Dynamic multi-scale fusion network for road damage detection in complex street views Authors: Hong Zhou, Yiyang Qin, Jiahuan Ren, Yan Zhang, Shangbing Gao Published: 2026 ABSTRACT AbstractWith the escalating deterioration of road surfaces, caused by surging traffic volumes and changing climatic conditions, road damage detection (RDD) has emerged as a critical task for ensuring traffic safety and prolonging pavement service life. However, in complex street scenes, the existing RDD methods still present numerous challenges: 1) The damage dimensions may vary drastically from slender cracks to large-area potholes, which leads to difficulties in multi-scale feature extraction; 2) narrow damages, such as cracks, usually exhibit curved and irregular geometric characteristics, which will complicate accurate localization. To tackle the above challenges, we propose a new RDD method based on YOLOv11 to improve the feature extraction capability in the backbone and neck, called dynamic multi-scale fusion network (DMSFNet). To make the detection of road damage more accurate on vast scale variation and slender morphological patterns, DMSFNet introduces a group fusion block (GFB) and a multi-branch feature fusion (MBFF) module. The GFB with multi-scale hybrid convolution architecture integrates grouping and aggregation mechanisms to explicitly model discrete scale-specific features, thereby significantly enhancing accuracy for damage of all sizes. MBFF employs three branches for the contextual features of the surrounding environment and geometric features of damage by dynamic snake convolution. Additionally, a lightweight efficient channel attention (ECA) is used for fusing the multi-branch features to enhance representation capability. Extensive experiments on the SVRDD, RDD2020, and USRDD datasets demonstrate the effectiveness of our method, which achieves a favorable trade-off between detection accuracy and computational efficiency. The employed datasets can be obtained by RDD2020 (https://data.mendeley.com/datasets/5ty2wb6gvg/1), SVRDD (https://zenodo.org/records/10100129) and USRDD (https://doi.org/10.6084/m9.figshare.21431547).Subject terms: Engineering, Mathematics and computing Introduction Roads are vital infrastructure that supports connectivity and trade, playing a key role in economic development. However, increasing traffic and changing climate conditions have led to various types of road damage, such as potholes, transverse cracks, longitudinal cracks, and alligator cracks1, as illustrated in Fig. 1. These defects can severely compromise driving safety, making effective road damage detection essential for timely maintenance. Some examples of the classic road damage. Early detection methods relied heavily on manual image annotation, which incurred significant labor and material costs2–5. With technological advances, automated approaches using laser scanning, infrared thermal imaging, and ground-penetrating radar have been proposed6,7. However, the high cost of multi-sensor inspection vehicles limits their routine application. In recent years, convolutional neural networks (CNNs) have been widely adopted for road damage detection due to their powerful feature extraction capabilities8–10. Early studies have demonstrated the effectiveness of CNNs for crack detection tasks. Cha et al.11 proposed a deep learning-based crack damage detection method using CNNs, which is considered one of the pioneering works in this field. These early explorations laid the foundation for subsequent research on automated pavement distress analysis. Methods generally fall into three categories: image classification-based12, semantic segmentation-based13,14, and object detection-based approaches15. Image classification-based methods typically take whole images or patches as input and use deep learning models to classify whether damage is present. For example, Wang16 applied a CNN with a sliding window to detect damage in masonry structures, while Alqethami et al.17 proposed RoadNet, a simple CNN architecture with customized layers for road damage classification. However, these methods struggle to accurately determine the location, extent, and shape of damage in real-world scenarios. Semantic segmentation-based methods partition input images into non-overlapping, semantically meaningful regions to extract pixel-level details of road defects, such as location, shape, and area. For instance, Fan et al.18 proposed an unsupervised disparity map segmentation method to differentiate between damaged and intact road surfaces. To address issues like limited receptive fields and information loss from pooling, Yang et al.19 introduced a Multi-Scale Triple-attention network (MST-Net) that integrates multi-scale inputs and attention mechanisms for more accurate crack segmentation. However, despite its precision, pixel-level segmentation remains time-consuming20. Compared with the above two types of methods, object detection based methods not only ascertain the presence of defects in an image but also precisely pinpoint their locations. Representative target detection models include Faster R-CNN21, SSD22, and the YOLO family23, etc. Cha et al.24 introduced architectures such as Faster R-CNN to achieve bounding box localization for multiple types of damages. However, the computational complexity of this method is high, and it is not effective in detecting small damages. Therefore25, proposed an enhanced SSD model, which employed receptive field block to improve detection speed and enhanced the ability of handling various cracks by multi-scale feature extraction. To obtain more accurate features of the large-scale variations and unobvious damages in complex environments26, employed Bidirectional Feature Pyramid Network (BiFPN)27 to reconstruct the neck structure and proposed Large Separable Kernel Attention (LSK attention) mechanism with dynamic large convolutional kernels to improve the detection accuracy28 proposed the Refined Switchable Atrous Convolution (RSAC) module to adaptively adjust receptive field to the scale of object. Although the above methods have achieved promising performance in detecting multi-scale road damage, the diversity and complexity of background disturbances in complex street views still pose some significant challenges: 1) Road damages usually vary widely in scale, ranging from millimeter-scale linear cracks to meter-level potholes, which requires models to be simultaneously sensitive to both fine-grained texture and macroscopic structural anomalies; 2) narrow cracks usually exhibit nonlinear, fragmented patterns with abrupt directional changes, which may significantly