted box score. Output the final detection results, including the score of the final predicted bounding box. The formula is as follows:

[0029] in, One-hot encoding representing the actual label; This represents the intersection-union ratio (IoU) between the predicted bounding box and the actual labeled bounding box.

[0030] This invention also provides a pavement distress detection system based on the aforementioned cross-level multi-scale feature fusion method, comprising: a multi-scale feature extraction module, a multi-level feature fusion module, and a dynamic label allocation and detection module, wherein: The multi-scale feature extraction module is used to extract hierarchical features from the input road surface distress image through a multi-scale feature extraction backbone network based on the Res2Net architecture, and output a multi-scale feature map containing different semantic levels. The multi-level feature fusion module is used to perform cross-level bidirectional feature fusion on the output multi-scale feature map through the multi-level feature fusion neck network, and to aggregate deep semantic features and shallow detail features at multiple scales to output a fused and enhanced multi-level feature map. The dynamic label allocation and detection module is used to send the output multi-level feature map into the detection head, and adaptively adjust the positive and negative sample division threshold in combination with the dynamic label allocation strategy to detect and locate road surface defects and output the detection results.

[0031] Compared with the prior art, the present invention has the following advantages: 1. This invention provides a road surface defect detection method based on cross-level and multi-scale feature fusion. By constructing a multi-scale feature extraction backbone network based on the Res2Net architecture, the input feature map is uniformly divided into multiple feature mapping subsets along the channel dimension. Multi-scale feature extraction is performed using a hierarchical residual connection method, which realizes the adaptability to defects of different sizes (from fine cracks to large potholes). It effectively alleviates the feature degradation problems caused by motion blur and insufficient image resolution, while preserving the fine-grained edge texture information of road surface defects and enhancing the recognition ability of targets of different scales and sizes.

[0032] 2. This invention provides a road surface defect detection method based on cross-level and multi-scale feature fusion. It performs cross-level bidirectional feature fusion through a multi-level feature fusion neck network, adopts a top-down path to simultaneously receive features from three different levels (high, medium, and low) for multi-scale fusion, and uses a bottom-up path for downsampling and splicing fusion. Combined with a reparameterized feature enhancement module and an enhanced feature alignment module, it achieves effective aggregation of deep semantic information and shallow detail information, eliminates the problem of defect target information loss caused by multiple downsampling, enhances the ability to distinguish background noise, and significantly improves the detection accuracy of defects with large scale spans and the detection ability of textured targets.

[0033] 3. The present invention provides a road surface defect detection method based on cross-level and multi-scale feature fusion. It combines a dynamic label allocation strategy to directly combine the intersection-union ratio (IU) of the predicted bounding box and the ground truth box, as well as the IU of the anchor box and the ground truth box. An adaptive dynamic threshold is obtained based on the mean and standard deviation of the fused IU, which realizes the adaptive adjustment of the threshold for dividing positive and negative samples. This overcomes the anchor box matching problem caused by edge feature degradation due to ambiguity, promotes the model to learn higher quality matching, and improves the detection efficiency and attention capability of positive samples.

[0034] 4. The present invention provides a road surface defect detection method based on cross-level and multi-scale feature fusion. During the training phase, the network feature space is expanded by a three-branch parallel residual topology through a structure reparameterization feature enhancement module. During the inference phase, the linearity of the convolutional layer is used to convert the multi-branch structure into an equivalent single-path linear operator structure. This effectively decouples the deep multi-band feature extraction during the training phase from the computational overhead elimination during the inference phase, thereby improving the deployment efficiency of the model while ensuring detection accuracy.

[0035] 5. The present invention provides a road surface defect detection method based on cross-level and multi-scale feature fusion. It integrates basic feature extraction branches and texture feature extraction branches through an enhanced feature alignment module, extracts local semantic information using standard convolution and enhances texture pattern representation using depthwise separable convolution, and performs uniform resampling of the feature space through a shuffle operation. This achieves the preservation of inter-channel correlation during downsampling, optimizes the feature extraction efficiency and representation quality of the convolutional neural network, and improves the texture extraction capability of the features.

[0036] In summary, by applying the technical solution of this invention, existing technologies address problems such as inaccurate detection due to large scale differences in similar road surface defects, decreased detection performance due to noise similar to defect edge features (e.g., water stains and shadows), and degradation of defect edge features due to motion blur of inspection equipment and limited image resolution. These problems are effectively solved through multi-scale feature extraction, cross-level bidirectional feature fusion, and dynamic label allocation strategies. Therefore, the technical solution of this invention solves the problems of insufficient accuracy in road surface defect detection under complex background interference, poor adaptability to targets of different scales, and high rates of missed and false detections of edge-degraded targets in existing technologies.

[0037] Based on the above reasons, this invention can be widely applied in fields such as intelligent inspection of road infrastructure, drone aerial road surface detection, autonomous driving road condition perception, and intelligent transportation system maintenance decision-making. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a structural diagram of the road surface defect detection method based on cross-level and multi-scale feature fusion of the present invention.

