A road damage detection method based on direction perception and feature equalization

Abstract

The application discloses a road damage detection method based on direction perception and feature equalization, characterized by constructing and training a road damage detection model comprising a backbone network, a neck network and a detection head, the backbone network extracts input images to obtain first, second and third feature maps; the neck network comprises a direction perception feature enhancement module and a multi-level feature equalization module, the former processes the first feature map to obtain an enhanced feature map, and the latter processes the enhanced feature map, the second feature map and the third feature map to obtain an equalized feature map, and the detection head predicts the equalized feature map to obtain a road damage detection result; the advantage is that different levels of features are spatially scaled and aligned and unified in channel dimension, a learnable spatial weight parameter is introduced to adjust the proportion of different levels of features, the feature suppression problem caused by scale difference in multi-scale feature fusion is alleviated, and thus the detection accuracy of all damage categories is balanced and improved.

Description

Technical Field

[0001] This invention relates to road damage detection methods, and more particularly to a road damage detection method based on direction perception and feature equalization. Background Technology

[0002] Currently, road freight transport accounts for 73.6% of the nation's total commercial freight volume, and road surface conditions directly affect driving safety and transportation efficiency. Under the influence of traffic loads, environmental erosion, and material aging, road surface damage is becoming increasingly serious, becoming a major cause of traffic accidents and resulting in significant economic losses annually. Therefore, developing efficient and accurate road surface damage detection algorithms is of great significance for optimizing maintenance decisions and ensuring traffic safety.

[0003] Currently, general-purpose object detection algorithms have achieved excellent performance in many fields, but they still cannot be directly applied to pavement damage detection. The root cause lies in the fact that their commonly used rectangular convolution lacks the ability to represent continuous and directional information, making it difficult to accurately capture the feature information of crack-like damage. Therefore, their detection accuracy is not high after direct transplantation. How to effectively extract special morphological information such as directionality from cracks will directly affect the final detection effect.

[0004] Furthermore, pavement damage exhibits complex morphologies and significant size differences between different types of damage. This necessitates models with high-precision detection capabilities for pavement damage of various shapes and scales. Simultaneously, existing models exhibit varying degrees of sensitivity to features across different damage categories, leading to distinct feature suppression issues. Therefore, feature balancing must be considered when improving accuracy. How to alleviate the feature suppression problem of existing models and thus achieve a balanced improvement in detection accuracy for all damage types is also one of the key research challenges in this field. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a road damage detection method based on direction perception and feature equalization, which can effectively extract directional texture features in road damage and equalize the expression of different types of damage features, thereby improving the detection accuracy of various types of road damage.

[0006] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows: a road damage detection method based on direction perception and feature equalization, comprising: firstly constructing a road damage detection model composed of a backbone network, a neck network and a detection head connected in sequence; then training the road damage detection model to obtain a trained road damage detection model; finally inputting the road image to be detected into the trained road damage detection model to obtain the road damage detection result; The specific operational process for obtaining road damage detection results is as follows: Step 1: The backbone network extracts features from the input road image to be detected, obtaining first feature maps, second feature maps, and third feature maps of different scales; Step 2: The neck network performs multi-scale feature fusion on the first, second, and third feature maps output by the backbone network to obtain a balanced feature map; Step 3: Perform bounding box coordinate regression, target confidence prediction, and damage category prediction on the balanced feature map output by the head-to-neck network to obtain road damage detection results containing the location box of the damaged target, damage category, and confidence level. The neck network includes a direction-aware feature enhancement module and a multi-level feature equalization module connected in sequence; The specific operation process of step 2 is as follows: Step 2-1: The orientation-aware feature enhancement module enhances the directional texture features in the first feature map to obtain an enhanced feature map; Step 2-2: The multi-level feature equalization module performs spatial scale alignment, channel dimension unification, and adaptive weighted fusion operations on the enhanced feature map, the second feature map, and the third feature map to obtain the equalized feature map.