hinder boundary accuracy, especially in low-contrast regions. Therefore, we proposed a Dynamic Multi-Scale Fusion Network (DMSFNet) for complex street views, which combines the Group Fusion Block (GFB) and the Multi-Branch Feature Fusion (MBFF) modules with the YOLOv1129. Specifically, to accurately extract multi-scale features of road damages with significant aspect ratio variations in complex street views, GFB designs a novel multi-scale hybrid convolution architecture to reorganize and group different fine-grained features generated by different convolutional kernels sizes. In addition, to capture complex crack features more flexibly, we propose MBFF with three branches, one of which introduces Dynamic Snake Convolution (DSConv)30 to dynamically adjust shape and position of the convolution kernel to accurately capture the morphological features of slender, curved cracks. The other two branches use residual connections and cross-stage feature fusion to establish hierarchical representation and balance holistic structural comprehension and detailed texture preservation. As shown in Fig. 2, the experimental results demonstrate that our DMSFNet can effectively reduce the background interference in complex street views and improve the ability of detection for multi-scale road damage. To sum up, the main contributions of this paper are described as follows: To address the challenge of detecting road damages with extreme scale variations, we propose a Group Fusion Block (GFB) with multi-scale hybrid convolution architecture, which utilizes multiple convolutional layers with different kernel sizes in parallel and adaptively adjusts the receptive field according to the scale of the object. Therefore, the features from different scales can be preserved and refined, and the spatial context of damages, from fine cracks to large potholes, can also be captured accurately. To address the challenge of detecting road damages with extreme scale variations, we propose a Group Fusion Block (GFB) with multi-scale hybrid convolution architecture, which utilizes multiple convolutional layers with different kernel sizes in parallel and adaptively adjusts the receptive field according to the scale of the object. Therefore, the features from different scales can be preserved and refined, and the spatial context of damages, from fine cracks to large potholes, can also be captured accurately. To flexibly extract the features of complex cracks, we design a Multi-Branch Feature Fusion (MBFF) block with three branches to enhance the ability of capturing information of cracks with elongated, curved patterns. To be specific, MBFF consists of three specialized branches, i.e., Contextual Aware Branch (CAB), Geometric Tracking Branch (GTB) and Identity Branch (IB). CAB replaces the standard Bottleneck structure with the Inverted Residual Block (IRBlock) to project features into a higher-dimensional space, which can enhance the representation capability for diverse and complex street-view damage. GTB combines Depth-wise Separable Convolution (DWConv) and Dynamic Snake Convolution (DSConv) in a two-stage refinement process to accurately capture fine, curved cracks by adaptively aligning receptive fields with learned continuous contours, enhancing the connectivity of narrow fractures while suppressing background noise. IB is mainly used to avoid possible loss or distortion of information in the other two branches. Therefore, MBFF can ensure that the ability to characterize complex crack features is enhanced while improving the robustness of the model in complex scenarios. To flexibly extract the features of complex cracks, we design a Multi-Branch Feature Fusion (MBFF) block with three branches to enhance the ability of capturing information of cracks with elongated, curved patterns. To be specific, MBFF consists of three specialized branches, i.e., Contextual Aware Branch (CAB), Geometric Tracking Branch (GTB) and Identity Branch (IB). CAB replaces the standard Bottleneck structure with the Inverted Residual Block (IRBlock) to project features into a higher-dimensional space, which can enhance the representation capability for diverse and complex street-view damage. GTB combines Depth-wise Separable Convolution (DWConv) and Dynamic Snake Convolution (DSConv) in a two-stage refinement process to accurately capture fine, curved cracks by adaptively aligning receptive fields with learned continuous contours, enhancing the connectivity of narrow fractures while suppressing background noise. IB is mainly used to avoid possible loss or distortion of information in the other two branches. Therefore, MBFF can ensure that the ability to characterize complex crack features is enhanced while improving the robustness of the model in complex scenarios. We propose a road damage detection model based on the YOLOv11 model called Dynamic Multi-Scale Fusion Network (DMSFNet), which can enhance the ability to extract the features of large-scale and irregular cracks, improving the accuracy of road damage detection. The SOTA performance of DMSFNet on three datasets is verified by extensive experiments. We propose a road damage detection model based on the YOLOv11 model called Dynamic Multi-Scale Fusion Network (DMSFNet), which can enhance the ability to extract the features of large-scale and irregular cracks, improving the accuracy of road damage detection. The SOTA performance of DMSFNet on three datasets is verified by extensive experiments. Comparison of detection mAP@0.5, network parameters and FLOPs of various advanced models on dataset SVRDD. The size of the circle indicates the model parameters. Note that our DMSFNet achieves higher accuracy. Related work Road damage detection and recognition methods Supervised based RDD methods Supervised deep learning methods for damage detection can be divided into two types: pixel-level damage detection and object-level detection. The first one involves damage segmentation, where each pixel in the image is classified as a damage or non-damage31. improved the accuracy and recall of crack detection by increasing the depth of CNN to enable the model to learn more complex features32 presented an improved U-Net structure combined with residual connections and attention gating to enhance deep feature learning and improve the accuracy of crack segmentation33 used capsule network to better capture the direction and shape information of damage, and the siamese network can learn the characteristics of similar diseases and enhance the recognition ability of the model for the same diseases in different video frames34 proposed a pavement defect segmentation with