A road defect detection method and system based on intelligent traffic principles for low-altitude multi-weather conditions

Abstract

This invention discloses a road defect detection method and system based on intelligent transportation principles and oriented towards low-altitude and multi-weather conditions. The method includes acquiring road images collected under different weather conditions; inputting the road images into a pre-constructed lightweight real-time detection model for processing to obtain defect detection results; wherein, in the downsampling stage of the backbone network of the lightweight real-time detection model, the input feature map is processed by a lightweight downsampling module; the feature fusion network performs content-aware upsampling and spatial alignment of high-level features through a dynamic adaptive upsampling module; the detection head performs differentiated feature enhancement on the high-resolution feature layer and the low-resolution feature layer respectively through a lightweight detail enhancement detection head, and outputs the road defect category prediction result and bounding box regression result based on the enhanced features. This invention achieves a good balance between road defect detection accuracy, model lightweightness, and operational efficiency under complex weather conditions.

Description

Technical Field

[0001] This invention relates to the fields of intelligent road defect detection and computer vision technology, and in particular to a road defect detection method and system based on the principle of intelligent transportation and oriented towards low-altitude and multi-weather conditions. Background Technology

[0002] Road defect detection is a crucial aspect of road maintenance and traffic safety. Current methods for road defect detection primarily include manual inspection, traditional image processing-based methods, and deep learning-based methods. While manual inspection is intuitive, it suffers from high labor intensity, low efficiency, strong subjectivity, and high operational risks, making it unsuitable for large-scale, high-frequency road inspections. Traditional image processing methods typically rely on manual feature extraction techniques such as edge detection, texture analysis, wavelet transform, and threshold segmentation. Although effective in certain scenarios, these methods are sensitive to changes in lighting, background noise, occlusion interference, and complex road textures, lacking robustness. Furthermore, different defect types often require redesigning features and rules, resulting in high development and maintenance costs.

[0003] In recent years, with the development of deep learning, object detection methods based on convolutional neural networks (CNN) and Transformers have gradually become the mainstream solutions for road defect identification. These methods have end-to-end learning capabilities, enabling automatic feature extraction and improved detection performance. However, in UAV road defect detection scenarios, existing technologies still have significant shortcomings: First, images acquired by UAVs are affected by flight altitude, shooting angle, and platform shaking, resulting in defect targets with large scale spans, irregular shapes, and small local proportions; second, under complex weather conditions such as rain, snow, fog, and low illumination, images are prone to problems such as blurring, reflection, shadow occlusion, and decreased contrast, causing weakened defect boundaries, loss of details, and enhanced background interference; third, UAV platforms have limited computing and storage resources, and existing models often face problems of large parameter count, high computational load, and slow inference speed when pursuing high detection accuracy, making it difficult to meet the needs of real-time deployment.

[0004] Furthermore, existing multi-scale feature fusion methods still have shortcomings in road defect detection tasks. Targets such as road cracks and potholes typically possess characteristics such as small size, slender structure, and irregular boundaries. While shallow features retain more edge and location information, their semantic expressive power is insufficient; deep features, although possessing strong semantic information, are prone to losing fine-grained structural information during continuous downsampling. At the same time, traditional fixed interpolation upsampling methods lack content adaptation capabilities, easily causing spatial misalignment, boundary blurring, and attenuation of key information when fusing high- and low-scale features, especially under complex weather conditions.

[0005] Therefore, existing technologies are unable to simultaneously address the issues of insufficient extraction of road defect features, distortion of cross-scale feature fusion, and insufficient lightweight deployment capabilities under complex meteorological scenarios. Summary of the Invention

[0006] To address the problems of image feature degradation, poor multi-scale feature fusion, and difficulty in balancing detection accuracy and real-time deployment efficiency in existing UAV road defect detection technologies under complex weather conditions, this invention proposes a road defect detection method and system based on intelligent transportation principles and oriented towards low-altitude and multi-weather conditions. The aim is to construct a lightweight road defect detection model that can adapt to complex environments such as rain, snow, fog, and low illumination.

[0007] On the one hand, to achieve the above objectives, this invention provides a road defect detection method based on intelligent transportation principles and oriented towards low-altitude and multi-weather conditions, comprising: Acquire road images under different weather conditions; The road image is input into the pre-built lightweight real-time detection model DDL-YOLO for processing to obtain defect detection results; The DDL-YOLO model comprises a backbone network, a feature fusion network, and a detection head. During the downsampling phase of the backbone network, the input feature map is processed by the lightweight downsampling DPCS_A module to output a multi-layer feature map, enhancing the ability to preserve fine-grained structural information and model channel dependencies. The feature fusion network uses the dynamic adaptive upsampling DSAUpSample module to perform content-aware upsampling and spatial alignment of high-level features in the multi-layer feature map, restoring the feature map resolution and generating multi-scale fused features aligned across scales. The detection head uses the lightweight detail enhancement detection head LDEH to perform differentiated feature enhancement on the high-resolution and low-resolution feature layers, and outputs the category prediction results and bounding box regression results for road defects based on the enhanced features.