[0040] Figure 2 This is a structural diagram of the Res2Net module of the present invention.

[0041] Figure 3 This is a structural diagram of the cross-level fusion module of the present invention.

[0042] Figure 4 This is a structural diagram of the structured reparameterized feature enhancement module of the present invention.

[0043] Figure 5 This is a diagram of the enhanced feature alignment module of the present invention.

[0044] Figure 6 This is a structural diagram of the dynamic tag allocation strategy of the present invention. Detailed Implementation

[0045] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0046] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product or device.

[0047] This invention provides a method for detecting pavement defects by fusing features across multiple levels and scales, including: S1. The input road surface distress image is processed by a multi-scale feature extraction backbone network based on the Res2Net architecture to extract hierarchical features and output a multi-scale feature map containing different semantic levels. S2. The multi-level feature fusion neck network is used to perform cross-level bidirectional feature fusion on the output multi-scale feature map, and the deep semantic features and shallow detail features are aggregated at multiple scales to output the fused and enhanced multi-level feature map. S3. The output multi-level feature map is sent to the detection head, and the positive and negative sample division threshold is adaptively adjusted in combination with the dynamic label allocation strategy to detect and locate road surface defects and output the detection results.

[0048] In this embodiment, the method of the present invention mainly includes a multi-scale feature extraction backbone network based on Res2net, a multi-level feature fusion neck network, and a detection head part, the network structure of which is as follows: Figure 1 As shown.

[0049] In a specific implementation, as a preferred embodiment of the present invention, step S1 includes: S11, through three consecutive 3 3. Convolution is used to extract features from the input road surface distress image and downsample the image resolution to 1 / 4 of the original resolution to output the initial feature map; S12. Configure the channel width of each Res2Net module and calculate the channel width of each Res2Net module. The formula is as follows:

[0050] in, Indicates the number of input channels. and The base width and number of groups for each group are set to 4 and 8 respectively; S13. The initial feature map is sequentially input into multiple Res2Net modules with configured channel widths for hierarchical feature extraction, outputting multi-scale feature maps of four semantic levels: P2, P3, P4, and P5. Each Res2Net module uniformly maps the input feature map along the channel dimension. Each feature mapping subset is used with... express, Each Having the same spatial size and the same number of channels as the original number of channels The feature subsets are processed using a hierarchical residual connection method. 3. Convolution processing ensures that each feature subset receives feature information from the preceding feature subset, and the output after residual connection... satisfy:

[0051] in, Indicates 3 3 convolution operations, each feature subset is processed The output results have a larger receptive field. The overall module output contains different numbers and combinations of receptive field sizes, and the output contains multi-scale feature maps with different receptive field combinations. S14. The output P5 level highest-level features are aggregated through multi-scale pooling using the spatial pyramid pooling module to output the aggregated high-level semantic feature map.

[0052] In specific implementation, as a preferred embodiment of the present invention, step S13 includes the following five stages: Stage 0: The initial feature map is processed by 3×3 depthwise convolution and 3×3 max pooling, and the output is a Stem-level feature map with a resolution of H / 4×W / 4 and 64 channels. Stage 1: By stacking 3 Bottle2neck modules [1×1,128; 3×3,32×4; 1×1,256], feature extraction is performed on the feature map output from Stage 0, and the output is a P2 level feature map with a resolution of H / 4×W / 4 and 256 channels. Stage 2: Features are extracted from the feature map output from Stage 1 by stacking 4 Bottle2neck modules [1×1,256; 3×3,64×4; 1×1,512], and the output is a P3 level feature map with a resolution of H / 8×W / 8 and 512 channels. Stage 3: Features are extracted from the feature map output from Stage 2 by stacking 6 Bottle2neck modules [1×1,512; 3×3,128×4; 1×1,1024], and the output is a P4 level feature map with a resolution of H / 16×W / 16 and 1024 channels. Stage 4: By stacking 3 Bottle2neck modules [1×1,1024; 3×3,256×4; 1×1,2048], feature extraction is performed on the feature map output from Stage 3, and the output is a P5 level feature map with a resolution of H / 32×W / 32 and 2048 channels.

[0053] In this embodiment, the multi-scale feature extraction backbone network based on Res2net is constructed based on the Res2net50 module, and through 3 3. Convolutional networks are used to construct residual connection networks, which can extract detailed features at different scales and enhance the model's ability to extract features of diseases at different scales. The overall backbone network design is shown in Table 1.

[0054]

[0055] During the initialization phase, the multi-scale feature extraction backbone network uses three consecutive 3×3 convolutions to reduce information loss and downsample the original image resolution to 1 / 4. In stages 1-4, the primary function of the first block is dimensionality transformation, employing average pooling with a stride of 2 combined with 1×1 convolutions to ensure that the input features can be element-wise added to the transformed output features. After the dimensionality transformation is completed, subsequent blocks in the same stage maintain a stride of 1, and the feature map resolution and number of channels remain unchanged.