[0007] Compared with existing technologies, the advantages of this invention are as follows: First, by setting a direction-aware feature enhancement module in the neck network, the directional texture features in the first feature map output by the backbone network are specifically enhanced, which significantly strengthens the continuity and directionality information of crack-like damage that is difficult to capture under rectangular convolution operations, thereby solving the problem of low crack detection accuracy caused by the lack of directional representation capabilities in existing technologies; Second, by using a multi-level feature equalization module to sequentially perform spatial scale alignment, channel dimension unification, and adaptive weighted fusion on the enhanced feature map and the second and third feature maps, not only can different scales be made more consistent, but also the directional texture features of crack-like damage that are difficult to capture under rectangular convolution operations are significantly enhanced. The proposed method achieves balanced damage features within a unified expression space and effectively suppresses feature suppression of different damage categories by employing an adaptive weighting mechanism. This avoids the problem of low detection accuracy for certain damage types due to differences in feature sensitivity. Experimental results on the public dataset RDD2022 show that the proposed method achieves an mAP@50 of 62.8%, a 5.8 percentage point improvement over the baseline model. The method also achieves a recall of 58.3% and an inference speed of 122 FPS, demonstrating that it maintains good real-time performance while improving damage detection accuracy. In summary, this invention significantly improves road damage detection performance through the effective combination of direction-aware feature enhancement and multi-level feature balancing, achieving high-precision identification, especially for slender crack targets, while also maintaining real-time processing capabilities. It can provide a reliable algorithmic reference for intelligent road damage inspection and maintenance decision-making.

[0008] Furthermore, in step 1, the backbone network includes a first CBS layer, a second CBS layer, a first C2f layer, a third CBS layer, a second C2f layer, a third C2f layer, a fourth CBS layer, and a fourth C2f layer connected in sequence. The specific operation process of step 1 is as follows: Step 1-1: The first CBS layer performs downsampling feature extraction on the input road image to be detected to obtain the first intermediate feature map; Steps 1-2: The second CBS layer performs downsampling feature extraction on the first intermediate feature map output by the first CBS layer to obtain the second intermediate feature map; Steps 1-3: The first C2f layer performs feature fusion processing on the second intermediate feature map output by the second CBS layer to obtain the third intermediate feature map; Steps 1-4: The third CBS layer performs downsampling feature extraction on the third intermediate feature map output by the first C2f layer to obtain the fourth intermediate feature map; Steps 1-5: The second C2f layer performs feature fusion processing on the fourth intermediate feature map output by the third CBS layer to obtain a first feature map with a size of 256×80×80. Steps 1-6: The third C2f layer performs feature fusion processing on the first feature map output by the second C2f layer to obtain a second feature map with a size of 512×40×40. Steps 1-7: The fourth CBS layer performs downsampling feature extraction on the second feature map output by the third C2f layer to obtain the fifth intermediate feature map; Steps 1-8: The fourth C2f layer performs feature fusion processing on the fifth intermediate feature map output by the fourth CBS layer to obtain a third feature map with a size of 512×20×20.

[0009] Furthermore, the CBS layer comprises a convolutional layer, a batch normalization layer, and a SiLU activation function layer connected in sequence; the C2f layer comprises a first convolutional layer, a feature splitting unit, at least one bottleneck unit, a feature concatenation unit, and a second convolutional layer.

[0010] Furthermore, the specific operation process of step 2-1 is as follows: Step 2-1-1: Perform channel compression on the first feature map using a 1×1 convolution to obtain a compressed feature map. , ,in, This represents a 1×1 convolution operation. Represents the first feature map. , Indicates the number of input channels. Indicates the height of the first feature map. This represents the width of the first feature map. , Number of output channels For expansion ratio; Step 2-1-2: Compress the feature map using depthwise separable strip convolution. Perform directional decomposition to obtain the lateral feature map. and vertical feature map , , ,in, This represents a depthwise separable stripe convolution operation. For the strip core size; Step 2-1-3, transform the horizontal feature map and vertical feature map After fusion, batch normalization, SiLU activation function, and 3×3 convolution operation are performed sequentially to obtain the integrated feature map. , ,in, This indicates element-wise addition. This indicates a batch normalization operation. This represents the SiLU activation function. This represents a 3×3 convolution operation; Step 2-1-4: When the number of input channels matches the number of output channels, combine the first feature map with the integrated feature map. The enhanced feature map is obtained by adding elements one by one; when the number of input channels is inconsistent with the number of output channels, the feature maps are integrated. As an enhancement of the feature map, the introduction of asymmetric strip convolution can enhance the model's ability to perceive lateral and longitudinal directional texture features, thereby improving the detection accuracy of road damage such as thin cracks; depthwise separable convolution reduces the computational cost in the directional feature extraction process, ensuring detection speed.

