A pavement damage classification method, system, device and storage medium

Abstract

The application provides a kind of road surface damage classification method, system, equipment and storage medium, belong to road management technical field, including obtaining road surface damage image;Road surface damage image is input into the damage detail fusion DDFM model trained, and the road surface damage classification result is obtained;DDFM model is with MambaOut model as main network, CAMFI-STEM layer and CAMFI-down sampling layer are used to replace the stem layer and down sampling layer of original MambaOut model, and the detail feature extraction related to damage is enhanced;And introduce damage information fusion DIF module to replace the gating CNN module of original MambaOut model, use variance guided fusion mechanism to integrate space, frequency domain and semantic clues, to improve fine-grained recognition ability.The efficiency and accuracy of road surface damage classification are improved.

Description

Technical Field

[0001] This invention belongs to the field of road management technology, specifically relating to a method, system, device, and storage medium for classifying road surface damage. Background Technology

[0002] With continuous economic development, dense and efficient road networks have become critical infrastructure. However, frequent road use leads to pavement damage, affecting driving safety and traffic efficiency. Therefore, achieving automatic detection and classification of pavement damage is crucial for timely maintenance. Traditional manual visual inspection methods suffer from high labor intensity, long processing times, and strong subjectivity, resulting in low detection efficiency and accuracy. With advancements in sensing technology, automated detection systems have gradually become a research hotspot. Vibration sensors and gyroscopes can detect pavement damage by monitoring abnormal vehicle movements, but these methods are only suitable for identifying large-scale irregular damage and often fail to detect early micro-cracks. To improve detection accuracy, three-dimensional laser scanning technology is introduced to calculate average texture depth, and algorithms based on minimum spanning trees are used to enhance damage feature analysis capabilities. However, three-dimensional scanning systems are costly and not suitable for lightweight or low-cost practical applications.

[0003] In recent years, the rapid development of deep learning has greatly promoted the advancement of image processing technology, providing effective solutions for road damage detection. Various methods for road damage detection using deep learning have been proposed, including: using Faster R-CNN to detect cracks and concrete defects; and proposing a multi-scale fusion model w-SegNet based on SegNet, which demonstrates good robustness. Meanwhile, the DeepLab series models using dilated convolutions have also achieved significant improvements in crack detection tasks. However, methods based on convolutional neural networks (CNNs), due to their reliance on local convolutions and downsampling operations, are mainly good at modeling short-range dependencies and struggle to capture global contextual information. Although dilated convolutions and attention mechanisms alleviate this problem to some extent, their core limitations remain. In contrast, Transformer models model spatial information through positional encoding and utilize stacked structures to achieve global dependency modeling, demonstrating strong potential in the field of image recognition. Existing technologies have improved the fine-grained capability of crack detection by leveraging attention mechanisms; Swin-Transformer, by fusing CNN and Transformer features, maintains high accuracy while taking into account both local and global information; a weakly supervised model based on SwinTransformer is also proposed, which uses teacher labels to generate pseudo-labels and learn discriminative features. Although the Transformer architecture is better at capturing long-range context and multimodal features, it still faces certain difficulties in dealing with large-scale changing scenes, which limits its performance under complex and diverse road damage conditions.

[0004] The methods described above explore the problem of road damage detection and classification from multiple perspectives. However, compared with general images, road damage images have unique challenges: the damaged area is small, the background is complex, and it is often accompanied by shadows or markings. The visual differences between different categories are also low. These factors seriously affect the recognition accuracy of the classification model. Summary of the Invention

[0005] To address the shortcomings of existing deep learning methods in road damage detection, this invention provides a road damage classification method, system, device, and storage medium.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for classifying road surface damage includes the following steps: Acquire images of road surface damage; The road damage image is input into the trained damage detail fusion DDFM model to obtain the road damage classification result. The DDFM model uses the MambaOut model as the backbone network, replaces the stem layer and downsampling layer of the original MambaOut model with CAMFI-STEM layer and CAMFI-downsampling layer, and introduces the damage information fusion DIF module to replace the gated CNN module of the original MambaOut model. In the DDFM model for damage detail fusion after training, the CAMFI-Stem layer is used to extract preliminary features from the input pavement damage image. The extracted preliminary features are then input into the DIF module for damage feature capture to obtain a fused feature representation. The CAMFI-downsampling layer and the DIF module are combined to perform multiple feature enhancements on the fused feature representation. The enhanced feature map is then input into the classification head to obtain the pavement damage classification result.