relevance-aware and cross reasoning mechanisms to segment defects merely using non-defective images for training. Object-level detection can be divided into one-stage based and two-stage based methods. For the two-stage based method, the first stage usually generates region proposals, which typically use selective search or Region Proposal Networks (RPNs) to identify regions that may contain targets. The second stage aims at classifying and refining these region proposals by establishing bounding boxes. A typical two-stage network is Faster-RCNN21, which employs RPN to quickly generate regional recommendations and share convolution features with subsequent target detection networks, greatly improving the speed and efficiency of target detection35 employed Faster-RCNN to identify the candidate crack regions, followed by a Bayesian integration algorithm to filter out false positives. Similarly36, adopted Fast-RCNN to improve the efficiency of feature extraction by the region of interest pooling mechanism. Although the above two-stage based methods demonstrate effectiveness in damage detection, their computational inefficiency limits real-time applicability. Different from the two-stage based methods, which requires first selecting the target area, the one-stage detector predicts the bounding box and category probabilities of targets directly on the image. Therefore, one-stage networks such as You Only Look Once (YOLO)23 have been used to improve perception speed significantly37 enhanced the feature extraction capability of the model by adjusting the network structure of YOLO and using an improved loss function to improve the detection accuracy of damaged areas38 presented a method for automated pothole detection with different YOLO models, such as YOLOv4, YOLOv4-tiny, and YOLOv526 introduced a novel approach for road defect detection using a modified version of the YOLOv8 algorithm named BL-YOLOv8, which used BiFPN, SimSPPF and LSK-attention to improve detection accuracy and reduce computational complexity39 adjusted the structure of the YOLOv5 network to make it more suitable for detecting pavement damage features, improving detection accuracy and speed. This method was applied to evaluate the pavement condition of asphalt highways in Liaoning Province, China, and achieved excellent performance. Unsupervised based RDD methods To reduce the reliance on supervised data, some unsupervised methods based on auto encoders40 have been proposed for road damage detection41,42. These methods train models on normal road images and detect damage by checking whether the reconstruction error of an input sample exceeds a threshold. Nonetheless, due to the insufficient supervisory information and similarities among road damage types, such methods may produce erroneous results due to feature confusion. To address this limitation43, proposed an adversarial network-based road damage detection method using Adversarial Image-to-Frequency Transform (AIFT) for road defect detection. To improve the performance of road defect segmentation, Yu et al. proposed a multi-source domain adaptation method44, which generates multi-source pseudo labels as supervision signals for unlabeled target domains by integrating the migration information of multiple labeled source domain models, and combines outlier relocalization to achieve inter-domain alignment, significantly improving the robustness45 proposed Domain Adaptive Road Damage Detection (DA-RDD), which realizes domain-invariant representation learning for cross-domain crack detection through image-level adversarial alignment and instance-level progressive alignment to improve the segmentation performance. Semi-supervised based methods Although the unsupervised methods do not rely on labeled data, their performance is often worse than supervised methods43–45. Therefore, some semi-supervised methods have been proposed46–48. References46 and47 used adversarial learning for semi-supervised road damage detection which employed adversarial learning to minimize domain shifts between a large-scale labelled dataset and a dataset with few labelled samples to reduce the differences between different data domains48 proposed a Pavement Anomaly Detection Network (PAD Net) based on generative adversarial networks, which adopts a semi-supervised framework of dual generators and multi-style discriminators, retains background features and corrects abnormal areas through multi-scale image relationship modeling, and combines end-to-end mask channels to achieve pixel-level precise positioning. Pixel-level segmentation based RDD methods While bounding-box-based detection enables efficient localization of damage instances, pixel-level semantic segmentation provides a more fine-grained alternative by assigning class labels to each pixel. Such dense predictions are particularly valuable for accurately quantifying damage severity, including crack length, width, and area, which are critical for maintenance prioritization and resource allocation. Early studies adapted general-purpose segmentation architectures, such as Fully Convolutional Networks (FCN)49 and U-Net14, to pavement distress analysis. Although these methods achieved reasonable performance under controlled conditions, they often struggled in real-world scenarios due to severe foreground-background class imbalance, complex textures, and the thin, discontinuous nature of crack structures. To address these challenges, more specialized architectures have been proposed. For instance, Choi et al.50 developed SDDNet, a real-time crack segmentation network composed of standard convolutions, densely connected separable convolution modules, a modified atrous spatial pyramid pooling module, and a decoder module, which effectively balances computational efficiency with segmentation accuracy. Kang et al.51 introduced a multi-stage hybrid architecture for crack detection and quantification, synergizing deep learning with traditional image processing. The pipeline initiates with a Faster R-CNN for region proposal, proceeds to a modified Tubularity Flow Field (TuFF) for fine-grained pixel segmentation, and culminates in a tailored Distance Transform Method (DTM) for geometric characterization of cracks. This hierarchical integration substantially improves model robustness and cross-domain generalization under diverse background scenarios. Subsequently, Kang et al.52 introduced a novel Semantic Transformer Representation Network (STRNet) for real-time pixel-level crack segmentation. STRNet employs a squeeze-and-excitation encoder, a multi-head attention decoder, and a focal-Tversky loss to achieve a robust balance