[0011] Furthermore, the specific operation process of step 2-2 is as follows: Step 2-2-1: Obtain the maximum spatial scale of the enhanced feature map, the second feature map, and the third feature map. Use nearest neighbor interpolation to upsample the enhanced feature map, the second feature map, and the third feature map to the maximum spatial scale to obtain the corresponding scale-aligned feature maps. Step 2-2-2: Use independent 1×1 convolutional layers to unify the number of channels in the scale-aligned feature maps to the same dimension, resulting in channel-unified feature maps. , ,in, This indicates the nearest neighbor interpolation operation. This represents a 1×1 convolution operation used to generate spatial weights. , Represents the enhanced feature map, This represents the second feature map. Represents the third feature map; Step 2-2-3: Perform concatenation along the channel dimension to obtain the concatenated feature map. , ,in, Indicates a splicing operation; Step 2-2-4, stitching feature maps Perform independent 1×1 convolution operations to generate the original weights. , ; Step 2-2-5: Adjust the original weights on the enhanced feature map, the second feature map, and the third feature map respectively. conduct The normalization operation yields the enhanced feature map, the second feature map, and the third feature map in spatial location, respectively. The corresponding spatial weights: , , Among them, satisfying , These represent the original weights corresponding to the enhanced feature map, the second feature map, and the third feature map, respectively. Step 2-2-6, will and Weighted fusion is performed to obtain the fused feature map. , ,in, For element-wise multiplication; Step 2-2-7, using a 1×1 convolutional layer... Channel integration is performed to obtain the equalization feature map. By aligning the spatial scale and unifying the channel dimensions of the input multi-level feature maps, and introducing a dynamic weighting mechanism to adjust the contribution of different level feature maps in the fusion process, the phenomenon of shallow detail information being suppressed by deep semantic features is alleviated, effectively improving the model's ability to detect all damage categories.

[0012] Furthermore, the specific steps for training the road damage detection model are as follows: Images corresponding to four types of targets—longitudinal cracks, transverse cracks, alligator cracks, and potholes—were selected from the public road damage dataset RDD2022, and 80% of the selected images were used as the training set. The training parameters are set as follows: a total of 200 training rounds, a batch size of 16, a weight decay of 0.05, and a momentum of 0.937. The total loss function is set as the sum of the bounding box loss and the classification loss. Based on the total loss function, the training set is input into the road damage detection model for training. The road damage detection model is updated by the AdamW optimizer until the total number of training rounds is reached, at which point the iterative update stops, and the trained road damage detection model is obtained. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the overall structure of the road damage detection model in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the categories of images selected in this embodiment of the invention; Figure 3 This is a thermodynamic analysis diagram of the strip convolution effect in an embodiment of the present invention; Figure 4 This is a comparison chart of the detection results using the method proposed in this invention and using existing technology in an embodiment of this invention. Detailed Implementation

[0014] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0015] To address the issues of insufficient detection accuracy for slender cracks and information suppression in multi-level feature fusion in road damage detection, a road damage detection method based on direction awareness and feature equalization is proposed. This method includes: firstly, constructing a road damage detection model consisting of a sequentially connected backbone network, a neck network, and a detection head, as shown below. Figure 1 As shown; then the road damage detection model is trained to obtain the trained road damage detection model; finally, the road image to be detected is input into the trained road damage detection model to obtain the road damage detection result; The specific steps for training the road damage detection model are as follows: Images corresponding to four types of targets—longitudinal cracks (D00), transverse cracks (D10), alligator cracks (D20), and potholes (D40)—were selected from the publicly available road damage dataset RDD2022. Longitudinal cracks mainly extend parallel to the road along the direction of wheel tracks; transverse cracks are distributed perpendicular to the road direction; alligator cracks present a network-like pattern covering the road surface; and potholes include depressions, ruts, and pavement peeling. Specific examples are shown below. Figure 2As shown, the selected images comprise 38,885 images from road scenes in six countries: China (drones and motorcycles), Czech Republic, India, Japan, Norway, and the United States. Among them, 24,267 images are labeled, and 14,618 are background images, for a total of 56,411 bounding boxes. The number of labeled images for each category is as follows: D00 (26,378), D10 (11,999), D20 (10,755), and D40 (7,279). Finally, 80% of the selected images were randomly selected as the training set (31,104 images), and 20% were selected as the test set (7,781 images). The training parameters are set as follows: a total of 200 training rounds, a batch size of 16, a weight decay of 0.05, and a momentum of 0.937. The total loss function is set as the sum of the bounding box loss and the classification loss. Based on the total loss function, the training set is input into the road damage detection model for training. The road damage detection model is updated by the AdamW optimizer until the total number of training rounds is reached, at which point the iterative update stops, and the trained road damage detection model is obtained.