[0056] In a specific implementation, as a preferred embodiment of the present invention, in step S2, the multi-level feature fusion neck network includes a top-down feature fusion path and a bottom-up feature fusion path, wherein: The top-down feature fusion path includes: S2a1. The enhanced feature alignment module transforms the dimensions of the output P5 level features and inputs them as high-level features into the cross-level fusion module. S2a2, The P5 level features after dimensionality transformation are input together with the P3 and P4 level features output by the backbone network into the cross-level fusion module for multi-scale feature fusion, and the fused P4 level features are output; wherein, the cross-level fusion module first performs a unified feature alignment operation on the features of the three different levels respectively:

[0057] in, Indicates the first The output features of each hierarchical feature after feature alignment operation; Indicates containing 1 1. Feature alignment operation of projection convolution and resampling operators to ensure that all output features They have the same dimensions and size; Indicates the input number of the first... Each level of feature; This represents a hierarchical index, with values... (lower floors) (Middle level) (High-level); S2a3: Concatenate the aligned features and reassemble them using a 1×1 convolution:

[0058] in, This represents the final output feature of the cross-level fusion module; This represents a 1×1 convolution operation; Indicates feature concatenation operation; This represents the aligned low-level features; This represents the aligned mid-level features; This represents the aligned high-level features; in this embodiment, this feature fusion can better preserve features from feature maps of different resolutions and construct a richer feature representation space.

[0059] S2a4. After adjusting the channels of the fused P4-level features through a 1×1 convolution, the features are enhanced by the reparameterized feature enhancement module, and the enhanced P4-level feature map is output. S2a5: The enhanced P4 level features are transformed in dimension by the enhanced feature alignment module and input into the next cross-level fusion module, where they are fused with the P1 / P2 and P3 level features output by the backbone network, and the fused P3 level features are output. S2a6. After the fused P3-level features are processed by 1×1 convolution, the P3-level feature map is output through the reparameterized feature enhancement module and used as the input of the small target detection head.

[0060] The bottom-up feature fusion path includes: S2b1: The output P3 level features are downsampled by a convolution with a stride of 2 through the enhanced feature alignment module, reducing the scale to P4 size. S2b2: The downsampled P3-level features are concatenated and fused with the output P4-level features, and then processed by the reparameterized feature enhancement module to output the fused and enhanced P4-level feature map as the input of the mid-target detection head; S2b3: The output P4 level features are downsampled by convolution with a stride of 2 through the enhanced feature alignment module, reducing the scale to P5 size; S2b4. The downsampled P4-level features are concatenated and fused with the output P5-level features, and then processed by the reparameterized feature enhancement module to output the fused and enhanced P5-level feature map as the input of the large target detection head.

[0061] In this embodiment, to eliminate the problem of disease target information loss caused by multiple downsampling of deep features, a multi-level feature fusion network composed of a cross-level feature fusion module and a reparameterized feature fusion module is designed as the neck network. Cross-scale feature fusion is performed through bidirectional feature fusion, and a structural reparameterized feature enhancement module is used to enhance the aggregation capability of multi-scale information. Furthermore, an enhanced feature alignment module is used to perform scale transformation to preserve the correlation between channels during downsampling, thereby improving the model's detection accuracy for multi-scale targets. Through this multi-scale feature fusion mechanism, the detection capability for diseases with large scale spans can be effectively improved by aggregating high-level semantic information and shallow detail information, as well as filtering surface similarity noise.

[0062] In the top-down feature fusion stage, the features output by the feature extraction network (SPPF-processed P5 level) are first transformed in dimensionality by an enhanced feature alignment module and then input as high-level features into the cross-level fusion module for multi-scale fusion with the P3 and P4 level features of the backbone network. The fused features are then adjusted by a 1×1 convolution, followed by feature enhancement through a reparameterized feature enhancement module, ultimately outputting a P4 level feature image. Next, the enhanced feature alignment module transforms the dimensions of the P4 level features output from the previous step again and inputs them into the next cross-level fusion module for fusion with the P1 / P2 and P3 level features of the backbone network. After fusion, they are again processed by a 1×1 convolution and then output as a P3 level feature image through a reparameterized feature enhancement module. This image is directly used as the input to the small target detection head.

[0063] In the bottom-up feature fusion stage, the output P3-level features are first downsampled (scaled down to P4 size) by performing a 2-stride convolution on them using an enhanced feature alignment module. The downsampled features are then directly fed into the stitching module and stitched together with the P4-level features generated in the previous top-down stage. The stitched features are then processed by a reparameterized feature enhancement module, outputting a fused and enhanced P4-level feature image, which serves as the input to the mid-target detection head. Finally, following this pattern, the P4 features from the previous step are downsampled again (scaled down to P5 size) by the enhanced feature alignment module. This downsampled feature is then fed into the stitching module and stitched together with the initially processed P5 high-level features. The fused signal is then finally extracted by the reparameterized feature enhancement module, outputting a fused and enhanced P5-level feature image, which serves as the input to the large target detection head.