[0008] Preferably, the DPCS_A module includes a hybrid pooling unit, a channel partitioning unit, a detail branch, a content branch, a channel splicing unit, a channel rearrangement unit, and a multi-scale channel attention enhancement unit connected in sequence. The hybrid pooling unit is used to fuse average pooling and max pooling operations on the input feature map to obtain a hybrid pooled feature map. The channel partitioning unit is used to divide the hybrid pooled feature map into a first sub-feature and a second sub-feature in the channel dimension; The detail branch is used to perform pointwise convolution on the first sub-feature to extract and retain local spatial location information; The content branch is used to sequentially perform max pooling and depthwise separable convolution on the second sub-feature to extract semantic content features; The channel splicing unit is used to splice and merge the outputs of the detail branch and the content branch in the channel dimension; The channel rearrangement unit is used to perform channel shuffling operations on the spliced ​​and fused feature map; The multi-scale channel attention enhancement unit is used to perform channel reweighting processing on the feature map after channel shuffling based on the fusion of global average pooling and global max pooling.

[0009] Preferably, the multi-scale channel attention enhancement unit extracts global channel descriptors through global average pooling and global max pooling, and adds the two to obtain a fused channel description, specifically: ; In the formula, This is a feature map after channel washing. Description of the fusion channel; The final output of the DPCS_A module is: ; In the formula, This is the final output of the DPCS_A module. This indicates channel-by-channel multiplication. This is a one-dimensional convolution operation. This is the Sigmoid activation function.

[0010] Preferably, the DSAUpSample module includes an offset prediction branch, an initial sampling position construction unit, and a content-aware reconstruction unit; The offset prediction branch is used to predict the sampling offset basis through depthwise separable convolution based on the input high-level feature map. The initial sampling position construction unit is used to construct a regular sub-pixel sampling template as the initial sampling position based on a preset upsampling ratio; The content-aware reconstruction unit is used to combine the initial sampling position with the sampling offset basis to obtain the final sampling offset, and to group and rearrange the input high-level feature map according to the preset number of groups to obtain several grouped features. Then, for each grouped feature, based on the corresponding final sampling offset, adaptive reconstruction is performed using a bilinear sampling function to generate an upsampled feature map.

[0011] Preferably, the DSAUpSample module further includes a dynamic range constraint branch, which is used to generate dynamic range weights based on the input high-level feature map; The content-aware reconstruction unit is used to multiply the dynamic range weights and the sampling offset basis channel by channel, and then combine them with the initial sampling position to obtain the final sampling offset.

[0012] Preferably, the LDEH includes a channel mapping unit, a scaled differential enhancement unit, and a decoupled prediction unit; The channel mapping unit is used to unify input feature maps of different scales to the same hidden dimension through 1×1 convolution; The scale-differentiation enhancement unit is used to enhance local texture and edge structure by employing depthwise separable convolution for high-resolution feature layers; and to supplement local structural information by employing shared detail enhancement convolution for low-resolution feature layers. The decoupled prediction unit is used to perform bounding box regression prediction and category prediction on the enhanced features, respectively.

[0013] On the other hand, to achieve the above objectives, the present invention also provides a road defect detection system based on intelligent transportation principles and oriented towards low-altitude and multi-weather conditions, comprising: The image acquisition module is used to acquire road images collected under different weather conditions; The result output module is used to input the road image into the pre-built lightweight real-time detection model DDL-YOLO for processing to obtain defect detection results; The DDL-YOLO model comprises a backbone network, a feature fusion network, and a detection head. During the downsampling phase of the backbone network, the input feature map is processed by the lightweight downsampling DPCS_A module to output a multi-layer feature map, enhancing the ability to preserve fine-grained structural information and model channel dependencies. The feature fusion network uses the dynamic adaptive upsampling DSAUpSample module to perform content-aware upsampling and spatial alignment of high-level features in the multi-layer feature map, restoring the feature map resolution and generating multi-scale fused features aligned across scales. The detection head uses the lightweight detail enhancement detection head LDEH to perform differentiated feature enhancement on the high-resolution and low-resolution feature layers, and outputs the category prediction results and bounding box regression results for road defects based on the enhanced features.