[0016] The specific operational process for obtaining road damage detection results is as follows: Step 1: The backbone network extracts features from the input road image to be detected, obtaining first feature maps, second feature maps, and third feature maps of different scales; the backbone network includes the first CBS layer, the second CBS layer, the first C2f layer, the third CBS layer, the second C2f layer, the third C2f layer, the fourth CBS layer, and the fourth C2f layer connected in sequence. The specific operation process of step 1 is as follows: Step 1-1: The first CBS layer performs downsampling feature extraction on the input road image to be detected to obtain the first intermediate feature map; Steps 1-2: The second CBS layer performs downsampling feature extraction on the first intermediate feature map output by the first CBS layer to obtain the second intermediate feature map; Steps 1-3: The first C2f layer performs feature fusion processing on the second intermediate feature map output by the second CBS layer to obtain the third intermediate feature map; Steps 1-4: The third CBS layer performs downsampling feature extraction on the third intermediate feature map output by the first C2f layer to obtain the fourth intermediate feature map; Steps 1-5: The second C2f layer performs feature fusion processing on the fourth intermediate feature map output by the third CBS layer to obtain a first feature map with a size of 256×80×80. Steps 1-6: The third C2f layer performs feature fusion processing on the first feature map output by the second C2f layer to obtain a second feature map with a size of 512×40×40. Steps 1-7: The fourth CBS layer performs downsampling feature extraction on the second feature map output by the third C2f layer to obtain the fifth intermediate feature map; Steps 1-8: The fourth C2f layer performs feature fusion processing on the fifth intermediate feature map output by the fourth CBS layer to obtain a third feature map with a size of 512×20×20. The CBS layer includes a convolutional layer, a batch normalization layer, and a SiLU activation function layer connected in sequence; the C2f layer includes a first convolutional layer, a feature splitting unit, at least one bottleneck unit, a feature concatenation unit, and a second convolutional layer. In this embodiment, both the CBS layer and the C2f layer adopt existing technologies.

[0017] Step 2: The neck network performs multi-scale feature fusion on the first feature map, second feature map and third feature map output by the backbone network to obtain a balanced feature map; the neck network includes a direction-aware strip conv (DSC) module and a multi-level feature equalization (MFE) module connected in sequence. Road cracks are elongated lines, typically only a few pixels wide, but extending over a considerable distance, exhibiting significant continuous linear geometric features. Traditional convolutional kernels with their square receptive fields respond well to isotropic textures but struggle to effectively match the elongated shape of cracks. This leads to discontinuous crack responses and the introduction of irrelevant background noise during feature extraction. The orientation-aware feature enhancement module, by introducing asymmetric strip convolution, extracts features from cracks that are nearly horizontal and vertical.

[0018] Step 2-1: The orientation-aware feature enhancement module enhances the directional texture features in the first feature map to obtain an enhanced feature map. The specific operation process is as follows: Step 2-1-1: Channel compression is performed on the first feature map using 1×1 convolution to reduce computational complexity while maintaining feature expressiveness, resulting in a compressed feature map. , ,in, This represents a 1×1 convolution operation. Represents the first feature map. , Indicates the number of input channels. Indicates the height of the first feature map. This represents the width of the first feature map. , Number of output channels For expansion ratio; Step 2-1-2: Compress the feature map using depthwise separable strip convolution. Perform directional decomposition to obtain the lateral feature map. and vertical feature map , , ,in, This represents a depthwise separable stripe convolution operation. For the strip core size, For odd numbers greater than or equal to 3, use 1× which has a strong ability to perceive the elongated features of transverse and longitudinal cracks. and ×1 Two-directional strip convolution replaces traditional × Convolutional layers capture horizontal and vertical textures separately, and the feature responses in both directions are fused through element-wise addition. This design enables the network to respond to features in both the horizontal and vertical directions simultaneously, potentially enhancing the response to local linear structures for diagonal or mesh-like crack-like damage (D20 class). In this embodiment... Choose 7 or 11; Step 2-1-3, transform the horizontal feature map and vertical feature map After fusion, batch normalization, SiLU activation function, and 3×3 convolution operation are performed sequentially to integrate the data, capturing local details and obtaining the integrated feature map. , ,in, This indicates element-wise addition. This indicates a batch normalization operation. This represents the SiLU activation function. This represents a 3×3 convolution operation; Step 2-1-4: Output via residual connection; the residual connection is activated only when the number of input and output channels are the same, that is, when the number of input channels is the same as the number of output channels, the first feature map and the integrated feature map are connected. The enhanced feature map is obtained by adding elements one by one; when the number of input channels is inconsistent with the number of output channels, the feature maps are integrated. As an enhanced feature map, the expression is: ,in, This represents the enhanced feature map.