between accuracy and speed. More recently, multimodal approaches have been explored to overcome the limitations of single-modality inputs. Hussain et al.53 proposed a cross-fusion framework that integrates RGB and depth information, enabling more robust crack segmentation and quantification under challenging conditions such as low contrast. Despite these advances, pixel-level segmentation methods still face several practical limitations. First, dense pixel-wise annotation is labor-intensive and costly compared to bounding-box labeling. Second, segmentation models typically incur higher computational overhead, which limits their deployment on resource-constrained platforms. Third, for large-scale road inspection tasks that prioritize rapid localization over precise measurement, such fine-grained predictions may introduce unnecessary computational burden. Therefore, detection-based and segmentation-based methods can be viewed as complementary paradigms. The former is more suitable for large-scale, real-time screening, while the latter excels in detailed damage analysis. In this work, we focus on the detection paradigm and propose DMSFNet to achieve robust and efficient road damage detection under complex conditions. Dynamic snake convolution Conventional convolution operations typically rely on fixed geometric structures, which may not perform well when processing data with irregular shapes or dynamic changes. To solve this problem, Dynamic Snake Convolution (DSConv) was designed30, which mimics the movements of snake to better capture local features and global structures by introducing deformable convolution kernels and adaptive path planning mechanisms. Specifically, the convolution kernels of DSConv can deform according to the local features of the input data. This deformation is achieved through learnable offset parameters, enabling the kernel to better adapt to the geometric structure of images. In addition, DSConv incorporates an attention-based path-planning strategy, allowing the convolution kernel to slide along an optimal path in the data space. Since the path-planning mechanism considers both local features and global contextual information, the accuracy and robustness of feature extraction can also be improved. Proposed method Road damage exhibits the characteristics of large scale variation and large aspect ratio. Besides, the complex street views containing a lot of redundant information increase difficulty in detecting road damage. To address these issues, we propose a Dynamic Multi-Scale Fusion Network (DMSFNet) for complex street views, which effectively improves the multi-scale feature extraction capability and the detection accuracy of slender cracks. In this section, we elaborate on our DMSFNet in detail. Overall pipeline As shown in Fig 3, the DMSFNet proposed in this paper is specifically designed based on the YOLOv11 algorithm model. The structure of DMSFNet consists of three parts: backbone, neck, and head. In the original YOLOv11 network, the backbone is used for extracting features from the input image. The neck is employed for feature fusion and enhancement by integrating feature information to provide more useful semantic information for the detection head. The head layer is designed to detect the objects and output results, which contains a series of convolutional layers that are used to predict the class of the target, the bounding box coordinates, and the confidence level. In this paper, we propose two structured modules: Group Fusion Block (GFB) module and Multi-Branch Feature Fusion (MBFF) module. To be specific, we replace the convolutional module in the original second layer with GFB. MBFF module is used to replace C3k2 in the original backbone. In the feature fusion stage, on the basis of using the bidirectional multi-scale feature pyramid fusion, we simultaneously employ GFB and MBFF to achieve feature interaction and enhancement, thereby capturing delicate features more flexibly. The overall structure of DMSFNet, which consists of backbone, neck and head. Group fusion block (GFB) is used to improve the receptive field of the model. Multi-branch feature fusion (MBFF) modules are used to generate more diverse and rich feature information, which significantly enhanced the network’s semantic understanding of damage features. Group fusion block To capture the damages with extreme scale variation and elongated shapes, we propose a Group Fusion Block (GFB) with multi-scale hybrid convolution architecture, which integrates grouping and aggregation mechanisms to explicitly model discrete scale-specific features. Unlike standard Parallel modules, such as Atrous Spatial Pyramid Pooling(ASPP)54 and Receptive Field Block (RFB)55, that merely concatenate multi-scale features, the GFB introduces a grouping and aggregation mechanism after parallel depth-wise convolutions. This design is theoretically based on the principle of preventing feature homogenization, which explicitly structures the heterogeneous multi-scale features into groups, forcing the network to learn complementary representations. Then, a subsequent lightweight group-wise convolution is used to facilitate cross-group information exchange. The structure of GFB is shown in Fig. 4(a). (a) The structure of our GFB has depth-wise separable convolutions and ordinary convolutions. We use a multi-scale hybrid convolution architecture that is proposed to make cross-scale detection more accurate. (b) The structure of our MBFF, which consists of three branches for capturing local /global information. The first branch exhibits high sensitivity to targets of varying dimensions. The second branch can accurately extract the local features of cracks. The third branch uses 1×1 bypass convolution to improve the network’s ability to recognize complex patterns and the efficiency of information flow. Instead of using standard convolutions with large kernels which are computationally expensive, we first employ parallel 3×3 and 5×5 depth-wise separable convolutions to balance receptive field size and computational efficiency. The 5×5 branch provides a larger receptive field to capture context and long-range dependencies crucial for identifying large damages, while the 3×3 branch focuses on finer local textures. The outputs from these parallel branches are concatenated on channel dimension to enrich the feature set. Then, we structure these heterogeneous features into several parallel sub-spaces (groups) through the grouping operation. This mechanism compels the network to learn distributed and complementary representations, thereby preventing feature homogenization and enhancing overall