[0014] Preferably, the DPCS_A module includes: The hybrid pooling unit is used to fuse average pooling and max pooling operations on the input feature map to obtain a hybrid pooled feature map; The channel partitioning unit is used to divide the hybrid pooled feature map into a first sub-feature and a second sub-feature in the channel dimension; The detail branch is used to perform pointwise convolution on the first sub-feature to extract and preserve local spatial location information; The content branch is used to sequentially perform max pooling and depthwise separable convolution on the second sub-feature to extract semantic content features; The channel splicing unit is used to splice and merge the outputs of the detail branch and the content branch in the channel dimension. The channel rearrangement unit is used to perform channel shuffling operations on the spliced ​​and fused feature map; A multi-scale channel attention enhancement unit is used to perform channel reweighting on the feature map after channel shuffling based on a fusion of global average pooling and global max pooling.

[0015] Preferably, the DSAUpSample module includes: The offset prediction branch is used to predict the sampled offset basis through depthwise separable convolution based on the high-level feature map of the input. An initial sampling position construction unit is used to construct a regular sub-pixel sampling template as the initial sampling position based on a preset upsampling ratio; The content-aware reconstruction unit is used to combine the initial sampling position with the sampling offset basis to obtain the final sampling offset, and to group and rearrange the input high-level feature map according to a preset number of groups to obtain several grouped features. Then, for each grouped feature, based on the corresponding final sampling offset, adaptive reconstruction is performed using a bilinear sampling function to generate an upsampled feature map.

[0016] Preferably, the LDEH comprises: The channel mapping unit is used to unify input feature maps of different scales to the same hidden dimension through 1×1 convolution; The scale-differentiated enhancement unit is used to enhance local texture and edge structure by employing depthwise separable convolution for high-resolution feature layers, and to supplement local structural information by employing shared detail enhancement convolution for low-resolution feature layers. The decoupled prediction unit is used to perform bounding box regression prediction and class prediction on the enhanced features respectively.

[0017] Compared with the prior art, the present invention has the following advantages and technical effects: (1) This invention can improve the ability to extract road defect features under complex weather conditions. By introducing a lightweight downsampling module in the downsampling stage of the backbone network, local detail information and high-level semantic information are preserved at the same time during feature compression, and information interaction and dependency modeling between channels are enhanced, thereby effectively improving the model's ability to perceive fine-grained road defect targets such as cracks and potholes, and reducing missed detections and false detections caused by complex backgrounds and severe weather interference.

[0018] (2) The dynamic adaptive upsampling module in this invention achieves adaptive reconstruction oriented towards feature content through dynamic offset learning, rule-based initial sampling position constraints, and content-aware sampling mechanism. Compared with the traditional fixed interpolation upsampling method, this invention can effectively alleviate the problems of spatial misalignment, boundary blurring, and detail attenuation in cross-scale fusion, and is especially beneficial for the detection of targets such as thin cracks and low-contrast pits.

[0019] (3) This invention can improve the discrimination capability of the detection head with lower computational overhead. The lightweight detail-enhanced detection head enhances the local texture modeling capability and the high-level semantic structure expression capability according to the resolution differences of feature layers at different scales, thereby improving the classification and localization performance of multi-scale road defect targets. This detection head reduces the number of model parameters and computational complexity through lightweight design, balancing detection accuracy and deployment efficiency.

[0020] (4) Through the collaborative design of a lightweight downsampling module, a dynamic adaptive upsampling module, and a lightweight detail enhancement detection head, this invention achieves a good balance between the accuracy of road defect detection, the lightweight nature of the model, and the operational efficiency under complex weather conditions, and has strong engineering application value. Attached Figure Description

[0021] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 The diagram below shows the structure of the lightweight real-time detection model DDL-YOLO according to an embodiment of the present invention. (a) is a schematic diagram of the lightweight downsampling DPCS_A module, (b) is a schematic diagram of the dynamic adaptive upsampling DSAUpSample module, and (c) is a schematic diagram of the lightweight detail enhancement detection head LDEH. Detailed Implementation

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0024] This embodiment improves the YOLO11 detection network by refining its downsampling feature extraction structure, upsampling feature reconstruction method, and detection head structure. This enhances the model's ability to identify and locate various types of road defects, such as cracks, potholes, loose surfaces, and repaired areas. While improving the feature expression capability under complex weather conditions, it also reduces the number of model parameters and computational load, improves the model's inference speed and deployment friendliness, and enables it to better meet the application needs of real-time road inspection on UAV platforms.