[0019] To verify the selective response principle of this module to directional features, Figure 3 A thermogram comparison of samples with longitudinal and transverse cracks is presented. For transverse crack samples, 1× Lateral strip convolution exhibits a strong activation response at the crack location, while ×1 longitudinal strip convolution exhibits a weak activation response; for longitudinal crack samples... ×1 longitudinal strip convolution exhibits a strong response at the crack, while 1× The transverse strip convolution response is weak. When the two are superimposed, both transverse and longitudinal cracks exhibit strong responses, demonstrating that the module can extract the geometric features of the cracks and perform feature enhancement.

[0020] Road damage detection reveals damage of various morphologies. Narrow, linear cracks are common; crazing presents as large, irregular cracks; potholes appear as irregular depressions or ruts. The scale and morphology of different types of damage vary significantly. Existing feature pyramid structures often directly splice or add features when fusing multi-level features, leading to feature information suppression between shallow details and deep semantics at the spatial and channel levels, thus preventing the full realization of feature fusion effects. The multi-level feature equalization module, through the design of spatial scale alignment and channel dimension unification operations, combined with a spatial weighting mechanism, dynamically adjusts the fusion contribution of features at different levels, aiming to alleviate the suppression of detailed feature information. Its core components include spatial scale alignment, channel dimension unification, and adaptive weighted fusion.

[0021] Step 2-2: The multi-level feature equalization module performs spatial scale alignment, channel dimension unification, and adaptive weighted fusion operations on the enhanced feature map, the second feature map, and the third feature map to obtain the equalized feature map. The specific operation process is as follows: Step 2-2-1: Obtain the maximum spatial scale from the enhanced feature map, the second feature map, and the third feature map. and , Nearest neighbor interpolation is used to upsample the enhanced feature map, the second feature map, and the third feature map to the maximum spatial scale, respectively, to obtain the corresponding scale-aligned feature maps. These represent the heights of the enhanced feature map, the second feature map, and the third feature map, respectively. These represent the widths of the enhanced feature map, the second feature map, and the third feature map, respectively. Nearest neighbor interpolation can reduce the edge smoothing effect that bilinear interpolation may cause, thereby preserving the edge details of small targets such as cracks as much as possible. Step 2-2-2: To eliminate the fusion bias caused by inconsistent channel numbers, independent 1×1 convolutional layers are used to unify the number of channels in the scale-aligned feature maps to the same dimension, resulting in channel-unified feature maps. , ,in, This indicates the nearest neighbor interpolation operation. This represents a 1×1 convolution operation used to generate spatial weights. , Represents the enhanced feature map, This represents the second feature map. Represents the third feature map; Step 2-2-3: Perform concatenation along the channel dimension to obtain the concatenated feature map. , ,in, This indicates a splicing operation. ; Step 2-2-4, stitching feature maps Perform independent 1×1 convolution operations to generate the original weights. , , , It contains 3 channels, which correspond to the response weights of the enhanced feature map, the second feature map, and the third feature map at different spatial locations; Step 2-2-5: Adjust the original weights on the enhanced feature map, the second feature map, and the third feature map respectively. conduct The normalization operation yields the enhanced feature map, the second feature map, and the third feature map in spatial location, respectively. The corresponding spatial weights: , , Among them, satisfying , These represent the original weights corresponding to the enhanced feature map, the second feature map, and the third feature map, respectively. Step 2-2-6, will and Weighted fusion is performed to obtain the fused feature map. , ,in, This is an element-wise multiplication method; this fusion method enables the model to adaptively adjust the contribution of features at different levels according to the feature requirements of different spatial locations; for detailed areas such as crack edges, shallow features can obtain higher weights; for semantic areas such as pits and cracks, deep features can play a greater role. Step 2-2-7, using a 1×1 convolutional layer... Channel integration is performed to obtain the equalization feature map. .

[0022] Step 3: Perform bounding box coordinate regression, target confidence prediction, and damage category prediction on the balanced feature map output by the head-to-neck network to obtain road damage detection results containing the location box of the damaged target, damage category, and confidence.