representational capacity. However, simply grouping features would lead to isolated information flows. To overcome this, we apply 1×1 depth-wise separable convolutions to each group. This design reduces dimensionality within each group to improve efficiency. More importantly, it enables the effective and lightweight fusion of information across all channels within a group, refining the features. Finally, we project the aggregated features through a 1×1 convolution and fuse them with the shortcut path, forming a residual connection that stabilizes training and improves generalization. The learning process of GFB is: where is the shallow feature of the first convolution through the network. and denote a 3×3 convolution and a 1×1 convolution, respectively. , and represent 3×3 depth-wise separable convolution, 5×5 depth-wise separable convolution and 1×1 depth-wise separable convolution56, respectively. is the output feature of 3×3 depth-wise separable convolution. is the output feature of 5×5 depth-wise separable convolution. is the concatenation of two feature maps along the channel dimension. is a grouping operation. is the grouped features. N=4 represents the number of groups. is the output features of GFB. Multi-branch feature fusion block To accurately identify various cracks on the road surface of complex street scenes, such as slender cracks to large-area potholes and narrow cracks, we propose a novel Multi-Branch Feature Fusion (MBFF) Block and the architecture of MBFF is detailed in Fig. 4(b), which consists of three specialized branches, i.e.,Contextual Aware Branch (CAB), Geometric Tracking Branch (GTB) and Identity Branch (IB). Next, we will introduce these three branches respectively. Contextual aware branch (CAB) To achieve a comprehensive understanding of damage extent and category, which requires rich contextual information, we adopt a branch based on the C2f structure. Note that the standard Bottleneck structure in C2f performs non-linear transformations in a compressed, low-dimensional space, which will be a limitation for capturing the highly diverse and complex features of street-view damages. Therefore, we propose to replace the standard Bottleneck with the Inverted Residual Block (IRBlock) to project features into a higher-dimensional space, providing a richer and more disentangled basis for the subsequent depth-wise convolution and enhancing the context awareness. Furthermore, to preserve critical feature details, a linear bottleneck (a linear layer after ReLU6) is used to prevent the non-linearity from corrupting information in the low-dimensional output space. After obtaining the feature from previous stage, the process of the CAB can be expressed as follows: where is the input feature. and are the split features. represent the output features of IRBlock. represents Inverted Residual Block. represents the output of the CAB. Geometric tracking branch (GTB) This branch combines the Depth-wise separable Convolution (DWConv) and Dynamic Snake Convolution in a two-stage refinement process to enhance the ability of geometric tracking. The first stage utilizes a DWConv, which consists of a depth-wise separate convolution for efficient extraction of spatial features and a point-wise convolution for channel compression and information integration. This design can enhance the representation of the extracted features and balance computational efficiency. Then, the obtained features can be delivered to the second stage, i.e., DSConv layer, which constrains convolutional sampling points along learnable continuous curve trajectories and enables precise tracking of the bending direction. By adaptively aligning its receptive field with the learned crack trajectories, DSConv not only enhances the connectivity of narrow fractures but also effectively suppresses irrelevant background noise. Specifically, the learning process of the GTB can be formulated as follows: where denotes a 3×3 DSConv and denotes a 3×3 DWConv. is the output features of Geometric Tracking Branch. Identity branch (IB) To ensure the stability of gradient flow and preserve the loss of fine-grained details in the other two branches, we incorporate a simple convolutional bypass to guarantee that critical low-level information can be directly propagated to the fusion point. The learning process of the IB can be formulated as where is the output features of Identity Branch. After the features of the three branches are obtained, we concatenate features , and on the channel dimension and can obtain the fused feature . Note that some channels may contain key information that is highly relevant to the target (such as cracks), while other channels may contain more irrelevant background noise. Thus, we use a lightweight ECA Module57 to adaptively weight the channels, so that the important information can be amplified, and redundant information can be suppressed. The ECA can be formulated as: where C is the number of channels of the input feature. and b are hyperparameters. represents a one-dimensional convolution operation with a kernel size of k. γ =2 and b=1, following the original ECA design. is global average pooling. is sigmoid activation function. is the output features of the ECA. The final features can be obtained by . Experimental results and analysis Experimental setting Datasets description In this paper, three road damage datasets are employed for evaluation, including SVRDD58, RDD202058 and USRDD58. The description of the above datasets is as follows: SVRDD58: SVRDD is the first public dataset based on street view images for road damage detection that contains 8000 street view images from Beijing City. The resolution of road images is 1024 × 1024. The categories of road damage addressed include six damage categories and one confusing non-concrete pavement, namely, Longitudinal Crack(LC), Transverse Crack(TC), Pothole(P), Alligator Crack(AC), Longitudinal Patch(LP), Transverse Patch(TP), and Manhole Cover(MC). For the SVRDD dataset, we first split the 8000 images into 7000 development images and 1000 testing images. From the 7000 development images, 1400 images are randomly selected as the validation set, while the remaining 5600 images are used as the original training set. Data augmentation strategies, including Mosaic, random horizontal flipping, and random rotation, are applied only to the training set. After augmentation, the effective number of training images is expanded to 7000. The validation and testing sets are not augmented and are used only for hyperparameter tuning and final evaluation, respectively. RDD202059: RDD2020 dataset consists of 26620 road damage images