[0025] Based on the above objectives, this embodiment proposes a road defect detection method based on intelligent transportation principles and oriented towards low-altitude and multi-weather conditions, including: Acquire road images under different weather conditions; The road image is input into the pre-built lightweight real-time detection model DDL-YOLO for processing to obtain defect detection results; The DDL-YOLO model comprises a backbone network, a feature fusion network, and a detection head. During the downsampling phase of the backbone network, the input feature map is processed by the lightweight downsampling DPCS_A module to output a multi-layer feature map, enhancing the ability to preserve fine-grained structural information and model channel dependencies. The feature fusion network uses the dynamic adaptive upsampling DSAUpSample module to perform content-aware upsampling and spatial alignment of high-level features in the multi-layer feature map, restoring the feature map resolution and generating multi-scale fused features aligned across scales. The detection head uses the lightweight detail enhancement detection head LDEH to perform differentiated feature enhancement on the high-resolution and low-resolution feature layers, and outputs the category prediction results and bounding box regression results for road defects based on the enhanced features.

[0026] Specifically, such as Figure 1 The DDL-YOLO model generally consists of three parts: a backbone network, a feature fusion network, and a detection head. By introducing a lightweight downsampling module (Double Path Channel Shuffle with EfficientMultiscale Channel Attention (DPCS_A), a dynamic self-adaptive upsampling module (DSAUpSample), and a lightweight detail-enhanced detection head (LDEH), the accuracy and real-time performance of road defect detection under complex weather conditions are improved.

[0027] Furthermore, the DPCS_A module includes a hybrid pooling unit, a channel partitioning unit, a detail branch, a content branch, a channel splicing unit, a channel rearrangement unit, and a multi-scale channel attention enhancement unit connected in sequence. The hybrid pooling unit is used to fuse average pooling and max pooling operations on the input feature map to obtain a hybrid pooled feature map. The channel partitioning unit is used to divide the hybrid pooled feature map into a first sub-feature and a second sub-feature in the channel dimension; The detail branch is used to perform pointwise convolution on the first sub-feature to extract and retain local spatial location information; The content branch is used to sequentially perform max pooling and depthwise separable convolution on the second sub-feature to extract semantic content features; The channel splicing unit is used to splice and merge the outputs of the detail branch and the content branch in the channel dimension; The channel rearrangement unit is used to perform channel shuffling operations on the spliced ​​and fused feature map; The multi-scale channel attention enhancement unit is used to perform channel reweighting processing on the feature map after channel shuffling based on the fusion of global average pooling and global max pooling.

[0028] Specifically, the backbone network is used to extract hierarchical features from the input UAV road images. A DPCS_A module is introduced during the downsampling stage of the backbone network to enhance the ability to preserve fine-grained structural information and model channel dependencies. Its specific structure is as follows: Figure 1 As shown in (a), the DPCS_A module first processes the input feature map through a hybrid pooling unit. Performing the MixPool operation combines average pooling and max pooling to preserve statistical and significant response information, resulting in a hybrid pooled feature map. : ; In the formula, Let α represent the Sigmoid function, where α is the adaptive learning parameter.

[0029] Based on the channel partitioning unit, the feature map after hybrid pooling is processed. Divided into two sub-features based on channel dimension. : ; at this time, As input for the detail branch This serves as input for the content branch. The purpose of this branch decoupling method is to model local structural information and high-level semantic information separately, avoiding interference between different types of features in a single path.

[0030] The detail branch corresponds to the Position Embedding path. It extracts relatively weak information through pointwise convolution while preserving a large amount of shallow spatial location information. Its main function is to retain local structural information during downsampling, specifically: ; In the formula, This represents the mapping relationship between spatial and semantic information. Output for detail branches (pointwise convolution to extract fine-grained features).

[0031] The content branch corresponds to the MaxPool and Depthwise Convolution (DWConv) paths in the structure diagram. This branch first expands the receptive field through max pooling, and then extracts semantic content features through depthwise separable convolution. It is used to supplement the shortcomings of the detail branch in modeling high-level semantics. Specifically: ; In the formula, Output the content branch.

[0032] After obtaining the outputs of the two branches, they are spliced ​​and merged based on the channel splicing unit.

[0033] The fused features are further processed through a Channel Shuffle operation to enhance information exchange between different channels: ; In the formula, This is a feature map after channel shuffle. Indicates the number of groups.