[0023] To verify the effectiveness of this invention, the following experiments were conducted: The experiments were performed on a workstation equipped with an Intel Xeon Gold 5218R processor (2.10GHz) and an NVIDIA GeForce RTX 3080 GPU. The software environment was Windows 10 operating system, and the deep learning framework was PyTorch 2.2.0 with CUDA 12.1. Model training used the AdamW optimizer, a cosine annealing learning rate scheduling strategy, a total of 200 epochs, a batch size of 16, a weight decay of 0.05, and a momentum of 0.937. Data augmentation strategies included Mosaic, Mixup, Copy-Paste, and label smoothing. The loss function weights were configured as bounding box loss of 8.0 and classification loss of 0.6, and mixed-precision training was enabled.

[0024] This experiment uses precision (P), recall (R), F1 score, and mean precision (mAP@50) to evaluate model performance. Precision measures prediction accuracy, recall measures detection completeness, and the F1 score assesses the balance between the two. mAP@50 at an IoU threshold of 0.5 is used as the core performance indicator. By comprehensively analyzing these indicators, the specific effectiveness of the model in road damage detection tasks can be evaluated in depth and objectively from different dimensions.

[0025] To verify the effectiveness of this invention in road damage detection, the proposed method was compared with several mainstream object detection algorithms, including Faster R-CNN, Cascade R-CNN, CenterNet, and the YOLO series algorithms YOLOv5s, YOLOv8n, YOLOv11n, and YOLOv8-DRM. All models were trained and tested under the same dataset partitioning and experimental environment, and were evaluated in terms of mAP@50, recall, and inference speed. The experimental results are shown in Table 1.

[0026] Table 1

[0027] As shown in Table 1, this invention achieves a balance between detection accuracy and real-time performance. Specifically, the mAP@50 of this invention reaches 62.8%, which is 5.8 percentage points higher than the benchmark model YOLOv8n (57.0%) and 2.1 percentage points higher than the similar algorithm YOLOv8 DRM (60.7%), verifying the effectiveness of the proposed DSC and MFE modules. The recall rate of this invention reaches 58.3%, higher than all compared methods, indicating that the road damage detection model has a stronger detection capability for road surface damage targets and a lower false negative rate. This invention achieves an inference speed of 122 FPS, which meets the real-time requirements of road detection tasks. In summary, the proposed method achieves a better balance between accuracy and speed, verifying the effectiveness of this invention in road surface damage detection tasks.

[0028] To analyze the respective contributions of the Direction Aware Feature Enhancement (DSC) module and the Multi-Level Feature Equalization (MFE) module to model performance, this invention designed an ablation experiment using YOLOv8n as the baseline model. While ensuring consistent training configuration, the DSC and MFE modules were integrated into the baseline model respectively. Detailed experimental results are shown in Table 2.

[0029] Table 2

[0030] As shown in Table 2, after introducing the DSC module alone, the model's mAP@50 increased from 57.0% to 61.7%, an improvement of 4.7 percentage points, and the recall increased from 52.4% to 57.4%, indicating that direction-aware strip convolution can enhance the model's ability to represent the directional texture of road damage. After introducing the MFE module alone, the model's mAP@50 reached 62.1%, an improvement of 5.1 percentage points compared to the baseline, indicating that spatial scale alignment, channel unification, and adaptive weighted fusion can effectively alleviate the information suppression problem in the multi-level feature fusion process.

[0031] When the DSC and MFE modules are introduced simultaneously, the model's P, R, and mAP@50 reach 69.8%, 58.3%, and 62.8%, respectively, all outperforming the baseline model and the single-module improved model. This result indicates that DSC and MFE improve detection performance from the perspectives of directional feature enhancement and multi-scale feature equalization, respectively, and that they have a certain complementary effect.

[0032] Table 3

[0033] As shown in Table 3, the present invention achieved performance improvements in all four types of road damage after introducing DSC alone. For D00 (longitudinal cracks) and D10 (transverse cracks), mAP@50 increased from 56.4% and 56.7% to 61.0% and 59.8%, respectively. This confirms that the design intent of the DSC module is consistent with the geometric characteristics of slender cracks. By designing asymmetric strip convolution to capture the directional extension structure of cracks, the network's sensitivity to directional features is enhanced, thereby improving the model's learning ability for slender cracks. At the same time, the introduction of depthwise separable convolution in the DSC module ensures computational efficiency, adding only about 0.33M parameters, with minimal impact on inference speed. Notably, the DSC module also achieved performance improvements of 5.5% and 5.4% for D20 (cratering) and D40 (potholes), respectively. From a morphological perspective, this is because the cracks are formed by a large number of local linear cracks interwoven together, and the edges of the pits exhibit obvious directional characteristics. Therefore, strip convolution also has an enhancing effect on this type of structure, bringing significant accuracy gains.