[0007] Preferably, the CAMFI-STEM layer uses the original stem layer as the main path, CAMixer as the branch path, and then uses MFII to fuse the two paths; the main path of the CAMFI module performs content-aware convolution, predicts the position to be focused on through a mask-guided attention mechanism, and adaptively focuses on the damage area; the CAMixer module extracts fine-grained features, and then performs cross-path fusion through the MFII module.

[0008] Preferably, the processing procedures of the CAMixer module and the MFII module in the CAMFI-STEM layer are as follows: , ; ; ; Where Q and K represent the Query and Key matrices used in attention calculation, respectively, and W Q and W K These are their respective projection weights, where d represents the dimension of the attention projection, and X... out X represents the final output of the CAMFI module. St This indicates the output of the main path, and M(·) represents the feature fusion operation performed by the MFII module.

[0009] Preferably, the DIF module uses the original gated CNN module as the main path, integrates detail enhancement convolution (DEConv), coordinate attention mechanism and Fourier transform as auxiliary branches, and then uses the SMFA module to fuse the outputs of the main path and auxiliary branches.

[0010] Preferably, in the DIF module, coarse feature extraction is performed using the original gated CNN module; in the auxiliary branch, DEConv differential learning is used to enhance and impair features, and the horizontal and vertical position information is encoded into channel attention calculation through the dynamic region awareness mechanism DRAM to obtain spatial features with position encoding. The channel attention is modeled in the frequency domain using the Fourier transform module to obtain frequency domain features; the SMFA module guides the feature fusion of the spatial features with position encoding and the frequency domain features based on the variance adjustment mechanism of region content change, and outputs the enhanced features.

[0011] Preferably, when training the damage detail fusion DDFM model, Mixup or CutMix is ​​applied to augment the training set data, and the model parameters are optimized by combining the cross-entropy loss function.

[0012] The present invention also provides a road surface damage classification system, specifically including: The data acquisition module is used to acquire images of road surface damage.

[0013] The damage classification module is used to input the road damage image into the trained damage details DDFM model to obtain the road damage classification result. The DDFM model uses the MambaOut model as the backbone network, replaces the stem layer and Dowunsampling layer of the original MambaOut model with CAMFI-STEM layer and CAMFI-Dowunsampling layer, and introduces the damage information fusion DIF module. In the DDFM model of damage detail fusion after training, the CAMFI-Stem layer is used to extract preliminary features from the input pavement damage image. The extracted preliminary features are then input into the DIF module for damage feature capture to obtain the fused feature representation. The CAMFI-downsampling layer and the DIF module are combined to perform multiple feature enhancements on the fused feature representation. The enhanced feature map is then input into the classification head to obtain the pavement damage classification result.

[0014] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps described in the road surface damage classification method.

[0015] The present invention also provides a computer-readable storage medium storing a computer program that, when loaded by a processor, can execute the steps described in the road surface damage classification method.

[0016] The road surface damage classification method provided by this invention has the following beneficial effects: This invention designs a Content-Aware Hybrid Feature Interaction Module (CAMFI), replacing the stem and downsampling layers of the original MambaOut model with CAMFI-STEM and CAMFI-Downsampling layers. CAMFI introduces an attention-guided fusion mechanism, enhancing the network's focus on subtle and localized damage regions, thereby improving feature quality and spatial discriminative ability. To address background noise, shadow interference, and the high visual similarity between crack types, a Damage Information Fusion Module (DIF) is further proposed. DIF fuses spatial, frequency, and dimensional cues and employs a variance-guided fusion strategy to enhance key damage features while suppressing redundant or misleading contextual information. By combining CAMFI and DIF, DDFM effectively addresses key challenges in pavement damage classification. Attached Figure Description

[0017] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is an overall architecture diagram of the DDFM model according to an embodiment of the present invention, wherein, Figure 1 (a) is the CAMFI-Stem layer; Figure 1 (b) is the DIF module; Figure 1 (c) is the CAMFI-downsampling layer.