[0064] Traditional feature pyramids typically only fuse features at adjacent scales, which can easily lead to semantic attenuation of shallow features during transmission. This paper introduces a cross-level bidirectional fusion module to improve the top-down feature fusion branch, simultaneously accepting features from three different levels (high, medium, and low). This effectively improves the semantic dilution problem caused by downsampling in deep networks during top-down feature extraction and enhances the ability to discriminate background noise. The structure diagram of the multi-level feature fusion neck network is shown below. Figure 3 .

[0065] In a specific implementation, as a preferred embodiment of the present invention, the feature enhancement performed by the reparameterized feature enhancement module in the top-down feature fusion path and the bottom-up feature fusion path specifically includes: During the training phase, a residual topology structure consisting of three parallel branches—a 3×3 main convolution branch, a 1×1 auxiliary convolution branch, and an identity branch—is used to extract multi-band features from the input features. During the inference deployment phase, leveraging the linear properties of convolutional layers, the parameters of the batch normalization layers in each branch are fused into the convolutional kernels of each branch. The batch normalization output is represented as:

[0066] in, This indicates the batch normalization operator; Indicates the input tensor; Indicates the mean of channel characteristics; Indicates the standard deviation of channel characteristics; Indicates the scaling parameter; Indicates the offset parameter; The equivalent convolutional kernel weights and biases after fusion are as follows:

[0067]

[0068] in, This represents the equivalent convolution kernel weight tensor after incorporating batch normalized parameters; Represents the original convolution kernel weight tensor; Indicates the first Scaling parameters for each output channel; Indicates the first The characteristic standard deviation of each output channel; This represents the local minimum constant, used to prevent the denominator from being zero; This represents the equivalent bias value after fusion; Indicates the first The characteristic mean of each output channel; Indicates the first Offset parameters for each output channel; The multi-branch structure is transformed into an equivalent single-path linear operator structure, where the equivalent convolution kernel weights and biases are the sum of the parameters after the branches are fused, as shown in the following formula:

[0069]

[0070] in, This represents the final equivalent convolution kernel weight tensor; This represents the weights of the 3×3 main branches after batch normalization. This represents the weight after merging the 1×1 auxiliary branches; This represents the weight after merging the identity branches; This represents the final equivalent bias vector; This represents the bias after merging the 3×3 main branches; This represents the bias after merging 1×1 auxiliary branches; This represents the bias after merging identity branches; Utilize the obtained equivalent single-path linear operator structure for input features Perform inference calculations and output the enhanced features:

[0071] in, This indicates a convolution operation.

[0072] In this embodiment, the core of the reparameterized feature enhancement module is to use a multi-branch parallel residual topology during training to expand the network's feature space and decouple the network topology between the training and inference phases. The overall network structure is as follows: Figure 4 As shown.

[0073] In specific implementation, as a preferred embodiment of the present invention, such as Figure 5 As shown, the transformation of dimensions through the enhanced feature alignment module in the top-down feature fusion path and the bottom-up feature fusion path specifically includes: Local semantic information is extracted using a 1×1 standard convolution while preserving original spatial detail features, resulting in an output channel count of [number missing]. Basic feature map :

[0074] in, This represents a 1×1 standard convolution; Indicates the input tensor; Nonlinear evolution of feature representation is achieved through a cascaded sequence of convolutional operators. First, 3×3 convolutions are used to map high-dimensional information between channels and establish cross-channel semantic dependencies. Then, depthwise separable convolutions decouple the correlation between spatial and channel dimensions. Finally, the signal is processed by the GELU activation function, resulting in an output with [number of channels]. Texture feature map :

[0075] in, Represents the activation function of the Gaussian error linear unit; The weights represent the depthwise separable convolutions; Represents the weights of a 3×3 dense convolution; Indicates the convolution operation; Basic feature map and Channel splicing is performed to obtain the fused spliced ​​feature map. :

[0076] in, This indicates a channel-level concatenation operation; splicing feature map The reshaping and shuffle operations are performed to uniformly resample the feature space, outputting the final enhanced feature-aligned feature map. :

[0077] in, This indicates a Shuffle operation, which is an operation that swaps positions within a group; This indicates a reshaping operation, which converts the spliced ​​feature map into two sets of feature structures consisting of the outputs of two branches.

[0078] In this embodiment, the enhanced feature alignment module primarily aims to transform the dimensionality of the high-level features output from the feature extraction network into the P5 dimension and enhance the texture features of the feature map. It optimizes the feature extraction efficiency and representation quality of the convolutional neural network by fusing features from regular convolution and depthwise convolution, thereby improving the feature texture extraction capability. The structure diagram is shown below. Figure 5 As shown, the enhanced feature alignment module mainly consists of two branches. The first is the basic feature extraction branch, which extracts local semantic information and preserves the original spatial detail features through a standard 1×1 convolution operator. The second is the texture extraction branch, which achieves nonlinear evolution of feature representation through a cascaded sequence of convolution operators. First, a 3×3 convolution operator is used to map high-dimensional information between channels to establish cross-channel semantic dependencies. Then, depthwise separable convolution decouples the correlation between spatial and channel dimensions, and effectively expands the receptive field through operator stacking without significantly increasing parameter redundancy, thereby enhancing the representation density of local texture patterns in the feature space.