[0034] After introducing the MFE module, the AP@50 of each category was further improved compared to the baseline model, especially the D20 (cracked) category, which increased from 64.8% to 71.8%. This indicates that multi-level feature balancing helps improve the feature representation of complex textures and multi-scale damage, thereby alleviating the phenomenon of shallow details being suppressed by deep semantics. Finally, the mAP@50 of the combined DSC and MFE modules reached 62.4%, 60.8%, 72.1%, and 55.9%, respectively, verifying the feature balancing effect of the proposed method.

[0035] In summary, the results of the ablation experiments demonstrate that the DSC module and MFE module proposed in this invention are both effective designs, improving the performance of the baseline model from the aspects of directional feature enhancement and multi-scale feature fusion, respectively, thus verifying the rationality of the proposed method.

[0036] like Figure 4 As shown, compared with existing methods, the method proposed in this invention exhibits more stable detection performance in typical road damage scenarios such as D00 longitudinal cracks, D10 transverse cracks, D20 alligator cracks, and D40 potholes. For slender crack targets, this invention can more accurately locate the crack region and improve detection confidence, indicating that the direction-aware feature enhancement module can effectively capture the continuity and directional texture of cracks. For damages with complex shapes and significant scale differences, such as alligator cracks and potholes, this invention can reduce missed detections and localization errors, indicating that the multi-level feature equalization module helps to enhance the expressive ability of damage features at different scales. Therefore, this invention has advantages in improving the accuracy of road damage detection, enhancing the ability to identify slender cracks, and improving the localization effect of complex damage.

Claims

1. A road damage detection method based on direction perception and feature equalization, comprising: First, a road damage detection model is constructed, consisting of a backbone network, a neck network, and a detection head connected in sequence. Then, the road damage detection model is trained to obtain a trained road damage detection model. Finally, the road image to be detected is input into the trained road damage detection model to obtain the road damage detection result. The specific operational process for obtaining road damage detection results is as follows: Step 1: The backbone network extracts features from the input road image to be detected, obtaining first feature maps, second feature maps, and third feature maps of different scales; Step 2: The neck network performs multi-scale feature fusion on the first, second, and third feature maps output by the backbone network to obtain a balanced feature map; Step 3: Perform bounding box coordinate regression, target confidence prediction, and damage category prediction on the balanced feature map output by the head-to-neck network to obtain road damage detection results containing the location box of the damaged target, damage category, and confidence level. The neck network is characterized by comprising a direction-aware feature enhancement module and a multi-level feature equalization module connected in sequence. The specific operation process of step 2 is as follows: Step 2-1: The orientation-aware feature enhancement module enhances the directional texture features in the first feature map to obtain an enhanced feature map; Step 2-2: The multi-level feature equalization module performs spatial scale alignment, channel dimension unification, and adaptive weighted fusion operations on the enhanced feature map, the second feature map, and the third feature map to obtain the equalized feature map.

2. The road damage detection method based on direction perception and feature equalization according to claim 1, characterized in that... In step 1, the backbone network includes a first CBS layer, a second CBS layer, a first C2f layer, a third CBS layer, a second C2f layer, a third C2f layer, a fourth CBS layer, and a fourth C2f layer connected in sequence. The specific operation process of step 1 is as follows: Step 1-1: The first CBS layer performs downsampling feature extraction on the input road image to be detected to obtain the first intermediate feature map; Steps 1-2: The second CBS layer performs downsampling feature extraction on the first intermediate feature map output by the first CBS layer to obtain the second intermediate feature map; Steps 1-3: The first C2f layer performs feature fusion processing on the second intermediate feature map output by the second CBS layer to obtain the third intermediate feature map; Steps 1-4: The third CBS layer performs downsampling feature extraction on the third intermediate feature map output by the first C2f layer to obtain the fourth intermediate feature map; Steps 1-5: The second C2f layer performs feature fusion processing on the fourth intermediate feature map output by the third CBS layer to obtain a first feature map with a size of 256×80×80. Steps 1-6: The third C2f layer performs feature fusion processing on the first feature map output by the second C2f layer to obtain a second feature map with a size of 512×40×40. Steps 1-7: The fourth CBS layer performs downsampling feature extraction on the second feature map output by the third C2f layer to obtain the fifth intermediate feature map; Steps 1-8: The fourth C2f layer performs feature fusion processing on the fifth intermediate feature map output by the fourth CBS layer to obtain a third feature map with a size of 512×20×20.