[0034] Finally, the feature map after channel rearrangement is input into the EMCA unit (i.e., multi-scale channel attention enhancement unit) based on ECA improvement. The EMCA unit first extracts global channel descriptors through global average pooling and global max pooling, and then adds the two to obtain the fused channel descriptor: ; In the formula, This is a feature map after channel shuffle. Description of the fusion channel; Compared with traditional ECA, EMCA unit places greater emphasis on multi-scale channel dependency modeling capabilities, thus enabling it to more effectively highlight road defect-related channels during feature reweighting while suppressing background interference information.

[0035] Subsequently, one-dimensional convolution is used to model local cross-channel dependencies, and channel weights are generated through Sigmoid activation. Finally, the input features are channel-weighted to obtain the module output. : ; In the formula, This is the final output of the DPCS_A module. This indicates channel-by-channel multiplication. This is a one-dimensional convolution operation. This is the Sigmoid activation function.

[0036] Furthermore, the DSAUpSample module includes an offset prediction branch, an initial sampling position construction unit, and a content-aware reconstruction unit; The offset prediction branch is used to predict the sampling offset basis through depthwise separable convolution based on the input high-level feature map. The initial sampling position construction unit is used to construct a regular sub-pixel sampling template as the initial sampling position based on a preset upsampling ratio; The content-aware reconstruction unit is used to combine the initial sampling position with the sampling offset basis to obtain the final sampling offset, and to group and rearrange the input high-level feature map according to the preset number of groups to obtain several grouped features. Then, for each grouped feature, based on the corresponding final sampling offset, adaptive reconstruction is performed using a bilinear sampling function to generate an upsampled feature map.

[0037] Specifically, after processing by the DPCS_A module, the multi-layer features output by the backbone network gain stronger fine-grained structural representation capabilities during the downsampling stage. Let a certain layer of high-semantic, low-resolution feature map be denoted as... This feature is then fed into a feature fusion network to participate in cross-scale feature fusion.

[0038] Since high-level features need to be restored to a higher resolution before being fused with shallow features, traditional fixed interpolation upsampling easily introduces spatial misalignment and boundary blurring during reconstruction, thereby weakening the representation of targets such as thin cracks and low-contrast pits. Therefore, this embodiment proposes the DSAUpSample module, whose structure is as follows: Figure 1 As shown in (b).

[0039] First, input the feature map. Predicting sampling offset basis via depthwise separable convolutional blocks: ; In the formula, denoted as the offset base feature map, s represents the upsampling factor, G represents the number of groups, and coefficient 2 corresponds to the horizontal and vertical coordinate offset in the two-dimensional plane.

[0040] To avoid completely unconstrained offset of sampling points, a regular sub-pixel sampling template is pre-constructed as the initial sampling position. For an upsampling factor s, the initial position is denoted as: ; In the formula, This is the initial position.

[0041] The initial position is copied along the grouping dimension and then unfolded to obtain an initial sampling position encoding consistent with the offset tensor. This initial position provides a rule reference for each output sub-pixel, allowing offset learning to be built on a stable sampling starting point. When the dynamic range branch is enabled, i.e., Branch=True, the input features also generate dynamic range weights through another DSConv branch: ; In the formula, This represents the Sigmoid activation function. This is the dynamic range weight matrix.

[0042] Therefore, the final sampling offset is: ; In the formula, This indicates channel-by-channel multiplication. This is the sampling offset.

[0043] When dynamic range branching is not enabled (Branch=False), the offset basis is constrained only by a fixed scaling factor. In this case, the final sampled offset degenerates into: .

[0044] After obtaining the offset, the module rearranges the input features according to the grouping method and assigns each input position to the grouping. Using the center coordinates as a reference, construct a regular grid coordinate system K: ; The normalized sampling coordinates are then: ; Subsequently, a bilinear approximation function is used to adaptively reconstruct the grouped features to obtain the final sampled output: ; In the formula, For normalized sampling coordinates, This is the final upsampled output feature map of the DSAUpSample module.

[0045] Unlike traditional fixed interpolation upsampling, DSAUpSample achieves adaptive reconstruction based on feature content by dynamically learning the sampling position. This effectively alleviates spatial misalignment, boundary blurring, and detail decay problems in multi-scale fusion, and improves the structure preservation capability in the high-resolution feature recovery stage.

[0046] Furthermore, the LDEH includes a channel mapping unit, a scale-based differential enhancement unit, and a decoupled prediction unit; The channel mapping unit is used to unify input feature maps of different scales to the same hidden dimension through 1×1 convolution; The scale-differentiation enhancement unit is used to enhance local texture and edge structure by employing depthwise separable convolution for high-resolution feature layers; and to supplement local structural information by employing shared detail enhancement convolution for low-resolution feature layers. The decoupled prediction unit is used to perform bounding box regression prediction and category prediction on the enhanced features, respectively.