[0079] In a specific implementation, as a preferred embodiment of the present invention, step S3 includes: S31. For each ground truth bounding box, select the frontmost bounding box whose center point is closest to the ground truth bounding box in each feature layer. One candidate anchor box; S32. Classification score based on network prediction The intersection-union ratio of candidate anchor boxes and true labeled boxes. Weighted processing is performed, and the crossover ratio (CRO) of the merged data is calculated. The mean of the CRO is then calculated. satisfy:

[0080] in, Indicates the number of candidate anchor boxes; S33. Calculate the standard deviation of the cross-union ratio (CUNR), and use the sum of the mean and standard deviation as the dynamic CUNR threshold. :

[0081] in, Indicates the first The crossover ratio of the candidate anchor boxes; S34. Select anchor boxes with a fusion intersection-union ratio higher than the dynamic threshold and whose center point falls within the ground truth bounding box as positive samples; if an anchor box corresponds to multiple ground truth bounding boxes, select the ground truth bounding box with the largest fusion intersection-union ratio for matching; S35. Based on the filtered positive samples, the detection head is used to classify the road surface defects and perform bounding box regression to obtain the final predicted box score. Output the final detection results, including the score of the final predicted bounding box. The formula is as follows:

[0082] in, One-hot encoding representing the actual label; This represents the intersection-union ratio (IoU) between the predicted bounding box and the actual labeled bounding box.

[0083] In this embodiment, the dynamic label allocation strategy directly combines the IoU between the predicted bounding box and the ground truth box, as well as the IoU between the anchor box and the ground truth box, to obtain an adaptive threshold based on the IoU, thereby improving the detection efficiency of positive samples. The positive sample threshold is no longer fixed after initialization but is dynamically adjusted during training, gradually increasing the positive sample standard as training progresses, prompting the model to learn higher-quality matches. The structure diagram of the dynamic label allocation strategy is shown below. Figure 6 As shown, the network receives a preset prior anchor box and the prediction confidence score of the current output of the network. ) and prediction box ( ), and real labels and bounding boxes ( , For each ground truth bounding box, first find the nearest neighbor to the center point in each feature layer. A pool of candidate anchor boxes (not predicted bounding boxes) is used, and combined with the network's prediction classification score, the IoU of the candidate anchors is weighted. The mean and standard deviation of the fused IoU are calculated to obtain a dynamic IoU threshold. Finally, anchor boxes with a fused IoU higher than the threshold and whose center points fall within the ground truth bounding boxes are selected as positive samples. If an anchor box corresponds to multiple ground truth bounding boxes, the ground truth bounding box with the largest fused IoU is selected to obtain the final predicted bounding box.

[0084] This invention also provides a pavement distress detection system based on the aforementioned cross-level multi-scale feature fusion method, comprising: a multi-scale feature extraction module, a multi-level feature fusion module, and a dynamic label allocation and detection module, wherein: The multi-scale feature extraction module is used to extract hierarchical features from the input road surface distress image through a multi-scale feature extraction backbone network based on the Res2Net architecture, and output a multi-scale feature map containing different semantic levels. The multi-level feature fusion module is used to perform cross-level bidirectional feature fusion on the output multi-scale feature map through the multi-level feature fusion neck network, and to aggregate deep semantic features and shallow detail features at multiple scales to output a fused and enhanced multi-level feature map. The dynamic label allocation and detection module is used to send the output multi-level feature map into the detection head, and adaptively adjust the positive and negative sample division threshold in combination with the dynamic label allocation strategy to detect and locate road surface defects and output the detection results.

[0085] The embodiments of the present invention are described simply because they correspond to those in the embodiments above. For any similarities, please refer to the descriptions in the embodiments above, which will not be elaborated here.

[0086] Example To verify the effectiveness of the method of this invention, two types of algorithms were introduced for comparative experiments: one type is a general target detection algorithm, and the other type is a road surface defect detection algorithm, such as RDD-YOLO, RTMDet-s, and MS-RDDNet. Different target detection algorithms were compared on the RDD2022, RDD2022-China subset, and UAVPDD2023 datasets. Details of each dataset are as follows: RDD2022 (Road Damage Detection 2022) is a large-scale, multi-national pavement damage dataset. As the official open-source dataset for the CRDDC 2022 challenge, it contains 47,420 real road images from six countries: Japan, India, the Czech Republic, Norway, the United States, and China. It provides detailed bounding box annotations for various pavement damage types, including longitudinal cracks, transverse cracks, network cracks (crazing), and potholes, offering an extremely rich sample of multi-national road environments. Its aim is to promote research into automated road infrastructure detection.