3. The road damage detection method based on direction perception and feature equalization according to claim 2, characterized in that, The CBS layer comprises a convolutional layer, a batch normalization layer, and a SiLU activation function layer connected in sequence; the C2f layer comprises a first convolutional layer, a feature splitting unit, at least one bottleneck unit, a feature concatenation unit, and a second convolutional layer.

4. The road damage detection method based on direction perception and feature equalization according to claim 1, characterized in that... The specific operation process of step 2-1 is as follows: Step 2-1-1: Perform channel compression on the first feature map using a 1×1 convolution to obtain a compressed feature map. , ,in, This represents a 1×1 convolution operation. Represents the first feature map. , Indicates the number of input channels. Indicates the height of the first feature map. This represents the width of the first feature map. , Number of output channels For expansion ratio; Step 2-1-2: Compress the feature map using depthwise separable strip convolution. Perform directional decomposition to obtain the lateral feature map. and vertical feature map , , ,in, This represents a depthwise separable stripe convolution operation. For the strip core size; Step 2-1-3, transform the horizontal feature map and vertical feature map After fusion, batch normalization, SiLU activation function, and 3×3 convolution operation are performed sequentially to obtain the integrated feature map. , ,in, This indicates element-wise addition. This indicates a batch normalization operation. This represents the SiLU activation function. This represents a 3×3 convolution operation; Step 2-1-4: When the number of input channels matches the number of output channels, combine the first feature map with the integrated feature map. The enhanced feature map is obtained by adding elements one by one; when the number of input channels is inconsistent with the number of output channels, the feature maps are integrated. As an enhanced feature map.

5. A road damage detection method based on direction perception and feature equalization according to claim 1, characterized in that... The specific operation process of step 2-2 is as follows: Step 2-2-1: Obtain the maximum spatial scale of the enhanced feature map, the second feature map, and the third feature map. Use nearest neighbor interpolation to upsample the enhanced feature map, the second feature map, and the third feature map to the maximum spatial scale to obtain the corresponding scale-aligned feature maps. Step 2-2-2: Use independent 1×1 convolutional layers to unify the number of channels in the scale-aligned feature maps to the same dimension, resulting in channel-unified feature maps. , ,in, This indicates the nearest neighbor interpolation operation. This represents a 1×1 convolution operation used to generate spatial weights. , Represents the enhanced feature map, This represents the second feature map. Represents the third feature map; Step 2-2-3: Perform concatenation along the channel dimension to obtain the concatenated feature map. , ,in, Indicates a splicing operation; Step 2-2-4, stitching feature maps Perform independent 1×1 convolution operations to generate the original weights. , ; Step 2-2-5: Adjust the original weights on the enhanced feature map, the second feature map, and the third feature map respectively. conduct The normalization operation yields the enhanced feature map, the second feature map, and the third feature map in spatial location, respectively. The corresponding spatial weights: , , Among them, satisfying , These represent the original weights corresponding to the enhanced feature map, the second feature map, and the third feature map, respectively. Step 2-2-6, will and Weighted fusion is performed to obtain the fused feature map. , ,in, For element-wise multiplication; Step 2-2-7, using a 1×1 convolutional layer... Channel integration is performed to obtain the equalization feature map. .

6. The road damage detection method based on direction perception and feature equalization according to claim 1, characterized in that... The specific steps for training the road damage detection model are as follows: Images corresponding to four types of targets—longitudinal cracks, transverse cracks, alligator cracks, and potholes—were selected from the public road damage dataset RDD2022, and 80% of the selected images were used as the training set. The training parameters are set as follows: a total of 200 training rounds, a batch size of 16, a weight decay of 0.05, and a momentum of 0.

937. The total loss function is set as the sum of the bounding box loss and the classification loss. Based on the total loss function, the training set is input into the road damage detection model for training. The road damage detection model is updated by the AdamW optimizer until the total number of training rounds is reached, at which point the iterative update stops, and the trained road damage detection model is obtained.

Images