[0019] Figure 2 This is a structural diagram of the CAMFI module according to an embodiment of the present invention, wherein, Figure 2 (a) is the content-aware module; Figure 2 (b) is the feature interaction module between the main path and the auxiliary path.

[0020] Figure 3 This is a structural diagram of the DIF module according to an embodiment of the present invention, wherein, Figure 3 (a) represents convolution; Figure 3 (b) represents coordinate attention; Figure 3 (c) is a Fourier model; Figure 3 (d) represents nonlocal branch extraction; Figure 3 (e) represents local branch extraction.

[0021] Figure 4 This is a comparison of visual heatmaps in the embodiments of the present invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0023] Example This invention provides a method for classifying road surface damage, specifically as follows: Figure 1 As shown, it includes the following steps: Step 1: Obtain road surface damage images.

[0024] Step 2: Constructing a Damage Details Fusion (DDFM) Model. Using the MambaOut model as the backbone network, a DDFM model was designed to adapt to the feature distribution of road surface damage. Two key modules were introduced, utilizing the Content-Aware Hybrid Feature Interaction (CAMFI) module (i.e.,...). Figure 1 The CAMFI-STEM layer and CAMFI-downsampling layer in the original MambaOut model replace the stem layer and downsampling layer, and the damage information fusion (DIF) module replaces the gated CNN module of the original MambaOut model.

[0025] The CAMFI module enhances attention to small-scale, localized defects. The DIF module integrates spatial, frequency, and semantic cues, effectively suppressing background noise and enhancing feature discrimination. The collaborative work of these two modules enables DDFM to achieve robust and accurate classification performance under diverse road surface conditions.

[0026] The structure of the Content-Aware Hybrid Feature Interaction (CAMFI) module is as follows: Figure 2As shown, CAMFI is designed to enhance the Stem and downsampling layers of MambaOut. In the original architecture, these layers use standard convolutions for feature extraction and resolution reduction. However, due to the low proportion of damaged areas and the small, elongated crack structures in road images, standard convolutions often fail to capture sufficient discriminative features. The main path of the CAMFI-stem layer is the original stem layer, using content-aware convolution CAMixer as a branch path to compensate for the fine-grained features ignored by the main path, and then using MFII to fuse the two paths. Road images typically consist of simple regions (denoted as Xs, mainly normal regions) and complex regions (denoted as Xh, corresponding to damaged regions). During feature extraction, the CAMFI module performs content-aware convolutions, adaptively focusing on damaged regions without significantly increasing computational overhead. This process is achieved through a mask m, which guides the attention mechanism to predict the locations to focus on, enabling the model to automatically highlight damaged regions. CAMixer serves as a parallel branch for extracting fine-grained features. This branch supplements the details that the standard convolutions relied upon by the main path might miss. The outputs of the two paths are then fused using the MFII module. The MFII module performs cross-path fusion, enabling parallel branches to recover fine-grained details missed by the main path, thereby enhancing the ability to distinguish similar crack types. Combined with CAMixer's attention filtering, this module reduces false positives from complex backgrounds. This process is formally defined in equations (1) to (3), where Q and K represent the Query and Key matrices used in the attention calculation, respectively, and W... Q and W K These are their respective projection weights, where d represents the dimension of the attention projection. X out X represents the final output of the CAMFI module. St This indicates the output of the main path, and M(·) represents the feature fusion operation performed by the MFII module.

[0027] , (1); (2); (3).

[0028] The structure of the Damage Information Fusion (DIF) module is as follows: Figure 3As shown, DIF is used to enhance the gated CNN module in the original MambaOut. The gated CNN module has difficulty handling complex situations such as overlapping damage and shadows, lane marking occlusion, or extremely fine cracks. To address this issue, the original gated CNN module is used as the main path, and detailed enhancement convolution (DEConv), coordinate attention mechanism, and Fourier transform are integrated as auxiliary branches. Then, the SMFA module is used to exchange information between the main path and the auxiliary branches, resulting in the Damage Information Fusion (DIF) module, which enhances feature extraction and fusion capabilities.