[0047] Specifically, after processing by the backbone network and the feature fusion network, this embodiment uses the Lightweight Detail Enhancement Detection Head (LDEH) to predict multi-scale features, and its structure is as follows: Figure 1 As shown in (c).

[0048] Compared to traditional detection heads, LDEH's core advantages lie in two aspects: First, through scale-based differential enhancement, high-resolution features can more fully preserve local texture and edge information, while low-resolution features can more effectively supplement semantic structure representation. Second, through lightweight convolutional design and shared enhancement structures, it controls the number of parameters and computational overhead while ensuring prediction quality.

[0049] After fine-grained feature enhancement by the DPCS_A module and adaptive upsampling reconstruction by the DSAUpSample module, the neck output yields a set of multi-scale features with more accurate cross-scale alignment and more complete structural recovery, denoted as follows: , and .in, It has higher spatial resolution, containing richer local textures and edge details. and It has stronger semantic expression and is more conducive to describing the overall structure and contextual information of the target.

[0050] First, the input features at the three scales are each processed through a 1×1 convolution to perform channel mapping, thereby unifying the hidden dimensions and reducing channel differences between different layers. Then, for high-resolution features... ,use Enhance local texture and edge structure representation; for low-resolution features and Adopting a shared approach The structure is enhanced to supplement local structural information in high-level semantic features.

[0051] After scale-based differentiation, the enhanced features can be uniformly represented as: ; In the formula, For the detection head of the first l Enhanced feature map of the layer.

[0052] As can be seen from the above formula, LDEH does not use the same enhancement method for the three scale features. Instead, it enhances local texture modeling and semantic structure modeling according to the resolution difference, thereby improving the detection head's adaptability to targets of different scales.

[0053] After feature enhancement, LDEH employs a decoupled prediction strategy for bounding box regression and class prediction, respectively. The regression branch incorporates a learnable scaling factor for dynamic adjustment after a 1×1 convolution, while the classification branch directly outputs the class response. For the... Each scale feature can be used to predict the following results: ; In the formula, For the detection head l Single-scale detection output of the layer.

[0054] During the training phase, LDEH directly outputs prediction results at three scales.

[0055] During the inference phase, multi-scale predictions are further integrated and bounding box decoding is completed.

[0056] Suppose that the prediction results at the three scales are rearranged and concatenated to obtain a unified representation, then the final output feature map Y can be written as: .

[0057] This embodiment also provides a road defect detection system based on intelligent transportation principles and oriented towards low-altitude and multi-weather conditions, including: The image acquisition module is used to acquire road images collected under different weather conditions; The result output module is used to input the road image into the pre-built lightweight real-time detection model DDL-YOLO for processing to obtain defect detection results; The DDL-YOLO model comprises a backbone network, a feature fusion network, and a detection head. During the downsampling phase of the backbone network, the input feature map is processed by the lightweight downsampling DPCS_A module to output a multi-layer feature map, enhancing the ability to preserve fine-grained structural information and model channel dependencies. The feature fusion network uses the dynamic adaptive upsampling DSAUpSample module to perform content-aware upsampling and spatial alignment of high-level features in the multi-layer feature map, restoring the feature map resolution and generating multi-scale fused features aligned across scales. The detection head uses the lightweight detail enhancement detection head LDEH to perform differentiated feature enhancement on the high-resolution and low-resolution feature layers, and outputs the category prediction results and bounding box regression results for road defects based on the enhanced features.

[0058] Furthermore, the DPCS_A module includes: The hybrid pooling unit is used to fuse average pooling and max pooling operations on the input feature map to obtain a hybrid pooled feature map; The channel partitioning unit is used to divide the hybrid pooled feature map into a first sub-feature and a second sub-feature in the channel dimension; The detail branch is used to perform pointwise convolution on the first sub-feature to extract and preserve local spatial location information; The content branch is used to sequentially perform max pooling and depthwise separable convolution on the second sub-feature to extract semantic content features; The channel splicing unit is used to splice and merge the outputs of the detail branch and the content branch in the channel dimension. The channel rearrangement unit is used to perform channel shuffling operations on the spliced ​​and fused feature map; A multi-scale channel attention enhancement unit is used to perform channel reweighting on the feature map after channel shuffling based on a fusion of global average pooling and global max pooling.