[0087] The RDD2022-China subset is a subset of RDD2022 specifically collected from the China region. This subset encompasses road images captured at different altitudes, including by drones (aerial view) and two-wheeled vehicles (ground view), realistically reflecting the unique road paving characteristics and complex environmental backgrounds of China. This subset is of significant reference value for evaluating the robustness of models in detecting changes in perspective and multi-scale defects (especially small-target defects) under specific regional disturbances.

[0088] UAVPDD2023 (Unmanned Aerial Vehicle Pavement Defect Detection 2023) is a high-resolution road surface defect dataset specifically designed for drone perspectives. Unlike traditional vehicle-mounted head-up data, the images in this dataset are primarily captured by drones from an aerial viewpoint. Since defects seen from a drone's perspective often exhibit large scales, extreme aspect ratios, and are highly susceptible to interference from complex road surface backgrounds (such as changes in lighting and shadow occlusion), introducing this dataset effectively validates the algorithm's ability to capture and accurately locate fine-grained, minute defect features in edge device aerial photography scenarios.

[0089] The experimental results are shown in the table below: Table 2 Performance metrics of different algorithms on the RDD2022 dataset

[0090] Table 3 Performance metrics of different algorithms on the RDD2022-China subset dataset

[0091] Table 4 Performance metrics of different algorithms on the UAVPDD2023 dataset

[0092] As shown in Tables 2 and 3, our algorithm achieved state-of-the-art results on both datasets. Due to the complex background of road surface defect detection and the irregular geometric shapes and high integration of defect edges with the background (normal road surface), the ground truth bounding boxes inherently possess semantic ambiguity. Our main focus is on improving mAP (maximum accuracy per second). On the RDD2022 dataset, the mAP reached 38.9%, 2.5 percentage points higher than the SEA-YOLO model. On the UAVDD2023 dataset, the mAP reached 63.5%, achieving the best result to date. On the RDD2022-China subset dataset, the mAP reached 58.6%, 1.5 percentage points higher than the best-performing RCD-DETR model on this dataset. Experimental results demonstrate that the proposed cross-level multi-scale feature fusion algorithm effectively enhances the model's feature extraction capability for complex road surface scenes. Through semantic alignment of multi-scale defect features, the accuracy of irregular defect bounding boxes is significantly improved, proving its high-quality boundary regression capability in the face of complex background interference.

[0093] To verify the effectiveness of each part of the algorithm, it was gradually modified based on the baseline model, and ablation experiments were conducted on the RDD2022-China dataset and the UAVPDD2023 dataset. The detection performance under different combinations was compared experimentally to verify the impact of each module on the algorithm's performance. The ablation experiment results on the RDD2022-China dataset are shown in Table 5, and the results on other datasets are similar.

[0094] Table 5 shows the ablation experiments on the RDD2022-China dataset.

[0095] As shown in Table 5, the experimental results demonstrate that, compared to the original baseline model, the detection performance of the model steadily improves after modifying each part of the network sequentially, verifying the effectiveness of each module in the algorithm presented in this chapter. The dynamic label assignment strategy enhances the model's ability to recognize positive samples; the multi-scale feature extraction network processes features at different scales, enhancing the model's ability to recognize targets of different scales and sizes; the asymmetric feature extraction network fuses feature information from different levels, improving the detection capability of targets with blurred textures; and the dynamic label assignment strategy increases the focus on positive samples. In summary, the three components of this network strengthen the network's feature extraction capability in road surface defect detection and improve the accuracy of target detection.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting pavement defects through cross-level, multi-scale feature fusion, characterized in that, include: S1. The input road surface distress image is processed by a multi-scale feature extraction backbone network based on the Res2Net architecture to extract hierarchical features and output a multi-scale feature map containing different semantic levels. S2. The multi-level feature fusion neck network is used to perform cross-level bidirectional feature fusion on the output multi-scale feature map, and the deep semantic features and shallow detail features are aggregated at multiple scales to output the fused and enhanced multi-level feature map. S3. The output multi-level feature map is sent to the detection head, and the positive and negative sample division threshold is adaptively adjusted in combination with the dynamic label allocation strategy to detect and locate road surface defects and output the detection results.

2. The method for detecting pavement defects by cross-level, multi-scale feature fusion according to claim 1, characterized in that, Step S1 includes: S11, through three consecutive 3 3. Convolution is used to extract features from the input road surface distress image and downsample the image resolution to 1 / 4 of the original resolution to output the initial feature map; S12. Configure the channel width of each Res2Net module and calculate the channel width of each Res2Net module. The formula is as follows: in, Indicates the number of input channels. and The base width and number of groups for each group are set to 4 and 8 respectively; S13. The initial feature map is sequentially input into multiple Res2Net modules with configured channel widths for hierarchical feature extraction, outputting multi-scale feature maps of four semantic levels: P2, P3, P4, and P5. Each Res2Net module uniformly maps the input feature map along the channel dimension. Each feature mapping subset is used with... express, Each Having the same spatial size and the same number of channels as the original number of channels The feature subsets are processed using a hierarchical residual connection method.