[0029] The original gated CNN module in MambaOut performs coarse feature extraction, but performs poorly when dealing with complex situations such as damage overlapping with shadows or lane markings. Uneven lighting and background noise blur the edges of damage, limiting the ability to express features. DEConv is introduced to enhance damage-related features through differential learning, focusing on fine-grained details rather than the entire image. At the same time, a Dynamic Region Awareness (DRAM) mechanism is further added, which combines spatial and channel attention to generate fine channel-specific spatial weights. To improve localization capabilities, coordinate attention is integrated, encoding horizontal and vertical position information into channel attention, guiding DRAM to generate position-aware weights, helping the network focus on the real damage area and suppress background noise, thereby enhancing feature discrimination capabilities. This process is formally defined in equations (4) to (6), where, Represents dynamic attention weights. Represents DRAM functions. Represents static weights. This represents a function that incorporates positional encoding into a dynamic attention mechanism.

[0030] (4); (5); (6); To address the classification difficulties caused by the high visual similarity of crack types, a Fourier transform module is introduced. Although DRAM and coordinate attention focus on the spatial and channel domains, they easily overlook frequency information, which is crucial for capturing subtle structural differences, especially in the mid-to-high frequency range. FCA models channel attention in the frequency domain, helping the network focus on channels with strong discriminative responses. In addition, the self-modulated feature aggregation (SMFA) module coordinates feature fusion between the main branch and the auxiliary branch. The main branch models the low-frequency nonlocal structure to obtain global context information; the auxiliary branch fuses spatial coordinates and frequency features to enhance the specific representation of local damage. A variance adjustment mechanism based on regional content changes guides the fusion process to generate the final output. This process is formally defined in formulas (7) to (10), where This represents spatial features with location encoding. This represents the corresponding frequency domain characteristics. For frequency coordinates. The term converts spatial information into a complex-valued frequency response. Capture channel attention in the frequency domain. and These represent the features used for local modeling and augmentation modeling, respectively. This represents a feature fusion function based on variance guidance.

[0031] (7); (8); (9); (10); The DIF module integrates location, frequency, and dimensionality information into the final feature representation, enhancing the model's ability to capture damage-related features and thus supporting high-precision classification tasks.

[0032] Step 3: Input the road damage image into the damage detail fusion model for training. In the first stage, preliminary feature extraction is performed through the CAMFI-Stem layer. The main path uses the original Stem structure, while the auxiliary path uses CAMixer to enhance the focus on fine-grained damage. The output is fused through MFII and then processed through DIF blocks, which enrich the features with spatial, frequency, and dimensional information. From the second to the fourth stage, downsampling is performed using the CAMFI-downsampling layer, which optimizes the original downsampling process by preserving the features related to the damage. Each stage of the CAMFI-downsampling layer is followed by a DIF block to strengthen the feature representation. The final output is generated by the classification head. The overall process is formalized in equations (11) to (13), where Indicates the input image. This is the output of the first stage. and These represent the original Stem and the downsampling operation, respectively. Indicates CAMFI processing, Indicates the DIF operation. This indicates feature fusion via MFII. It is a classification header. The final predicted category label includes normal, longitudinal crack, transverse crack, and repair.

[0033] (11); (12); (13); To improve generalization in complex scenarios, Mixup or CutMix is ​​applied during training, combined with cross-entropy loss. The soft labels generated by these data augmentation methods can better reflect instances with blurred boundaries or mixed categories. The loss calculation formula is shown in Equation (14).

[0034] (14).

[0035] The proposed damage detail fusion model is compared with mainstream image classification models. Multiple datasets of different sizes, image types, and damage categories are used, including CQU-BPMDD, Crack500-PDD, and CQU-BPMDD. Detailed information about the datasets is shown in Table 1.