from Japan, India, and the Czech Republic. The resolution of road images in Japan and the Czech Republic is 600 × 600, and the resolution of road images in India is 720 × 720. There are four types of road damage: longitudinal cracks (D00), transverse cracks (D10), alligator cracks (D20) and potholes (D40). In this paper, we partitioned the dataset into 18500 images for training, 3700 for validation, and 2541 for testing. USRDD60: The USRDD is the United States road damage dataset in the 2022 crowdsensing-based road damage detection challenge (CRDDC). All images are 640 × 640 Google Street View images. Similarly, it contains four types of damages: longitudinal cracks (D00), transverse cracks (D10), alligator cracks (D20), and potholes (D40). Following the same rigorous protocol applied to the SVRDD dataset, we partitioned the available data into 4000 images for training, 805 images for validation, and 1200 images for testing. Training setting The experimental environment in our work consists of Windows 11 operating system with one NVIDIA RTX 4090 GPU, and the version of CUDA is 11.8. For fairness in data comparison, we kept the hyperparameter settings in our experiments consistent with those of the YOLOv11 baseline unless otherwise specified. All images were resized to 640 × 640 during both training and testing phases. We trained our models using Stochastic Gradient Descent (SGD) with an initial learning rate of 0.01, a momentum of 0.937, and a weight decay of 0.0005. The Batch Size was set to 32, and models were trained for 300 epochs. The data augmentation strategies included Mosaic augmentation, horizontal random flip, and random rotation, etc., which are consistent with the baseline for a fair comparison. Implementation details To ensure full reproducibility, we provide the following specifications that complement the training settings described above. GFB configuration. In the Group Fusion Block (GFB), the number of feature groups N is set to 4. This value was determined through preliminary grid search over , with N=4 offering the optimal balance between representational diversity and computational efficiency. The depth-wise separable convolutions employ a stride of 1 and padding consistent with kernel size to preserve spatial resolution. MBFF configuration. Within the Multi-Branch Feature Fusion (MBFF) module, the Contextual Aware Branch (CAB) employs n=2 Inverted Residual Blocks (IRBlocks). Each IRBlock uses an expansion factor of 6 and a stride of 1. The Geometric Tracking Branch (GTB) utilizes a 3× 3 Dynamic Snake Convolution (DSConv) with 9 sampling points constrained along a learnable continuous curve trajectory. The Efficient Channel Attention (ECA) module is configured with hyperparameters γ =2 and b=1, following the original implementation in57. Random seed and initialization. A fixed random seed of 42 is used across all experiments for NumPy, PyTorch, and CUDA operations to ensure deterministic behavior. All models, including both the backbone and newly introduced modules (GFB and MBFF), are trained from scratch without any pre-trained weights. Network parameters are initialized using Kaiming uniform initialization for convolutional layers and batch normalization layers are initialized with unit scale and zero shift. FPS evaluation protocol. Inference speed (Frames Per Second, FPS) is measured on a single NVIDIA RTX 4090 GPU with CUDA 11.8. Following standard practice, we report the average latency over 1000 forward passes with a batch size of 1 after a 100-iteration warm-up phase. Input images are resized to 640× 640. Post-processing steps, including Non-Maximum Suppression (NMS), are included in the latency measurement to reflect end-to-end deployment conditions. Data augmentation details. As noted in the training setting, we employ Mosaic augmentation (probability 1.0), random horizontal flipping (probability 0.5), random rotation (), and HSV color space augmentation (hue ± 0.015, saturation ± 0.7, value ± 0.4). All augmentations are applied during training. Evaluation metrics To evaluate the performance of our proposed DMSFNet, score, Precision, Recall, mAP@0.5 , Params and FPS are employed. score is the harmonic average of the precision and recall, which can measure the overall performance of the model and can be defined as: where P stands for Precision and R stands for Recall, which are obtained by Eq. 7 and Equation 8 respectively. where TP, FP and FN indicate the number of true positives, false positives and false negatives, respectively. AP is the average precision obtained by integrating the Precision-Recall curve for each class and can be calculated as where P(R) denotes the Precision-Recall curve. mAP is the mean AP values of all categories and can be calculated as where M is the number of classes in the object detection datasets, M= 7 in the SVRDD dataset, M= 4 in the RDD2020 dataset and USRDD dataset. When the threshold of Intersection over Union (IOU) for dividing positive and negative samples is set to 0.5, the mean average precision is expressed as mAP@0.5. Params is the sum of the parameters of the model. FPS is the number of frames processed in one second. Compared methods To further verify the superiority of the proposed method, we compare our DMSFNet with several mainstream target detection algorithms, including RT-DETR61, RT-DETRv262, DEIM64 and YOLO series as the experimental algorithms. For fair comparison, we retrain all compared methods on the three used datasets. Results on the SVRDD dataset We evaluate each method on the SVRDD dataset, with numerical results reported in Table 1. Observations are as follows: 1) Our method achieves the highest overall mAP@0.5, outperforming both YOLOv11 and Mamba-YOLO. Under the stricter mAP@0.5:0.95 metric, DMSFNet also attains the best performance (51.1%), surpassing YOLOv11 (48.4%)–further validating the precision of our bounding box regression. 