3. Convolution processing ensures that each feature subset receives feature information from the preceding feature subset, and the output after residual connection... satisfy: in, Indicates 3 3 convolution operations, each feature subset is processed The output results have a larger receptive field. The overall module output contains different numbers and combinations of receptive field sizes, and the output contains multi-scale feature maps with different receptive field combinations. S14. The output P5 level highest-level features are aggregated through multi-scale pooling using the spatial pyramid pooling module to output the aggregated high-level semantic feature map.

3. The method for detecting pavement defects by cross-level, multi-scale feature fusion according to claim 2, characterized in that, Step S13 includes the following five stages: Stage 0: The initial feature map is processed by 3×3 depthwise convolution and 3×3 max pooling, and the output is a Stem-level feature map with a resolution of H / 4×W / 4 and 64 channels. Stage 1: By stacking 3 Bottle2neck modules [1×1,128; 3×3,32×4; 1×1,256], feature extraction is performed on the feature map output from Stage 0, and the output is a P2 level feature map with a resolution of H / 4×W / 4 and 256 channels. Stage 2: Features are extracted from the feature map output from Stage 1 by stacking 4 Bottle2neck modules [1×1,256; 3×3,64×4; 1×1,512], and the output is a P3 level feature map with a resolution of H / 8×W / 8 and 512 channels. Stage 3: Features are extracted from the feature map output from Stage 2 by stacking 6 Bottle2neck modules [1×1,512; 3×3,128×4; 1×1,1024], and the output is a P4 level feature map with a resolution of H / 16×W / 16 and 1024 channels. Stage 4: By stacking 3 Bottle2neck modules [1×1,1024; 3×3,256×4; 1×1,2048], feature extraction is performed on the feature map output from Stage 3, and the output is a P5 level feature map with a resolution of H / 32×W / 32 and 2048 channels.

4. The method for detecting pavement defects by cross-level, multi-scale feature fusion according to claim 1, characterized in that, In step S2, the multi-level feature fusion neck network includes a top-down feature fusion path and a bottom-up feature fusion path, wherein: The top-down feature fusion path includes: S2a1. The enhanced feature alignment module transforms the dimensions of the output P5 level features and inputs them as high-level features into the cross-level fusion module. S2a2, The P5 level features after dimensionality transformation are input together with the P3 and P4 level features output by the backbone network into the cross-level fusion module for multi-scale feature fusion, and the fused P4 level features are output; wherein, the cross-level fusion module first performs a unified feature alignment operation on the features of the three different levels respectively: in, Indicates the first The output features of each hierarchical feature after feature alignment operation; Indicates containing 1 1. Feature alignment operation of projection convolution and resampling operators to ensure that all output features They have the same dimensions and size; Indicates the input number of the first... Each level of feature; This represents a hierarchical index, with values... , , ; S2a3: Concatenate the aligned features and reassemble them using a 1×1 convolution: in, This represents the final output feature of the cross-level fusion module; This represents a 1×1 convolution operation; Indicates feature concatenation operation; This represents the aligned low-level features; This represents the aligned mid-level features; This represents the aligned high-level features; S2a4. After adjusting the channels of the fused P4-level features through a 1×1 convolution, the features are enhanced by the reparameterized feature enhancement module, and the enhanced P4-level feature map is output. S2a5: The enhanced P4 level features are transformed in dimension by the enhanced feature alignment module and input into the next cross-level fusion module, where they are fused with the P1 / P2 and P3 level features output by the backbone network, and the fused P3 level features are output. S2a6. After the fused P3-level features are processed by 1×1 convolution, the P3-level feature map is output through the reparameterized feature enhancement module and used as the input of the small target detection head. The bottom-up feature fusion path includes: S2b1: The output P3 level features are downsampled by a convolution with a stride of 2 through the enhanced feature alignment module, reducing the scale to P4 size. S2b2: The downsampled P3-level features are concatenated and fused with the output P4-level features, and then processed by the reparameterized feature enhancement module to output the fused and enhanced P4-level feature map as the input of the mid-target detection head; S2b3: The output P4 level features are downsampled by convolution with a stride of 2 through the enhanced feature alignment module, reducing the scale to P5 size; S2b4. The downsampled P4-level features are concatenated and fused with the output P5-level features, and then processed by the reparameterized feature enhancement module to output the fused and enhanced P5-level feature map as the input of the large target detection head.