[0059] Furthermore, the DSAUpSample module includes: The offset prediction branch is used to predict the sampled offset basis through depthwise separable convolution based on the high-level feature map of the input. An initial sampling position construction unit is used to construct a regular sub-pixel sampling template as the initial sampling position based on a preset upsampling ratio; The content-aware reconstruction unit is used to combine the initial sampling position with the sampling offset basis to obtain the final sampling offset, and to group and rearrange the input high-level feature map according to a preset number of groups to obtain several grouped features. Then, for each grouped feature, based on the corresponding final sampling offset, adaptive reconstruction is performed using a bilinear sampling function to generate an upsampled feature map.

[0060] Further, the LDEH includes: The channel mapping unit is used to unify input feature maps of different scales to the same hidden dimension through 1×1 convolution; The scale-differentiated enhancement unit is used to enhance local texture and edge structure by employing depthwise separable convolution for high-resolution feature layers, and to supplement local structural information by employing shared detail enhancement convolution for low-resolution feature layers. The decoupled prediction unit is used to perform bounding box regression prediction and class prediction on the enhanced features respectively.

[0061] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A road defect detection method based on intelligent transportation principles and oriented towards low-altitude and multi-weather conditions, characterized in that, include: Acquire road images under different weather conditions; The road image is input into the pre-built lightweight real-time detection model DDL-YOLO for processing to obtain defect detection results; The DDL-YOLO model comprises a backbone network, a feature fusion network, and a detection head. During the downsampling phase of the backbone network, the input feature map is processed by the lightweight downsampling DPCS_A module to output a multi-layer feature map, enhancing the ability to preserve fine-grained structural information and model channel dependencies. The feature fusion network uses the dynamic adaptive upsampling DSAUpSample module to perform content-aware upsampling and spatial alignment of high-level features in the multi-layer feature map, restoring the feature map resolution and generating multi-scale fused features aligned across scales. The detection head uses the lightweight detail enhancement detection head LDEH to perform differentiated feature enhancement on the high-resolution and low-resolution feature layers, and outputs the category prediction results and bounding box regression results for road defects based on the enhanced features.

2. The road defect detection method based on intelligent transportation principles and oriented towards low-altitude multi-meteorological conditions as described in claim 1, characterized in that, The DPCS_A module includes a hybrid pooling unit, a channel partitioning unit, a detail branch, a content branch, a channel splicing unit, a channel rearrangement unit, and a multi-scale channel attention enhancement unit connected in sequence. The hybrid pooling unit is used to fuse average pooling and max pooling operations on the input feature map to obtain a hybrid pooled feature map. The channel partitioning unit is used to divide the hybrid pooled feature map into a first sub-feature and a second sub-feature in the channel dimension; The detail branch is used to perform pointwise convolution on the first sub-feature to extract and retain local spatial location information; The content branch is used to sequentially perform max pooling and depthwise separable convolution on the second sub-feature to extract semantic content features; The channel splicing unit is used to splice and merge the outputs of the detail branch and the content branch in the channel dimension; The channel rearrangement unit is used to perform channel shuffling operations on the spliced ​​and fused feature map; The multi-scale channel attention enhancement unit is used to perform channel reweighting processing on the feature map after channel shuffling based on the fusion of global average pooling and global max pooling.

3. The road defect detection method based on intelligent transportation principles and oriented towards low-altitude multi-weather conditions as described in claim 2, characterized in that, The multi-scale channel attention enhancement unit extracts global channel descriptors through global average pooling and global max pooling, and adds the two to obtain a fused channel descriptor, specifically: ； In the formula, This is a feature map after channel washing. Description of the fusion channel; The final output of the DPCS_A module is: ； In the formula, This is the final output of the DPCS_A module. This indicates channel-by-channel multiplication. This is a one-dimensional convolution operation. This is the Sigmoid activation function.

4. The road defect detection method based on intelligent transportation principles and oriented towards low-altitude multi-meteorological conditions as described in claim 1, characterized in that, The DSAUpSample module includes an offset prediction branch, an initial sampling position construction unit, and a content-aware reconstruction unit. The offset prediction branch is used to predict the sampling offset basis through depthwise separable convolution based on the input high-level feature map. The initial sampling position construction unit is used to construct a regular sub-pixel sampling template as the initial sampling position based on a preset upsampling ratio; The content-aware reconstruction unit is used to combine the initial sampling position with the sampling offset basis to obtain the final sampling offset, and to group and rearrange the input high-level feature map according to the preset number of groups to obtain several grouped features. Then, for each grouped feature, based on the corresponding final sampling offset, adaptive reconstruction is performed using a bilinear sampling function to generate an upsampled feature map.