2) For Longitudinal Cracks (LC) and Transverse Cracks (TC), DMSFNet improves mAP@0.5 by 3.4% and 9.1% over YOLOv11, respectively, demonstrating significant gains in detecting elongated cracks. Despite the small size and visual ambiguity of potholes and manhole covers, DMSFNet achieves strong results with mAP@0.5 of 76.9% for Potholes (P) and 77.0% for Manhole Covers (MC), further confirming its robustness across diverse damage types. Fig. 5 presents visual comparisons. The first two rows highlight detections of slender cracks and damages with large scale variations, where DMSFNet identifies more road damages with higher confidence and localization accuracy. Table 1The numerical detection results in terms of mAP@0.5 and mAP@0.5:0.95 on the SVRDD dataset.MethodFLOPs(G)Param(M)OverallOverallmAP@0.5% on some categoriesmAP@0.5%mAP@0.5:0.95%LCTCPACLPTPMCRT-DETR 6160.020.065.441.568.266.563.167.864.965.761.6RT-DETRv2 6260.020.066.142.269.067.363.868.565.666.462.1RT-DETRv3 6360.020.066.942.970.168.264.571.366.467.160.7DEIM 647.04.065.742.868.566.863.368.165.266.061.8DEIMv2 656.83.666.243.569.267.564.068.865.566.162.3YOLOv5 664.51.960.336.565.763.257.562.956.257.459.3YOLOv88.73.262.939.266.464.860.163.461.060.964.0YOLOv10 676.72.366.343.065.367.664.768.968.466.862.4YOLOv11 296.52.674.648.477.577.770.678.368.077.972.0YOLOv12 686.52.671.446.072.971.571.074.869.970.768.8YOLOv13 696.42.571.646.676.873.165.272.271.772.669.8YOLO-MS 7017.44.569.044.371.870.166.671.468.569.365.0Mamba-YOLO 7113.25.873.347.576.174.470.975.772.873.669.3FBRT-YOLO 726.70.972.046.874.873.169.676.471.570.368.0Gold-YOLO 7312.15.672.547.075.373.670.174.972.072.868.5DMSFNet (ours)11.73.380.451.180.986.876.984.477.878.977.0 The numerical detection results in terms of mAP@0.5 and mAP@0.5:0.95 on the SVRDD dataset. The visualization results on the SVRDD: (a) Initial image, (b) RT-DETR, (c) DEIM, (d) YOLOv11, (e) FBRT-YOLO, (f) DMSFNet (ours). Results on the RDD2020 dataset We evaluate all methods on the RDD2020 dataset, with the quantitative results presented in Table 2. Key observations are as follows: 1) Our method achieves the highest overall mAP@0.5, whereas RT-DETR and RT-DETRv2 perform the worst. It also ranks first in the stricter mAP@0.5:0.95 metric (29.1%), surpassing YOLOv11 (23.6%) by a clear margin. This confirms the robustness of our bounding box localization. 2) Compared to YOLOv11, DMSFNet improves mAP@0.5 by 7.2%, 9.3%, and 5.2% on the D00, D10, and D20 categories, respectively, demonstrating clear gains in detecting elongated cracks. Fig. 6 provides visual comparisons. The first two rows highlight detections of slender cracks and damages with large-scale variations under poor lighting conditions. DMSFNet consistently detects more road damages with higher confidence and accuracy. Notably, the RDD2020 dataset contains numerous low-light images, which partly explains the relatively lower overall accuracy–highlighting a valuable direction for future research in low-light road damage detection. Table 2The numerical detection results in terms of mAP@0.5 and mAP@0.5:0.95 on the RDD2020 dataset.MethodFLOPs(G)Param(M)OverallOverallmAP@0.5% on some categoriesmAP@0.5%mAP@0.5:0.95%D00D10D20D40RT-DETR6160.020.049.622.550.245.855.147.3RT-DETRv26260.020.048.721.849.344.954.246.4RT-DETRv36360.020.050.123.050.746.355.647.8DEIM647.04.051.323.851.947.456.849.1DEIMv2656.83.652.524.553.148.558.050.4YOLOv5664.51.950.423.246.541.360.553.4YOLOv88.73.254.325.546.348.366.955.6YOLOv10676.72.352.224.252.848.257.750.1YOLOv11296.52.652.023.649.044.562.552.1YOLOv12686.52.653.825.054.449.759.351.8YOLOv13696.42.552.924.353.548.858.450.9YOLO-MS7017.44.551.123.451.747.256.649.0Mamba-YOLO7113.25.855.426.256.051.160.953.6FBRT-YOLO726.70.956.026.856.151.762.054.2Gold-YOLO7312.15.654.225.454.850.059.752.4DMSFNet (ours)11.73.359.329.156.253.867.759.6 The numerical detection results in terms of mAP@0.5 and mAP@0.5:0.95 on the RDD2020 dataset. The visualization results on the RDD2020: (a) Initial image, (b) RT-DETR, (c) DEIM, (d) YOLOv11, (e) FBRT-YOLO, (f) DMSFNet (ours). Results on the USRDD dataset We also evaluate our method on the USRDD dataset. The quantitative detection results and visual detection results are described in Table 3 and Fig. 7, respectively. It can be observed that our proposed DMSFNet performs consistently well on other datasets: 1) The results of mAP@0.5 on each category are still higher than the other compared methods, and it also leads under the mAP@0.5:0.95 metric, indicating improved bounding box localization precision. 2) As shown in Fig. 7, we see that DMSFNet can also detect slender cracks and large variations well in damage scales, despite the interference from tree projections. For a more fine-grained comparison, Fig. 8 visualizes the per-category mAP@0.5 via radar charts on these three datasets. DMSFNet consistently demonstrates broader coverage across these charts, underscoring its balanced performance. Table 3The numerical detection results in terms of mAP@0.5 and mAP@0.5:0.95 on the USRDD dataset.MethodFLOPs(G)Param(M)OverallOverallmAP@0.5% on some categoriesmAP@0.5%mAP@0.5:0.95%D00D10D20D40RT-DETR6160.020.058.832.569.563.370.132.3RT-DETRv26260.020.060.335.171.264.871.833.4RT-DETRv36360.020.061.537.172.666.273.234.0DEIM647.04.060.036.370.864.571.533.2DEIMv2656.83.660.437.570.364.471.735.1YOLOv566 4.5 1.957.732.868.362.268.831.5YOLOv88.73.258.734.069.563.270.032.1YOLOv10676.72.361.137.872.165.772.733.9YOLOv11296.52.659.436.570.263.970.832.7YOLOv12686.52.660.136.871.064.671.633.2YOLOv13696.42.561.638.272.766.373.334.1YOLO-MS7017.44.559.035.069.863.670.432.2Mamba-YOLO7113.25.859.037.571.865.472.433.6FBRT-YOLO726.7 0.9 60.837.673.266.773.734.4Gold-YOLO7312.15.660.937.972.065.172.633.8DMSFNet (ours)11.73.3 66.4 39.7 79.3 74.5 76.2 35.4 The numerical detection results in terms of mAP@0.5 and mAP@0.5:0.95 on the USRDD dataset. The visualization results on the USRDD: (a) Initial image, (b) RT-DETR, (c) DEIM, (d) YOLOv11, (e) FBRT-YOLO, (f) DMSFNet (ours). Radar chart visualization of mAP@0.5 across categories on three datasets. From left to right are the SVRDD, RDD2020, and USRDD datasets. To provide a clearer illustration, we selected 6 representative methods from the original 16. Comparison with other RDD methods In this section, we provide a reference-based comparison between DMSFNet and several existing road damage detection methods on the RDD2020 dataset. The results of the compared methods in sTable 4 are collected from 28 and their respective original papers, rather than being reproduced under a fully unified experimental protocol. Table 4Comparison with other RDD methods on RDD2020 dataset.MethodF1%Overall mAP@0.5%ES-RDD7456.553.6MDL-RDDC7536.9-AL-RDD7661.4-PDC7758.1-RDD-FRCNN7848.738.6RDD-YOLOv57956.653.6YOLO-LRDD8058.757.6YOLOX-RDD2860.058.1DMSFNet (ours)60.659.3 Comparison with other RDD methods on RDD2020 dataset. As shown in Table 4, DMSFNet achieves an overall mAP@0.5 of 59.3%, which is higher than the reported results of the compared methods. In terms of F1 score, DMSFNet obtains 60.6%, showing competitive performance compared with the best reported F1 score. These results suggest that the proposed GFB and MBFF modules are effective for road damage detection; however, a more rigorous comparison under a unified training and evaluation protocol would be valuable for future work. Notably, compared to