5. The method for detecting pavement defects by cross-level, multi-scale feature fusion according to claim 4, characterized in that, In the top-down and bottom-up feature fusion paths, the reparameterized feature enhancement module performs feature enhancement specifically including: During the training phase, a residual topology structure consisting of three parallel branches—a 3×3 main convolution branch, a 1×1 auxiliary convolution branch, and an identity branch—is used to extract multi-band features from the input features. During the inference deployment phase, leveraging the linear properties of convolutional layers, the parameters of the batch normalization layers in each branch are fused into the convolutional kernels of each branch. The batch normalization output is represented as: in, This indicates the batch normalization operator; Indicates the input tensor; Indicates the mean of channel characteristics; Indicates the standard deviation of channel characteristics; Indicates the scaling parameter; Indicates the offset parameter; The equivalent convolutional kernel weights and biases after fusion are as follows: in, This represents the equivalent convolution kernel weight tensor after incorporating batch normalized parameters; Represents the original convolution kernel weight tensor; Indicates the first Scaling parameters for each output channel; Indicates the first The characteristic standard deviation of each output channel; This represents the local minimum constant, used to prevent the denominator from being zero; This represents the equivalent bias value after fusion; Indicates the first The characteristic mean of each output channel; Indicates the first Offset parameters for each output channel; The multi-branch structure is transformed into an equivalent single-path linear operator structure, where the equivalent convolution kernel weights and biases are the sum of the parameters after the branches are fused, as shown in the following formula: in, This represents the final equivalent convolution kernel weight tensor; This represents the weights of the 3×3 main branches after batch normalization. This represents the weight after merging the 1×1 auxiliary branches; This represents the weight after merging the identity branches; This represents the final equivalent bias vector; This represents the bias after merging the 3×3 main branches; This represents the bias after merging 1×1 auxiliary branches; This represents the bias after merging identity branches; Utilize the obtained equivalent single-path linear operator structure for input features Perform inference calculations and output the enhanced features: in, This indicates a convolution operation.

6. The method for detecting pavement defects by cross-level, multi-scale feature fusion according to claim 4, characterized in that, In the top-down and bottom-up feature fusion paths, the dimensionality transformation through the enhanced feature alignment module specifically includes: Local semantic information is extracted using a 1×1 standard convolution while preserving original spatial detail features, resulting in an output channel count of [number missing]. Basic feature map : in, This represents a 1×1 standard convolution; Indicates the input tensor; Nonlinear evolution of feature representation is achieved through a cascaded sequence of convolutional operators. First, 3×3 convolutions are used to map high-dimensional information between channels and establish cross-channel semantic dependencies. Then, depthwise separable convolutions decouple the correlation between spatial and channel dimensions. Finally, the signal is processed by the GELU activation function, resulting in an output with [number of channels]. Texture feature map : in, Represents the activation function of the Gaussian error linear unit; The weights represent the depthwise separable convolutions; Represents the weights of a 3×3 dense convolution; Indicates the convolution operation; Basic feature map and Channel splicing is performed to obtain the fused spliced ​​feature map. : in, This indicates a channel-level concatenation operation; splicing feature map The reshaping and shuffle operations are performed to uniformly resample the feature space, outputting the final enhanced feature-aligned feature map. : in, This indicates a Shuffle operation, which is an operation that swaps positions within a group; This indicates a reshaping operation, which converts the spliced ​​feature map into two sets of feature structures consisting of the outputs of two branches.

7. The method for detecting pavement defects by cross-level, multi-scale feature fusion according to claim 1, characterized in that, Step S3 includes: S31. For each ground truth bounding box, select the frontmost bounding box whose center point is closest to the ground truth bounding box in each feature layer. One candidate anchor box; S32. Classification score based on network prediction The intersection-union ratio of candidate anchor boxes and true labeled boxes. Weighted processing is performed, and the crossover ratio (CRO) of the merged data is calculated. The mean of the CRO is then calculated. satisfy: in, Indicates the number of candidate anchor boxes; S33. Calculate the standard deviation of the cross-union ratio (CUNR), and use the sum of the mean and standard deviation as the dynamic CUNR threshold. : in, Indicates the first The crossover ratio of the candidate anchor boxes; S34. Select anchor boxes with a fusion intersection-union ratio higher than the dynamic threshold and whose center point falls within the ground truth bounding box as positive samples; if an anchor box corresponds to multiple ground truth bounding boxes, select the ground truth bounding box with the largest fusion intersection-union ratio for matching; S35. Based on the filtered positive samples, the detection head is used to classify the road surface defects and perform bounding box regression to obtain the final predicted box score. Output the final detection results, including the score of the final predicted bounding box. The formula is as follows: in, One-hot encoding representing the actual label; This represents the intersection-union ratio (IoU) between the predicted bounding box and the actual labeled bounding box.

8. A pavement distress detection system based on the pavement distress detection method of cross-level multi-scale feature fusion as described in any one of claims 1-7, characterized in that, include: The module comprises a multi-scale feature extraction module, a multi-level feature fusion module, and a dynamic label allocation and detection module, among which: The multi-scale feature extraction module is used to extract hierarchical features from the input road surface distress image through a multi-scale feature extraction backbone network based on the Res2Net architecture, and output a multi-scale feature map containing different semantic levels. The multi-level feature fusion module is used to perform cross-level bidirectional feature fusion on the output multi-scale feature map through the multi-level feature fusion neck network, and to aggregate deep semantic features and shallow detail features at multiple scales to output a fused and enhanced multi-level feature map. The dynamic label allocation and detection module is used to send the output multi-level feature map into the detection head, and adaptively adjust the positive and negative sample division threshold in combination with the dynamic label allocation strategy to detect and locate road surface defects and output the detection results.

Images