5. The road defect detection method based on intelligent transportation principles and oriented towards low-altitude multi-meteorological conditions according to claim 4, characterized in that, The DSAUpSample module also includes a dynamic range constraint branch, which is used to generate dynamic range weights based on the input high-level feature map. The content-aware reconstruction unit is used to multiply the dynamic range weights and the sampling offset basis channel by channel, and then combine them with the initial sampling position to obtain the final sampling offset.

6. The road defect detection method based on intelligent transportation principles and oriented towards low-altitude multi-meteorological conditions according to claim 1, characterized in that, The LDEH includes a channel mapping unit, a scale-based differential enhancement unit, and a decoupled prediction unit; The channel mapping unit is used to unify input feature maps of different scales to the same hidden dimension through 1×1 convolution; The scale-differentiation enhancement unit is used to enhance local texture and edge structure by employing depthwise separable convolution for high-resolution feature layers; and to supplement local structural information by employing shared detail enhancement convolution for low-resolution feature layers. The decoupled prediction unit is used to perform bounding box regression prediction and category prediction on the enhanced features, respectively.

7. A road defect detection system based on intelligent transportation principles and oriented towards low-altitude and multi-weather conditions, used to implement the road defect detection method based on intelligent transportation principles and oriented towards low-altitude and multi-weather conditions as described in any one of claims 1-6, characterized in that, include: The image acquisition module is used to acquire road images collected under different weather conditions; The result output module is used to input the road image into the pre-built lightweight real-time detection model DDL-YOLO for processing to obtain defect detection results; The DDL-YOLO model comprises a backbone network, a feature fusion network, and a detection head. During the downsampling phase of the backbone network, the input feature map is processed by the lightweight downsampling DPCS_A module to output a multi-layer feature map, enhancing the ability to preserve fine-grained structural information and model channel dependencies. The feature fusion network uses the dynamic adaptive upsampling DSAUpSample module to perform content-aware upsampling and spatial alignment of high-level features in the multi-layer feature map, restoring the feature map resolution and generating multi-scale fused features aligned across scales. The detection head uses the lightweight detail enhancement detection head LDEH to perform differentiated feature enhancement on the high-resolution and low-resolution feature layers, and outputs the category prediction results and bounding box regression results for road defects based on the enhanced features.

8. The road defect detection system based on intelligent transportation principles and oriented towards low-altitude multi-weather conditions as described in claim 7, characterized in that, The DPCS_A module includes: The hybrid pooling unit is used to fuse average pooling and max pooling operations on the input feature map to obtain a hybrid pooled feature map; The channel partitioning unit is used to divide the hybrid pooled feature map into a first sub-feature and a second sub-feature in the channel dimension; The detail branch is used to perform pointwise convolution on the first sub-feature to extract and preserve local spatial location information; The content branch is used to sequentially perform max pooling and depthwise separable convolution on the second sub-feature to extract semantic content features; The channel splicing unit is used to splice and merge the outputs of the detail branch and the content branch in the channel dimension. The channel rearrangement unit is used to perform channel shuffling operations on the spliced ​​and fused feature map; A multi-scale channel attention enhancement unit is used to perform channel reweighting on the feature map after channel shuffling based on a fusion of global average pooling and global max pooling.

9. The road defect detection system based on intelligent transportation principles and oriented towards low-altitude multi-weather conditions as described in claim 7, characterized in that, The DSAUpSample module includes: The offset prediction branch is used to predict the sampled offset basis through depthwise separable convolution based on the high-level feature map of the input. An initial sampling position construction unit is used to construct a regular sub-pixel sampling template as the initial sampling position based on a preset upsampling ratio; The content-aware reconstruction unit is used to combine the initial sampling position with the sampling offset basis to obtain the final sampling offset, and to group and rearrange the input high-level feature map according to a preset number of groups to obtain several grouped features. Then, for each grouped feature, based on the corresponding final sampling offset, adaptive reconstruction is performed using a bilinear sampling function to generate an upsampled feature map.

10. The road defect detection system based on intelligent transportation principles and oriented towards low-altitude multi-weather conditions according to claim 7, characterized in that, The LDEH includes: The channel mapping unit is used to unify input feature maps of different scales to the same hidden dimension through 1×1 convolution; The scale-differentiated enhancement unit is used to enhance local texture and edge structure by employing depthwise separable convolution for high-resolution feature layers, and to supplement local structural information by employing shared detail enhancement convolution for low-resolution feature layers. The decoupled prediction unit is used to perform bounding box regression prediction and class prediction on the enhanced features respectively.

Images