[0036] Table 1 Dataset Information To evaluate the individual impacts of the CAMFI and DIF modules on pavement damage classification, as well as their overall contribution to the performance of the proposed model, an ablation study was conducted. Table 2 presents the ablation results for the classification task. CAMFI and DIF each improve performance individually, and their combination achieves the best results in DDFM compared to baseline methods. As shown in Table 3, DDFM consistently outperforms baseline methods on datasets of varying sizes, including detection tasks on smaller datasets. These results confirm that the joint integration of CAMFI and DIF significantly improves the performance and adaptability of DDFM.

[0037] Table 2 Results of Classification Task Ablation Experiment Table 3. Results of ablation experiments for detection tasks To evaluate the detection performance of the DDFM model, it was compared with representative models PicT, DMTC, and DDACDN. As shown in Table 4, DDFM achieved a 4.16% improvement in AUC on CQU-BPMDD and a 1.96% improvement over the baseline on CQU-BPDD, with improvements in all other evaluation metrics. These results confirm the superior performance of DDFM compared to similar models in pavement damage detection tasks.

[0038] Table 4 Experimental Results of the Detection Task For the classification task, DDFM was evaluated on the CQU-BPMDD dataset and compared with several mainstream models. The results are shown in Table 5. Due to the unique challenges of pavement damage classification, general-purpose models typically perform poorly. In contrast, DDFM introduces two tailored modules: CAMFI, which leverages content-aware attention to enhance damage feature extraction; and DIF, which fuses spatial, frequency, and dimensional cues to provide richer representations. DDFM outperforms all competing models, improving accuracy by 3.16%, precision by 3.91%, and F1-score by 3.83% compared to the baseline, validating its effectiveness in this task.

[0039] Table 5 Experimental Results of Mainstream Classification Models In addition to comparisons with mainstream models, DDFM was also compared with task-specific methods, including PicT, DMTC, and DDACDN, and evaluated using dual datasets CQU-BPMDD and CQU-BPDD. As shown in Table 6, DDFM consistently outperformed all peer models across all metrics on both datasets. On CQU-BPDD, it achieved a 2.16% improvement in accuracy, a 1.15% improvement in precision, and a 1.8% improvement in F1-score compared to the strongest baseline, validating its robustness and generalization ability.

[0040] Table 6 Experimental Results of Road Surface Damage Classification Model like Figure 4 As shown, Grad-CAM is used to visualize the decision-making processes of DDFM and MambaOut using heatmaps. These heatmaps highlight the regions each model focuses on during the classification process. In challenging cases, DDFM demonstrates better localization and recognition capabilities, accurately identifying damage areas and outperforming MambaOut.

[0041] The present invention also provides a road surface damage classification system, comprising: The data acquisition module is used to acquire images of road surface damage.

[0042] The damage classification module is used to input road damage images into the trained damage details DDFM model to obtain road damage classification results. The DDFM model uses the MambaOut model as the backbone network, replaces the stem layer and Dowunsampling layer of the original MambaOut model with CAMFI-STEM layer and CAMFI-Dowunsampling layer, and introduces the damage information fusion DIF module. In the DDFM model of damage detail fusion after training, the CAMFI-Stem layer is used to extract preliminary features from the input pavement damage image. The extracted preliminary features are then input into the DIF module for damage feature capture to obtain the fused feature representation. The CAMFI-downsampling layer and the DIF module are combined to perform multiple feature enhancements on the fused feature representation. The enhanced feature map is then input into the classification head to obtain the pavement damage classification result.

[0043] The modules in the aforementioned pavement damage classification system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0044] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps in an embodiment of a road surface damage classification method. Specific implementation methods can be found in the method embodiments, and will not be repeated here.

[0045] Furthermore, the present invention also provides a non-transitory computer-readable storage medium containing instructions on which a computer program is stored. For example, a memory containing instructions that can be executed by a processor of a computer device to perform the above-described method. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc. When the computer program is executed by the processor, it can implement the steps in an embodiment of a road surface damage classification method. Specific implementation methods can be found in the method embodiments, which will not be repeated here.

[0046] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0047] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0048] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0049] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0050] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail in this specification and embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered within the protection scope of the present invention patent. No reference numerals in the claims should be construed as limiting the scope of the claims. Any simple variations or equivalent substitutions of technical solutions that can be readily obtained by those skilled in the art within the scope of the technology disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A method for classifying road surface damage, characterized in that, Includes the following steps: Acquire images of road surface damage; The road damage image is input into the trained damage detail fusion DDFM model to obtain the road damage classification result. The DDFM model uses the MambaOut model as the backbone network, replaces the stem layer and downsampling layer of the original MambaOut model with CAMFI-STEM layer and CAMFI-downsampling layer, and introduces the damage information fusion DIF module to replace the gated CNN module of the original MambaOut model. In the DDFM model for fusion of damage details after training, the CAMFI-Stem layer is used to extract preliminary features from the input pavement damage image. The extracted preliminary features are then input into the DIF module for damage feature capture to obtain the fused feature representation. The fused feature representation is enhanced multiple times using a combination of CAMFI-downsampling layer and DIF module. The enhanced feature map is then input into the classification head to obtain the road damage classification result.

2. The method for classifying road surface damage according to claim 1, characterized in that, The CAMFI-STEM layer uses the original stem layer as the main path, CAMixer as the branch path, and then uses MFII to fuse the two paths. The main path of the CAMFI module performs content-aware convolution, predicts the position to be focused on through a mask-guided attention mechanism, and adaptively focuses on the damage area. The CAMixer module extracts fine-grained features, and then performs cross-path fusion through the MFII module.

3. The method for classifying road surface damage according to claim 2, characterized in that, The specific processing procedures of the CAMixer module and MFII module in the CAMFI-STEM layer are as follows: , ； ； ； Where Q and K represent the Query and Key matrices used in attention calculation, respectively, and W Q and W K These are their respective projection weights, where d represents the dimension of the attention projection, and X... out X represents the final output of the CAMFI module. St This indicates the output of the main path, and M(·) represents the feature fusion operation performed by the MFII module.

4. The method for classifying road surface damage according to claim 1, characterized in that, The DIF module uses the original gated CNN module as the main path, integrates detail enhancement convolution (DEConv), coordinate attention mechanism and Fourier transform as auxiliary branches, and then uses the SMFA module to fuse the outputs of the main path and auxiliary branches.

5. The method for classifying road surface damage according to claim 4, characterized in that, In the DIF module, coarse feature extraction is performed using the original gated CNN module; in the auxiliary branch, DEConv differential learning is used to enhance and impair features, and the horizontal and vertical position information is encoded into the channel attention calculation through the dynamic region awareness mechanism DRAM to obtain spatial features with position encoding. The channel attention is modeled in the frequency domain using the Fourier transform module to obtain frequency domain features. The SMFA module guides the feature fusion of the spatial features with location coding and the frequency domain features based on the variance adjustment mechanism of regional content changes, and outputs the enhanced features.

6. The method for classifying road surface damage according to claim 1, characterized in that, When training the DDFM model for fusion of damage details, Mixup or CutMix is ​​applied to augment the training data, and the model parameters are optimized by combining the cross-entropy loss function.

7. A road surface damage classification system, characterized in that, include: The data acquisition module is used to acquire images of road surface damage. The damage classification module is used to input the road damage image into the trained damage details DDFM model to obtain the road damage classification result. The DDFM model uses the MambaOut model as the backbone network, replaces the stem layer and Dowunsampling layer of the original MambaOut model with CAMFI-STEM layer and CAMFI-Dowunsampling layer, and introduces the damage information fusion DIF module. In the DDFM model of damage detail fusion after training, the CAMFI-Stem layer is used to extract preliminary features from the input road damage image. The extracted preliminary features are then input into the DIF module for damage feature capture to obtain the fused feature representation. The fused feature representation is enhanced multiple times using a combination of CAMFI-downsampling layer and DIF module. The enhanced feature map is then input into the classification head to obtain the road damage classification result.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to perform the steps of the method according to any one of claims 1 to 6.

Images