general-purpose detectors (Tables 1–3), these task-specific methods employ various domain-specific optimizations such as customized data augmentation, multi-scale training, and ensemble strategies. Despite these specialized designs, DMSFNet still achieves promising and highly competitive performance, demonstrating the effectiveness and generalizability of our proposed GFB and MBFF modules for road damage detection in complex scenarios. Ablation study Module ablation experiment To validate the effectiveness of the core components in DMSFNet–GFB and MBFF–we conduct three ablation experiments on the SVRDD dataset by removing each module. Results are shown in Table 5. Both modules contribute to overall performance improvement: GFB increases mAP@0.5 by 3.8% and F1-score by 4.2%, while MBFF improves mAP@0.5 by 5.3% and Recall by 4.6%.Fig. 9 visualizes feature maps from the neck network. Compared to YOLOv11, DMSFNet extracts more salient road damage features. These results demonstrate that GFB and MBFF effectively enhance feature representation and detection accuracy. Table 5Module ablation experiments on the SVRDD dataset.MethodPrecision%Recall%F1%mAP@0.5%Params(M)FPSYOLOv1173.167.670.274.62.681YOLOv11-GFB77.971.374.478.42.768YOLOv11-MBFF79.772.275.879.93.359YOLOv11-GFB-MBFF(ours)80.772.376.280.43.349 Module ablation experiments on the SVRDD dataset. The visualization results of the feature map output by neck network on SVRDD dataset: (a) Input image from SVRDD dataset, (b) feature map of YOLOv11 baseline, (c) feature map of DMSFNet. Branch ablation study We further ablate the three branches of MBFF, i.e., CAB, GTB and IB with results in Table 6. Removing CAB causes the most significant drop, especially in Pothole mAP@0.5 (from 76.9% to 60.6%), highlighting its importance in modeling global context. Ablating GTB reduces Crack mAP@0.5 from 83.9% to 75.8%, confirming DSconv’s necessity for capturing slender crack patterns. Removing IB also degrades Overall and Crack mAP@0.5 (from 80.4% to 78.9% and 83.9% to 80.3%, respectively), demonstrating its positive contribution to feature extraction. Table 6Branch ablation experiments on the SVRDD dataset.MethodOverall mAP@0.5%Crack mAP@0.5%Pothole mAP@0.5%Params(M)FPSw/o CAB71.472.960.63.259w/o GTB74.375.865.23.251w/o IB78.980.373.33.349Our model80.483.976.93.349 Branch ablation experiments on the SVRDD dataset. Convolution ablation experiment To further validate the effectiveness of DSConv within the Geometric Tracking Branch, we conduct a controlled ablation study by replacing DSConv with standard convolution (Conv), depth-wise convolution (DWConv), and Deformable Convolution (DeformConv)81 under the same architectural setting. The results are reported in Table 7. Table 7Convolution ablation experiment on SVRDD dataset.MethodOverall mAP@0.5%Crack mAP@0.5%Params(M)FPSDWConv+Conv77.678.13.353DWConv+DWConv78.879.23.352DWConv+DeformConv80.980.63.345DWConv+DSConv(ours)80.483.93.349 Convolution ablation experiment on SVRDD dataset. As shown, replacing DSConv with standard convolution leads to a clear degradation in both overall detection performance and crack-specific accuracy. In particular, the Crack mAP@0.5 drops from 83.9% to 78.1%, indicating that standard convolution lacks the ability to effectively capture elongated and irregular crack structures. Compared with DWConv, DSConv further improves both overall mAP and crack detection performance, demonstrating that simply reducing computational cost via depth-wise convolution is insufficient for modeling complex geometric patterns. Although DeformConv achieves the highest overall mAP@0.5, its Crack mAP@0.5 is still inferior to that of DSConv. This suggests that while DeformConv enhances general feature adaptability, it is less effective in capturing fine-grained crack morphology. In addition, DeformConv introduces higher computational overhead, resulting in a lower inference speed. In contrast, DSConv achieves the best Crack mAP@0.5 while maintaining competitive overall performance and higher efficiency than DeformConv. These results demonstrate that the structurally constrained adaptive receptive field of DSConv is particularly suitable for modeling slender and curved crack patterns, providing a better trade-off between accuracy and efficiency. Fusion strategy ablation experiment Table 8 reports the impact of different channel attention mechanisms on multi-branch feature fusion. While Coordinate Attention (CA) attains the highest overall mAP@0.5 (81.5%), we adopt Efficient Channel Attention (ECA)57 in the final DMSFNet design as a deliberate accuracy–efficiency trade-off. Specifically, ECA achieves competitive accuracy (80.4% mAP@0.5) while maintaining a higher inference speed (49 FPS vs. 45 FPS for CA), which is more suitable for real-time deployment scenarios. In addition, CA introduces extra spatial coordinate encoding, increasing computational overhead without providing proportional gains for crack detection. In contrast, ECA employs a lightweight 1D convolution to capture local cross-channel interactions, enabling effective feature recalibration with minimal complexity.Although CA yields a marginal improvement of 1.1% in mAP@0.5, this gain is offset by reduced efficiency. Furthermore, within the complete DMSFNet framework, the performance gap is mitigated by the strong feature representation capability of the GFB and MBFF modules. Therefore, ECA is selected as the preferred fusion strategy to achieve a better balance between detection accuracy and computational efficiency. Table 8Fusion strategy ablation experiment on the SVRDD dataset.MethodOverall mAP@0.5%Params(M)FPSConcat+Conv(Baseline)77.23.254Concat+SE+Conv79.53.350Concat+CA+Conv81.53.345Concat+ECA+Conv(ours)80.43.349 Fusion strategy ablation experiment on the SVRDD dataset. Limitation and deployment discussion While the proposed DMSFNet achieves superior accuracy across multiple benchmarks, we acknowledge that its architectural enhancements—specifically the Group Fusion Block (GFB) and Multi-Branch Feature Fusion (MBFF) module—incur additional computational overhead relative to the YOLOv11 baseline. As reported in Table 1, FLOPs increase from 6.5G to 11.7G, accompanied by a reduction in inference speed from 81 FPS to 49 FPS. From the results in Fig. 2 and Tables 1–3, 11.7G FLOPs of our method remain substantially lower than those of competing high-performance detectors, such as transformer-based RT-DETR (60G FLOPs) and the multi-scale YOLO-MS (17.4G FLOPs), while delivering the highest mAP@0.5 among all evaluated methods. Notably, compared to Mamba-YOLO (13.2G FLOPs), DMSFNet attains markedly higher accuracy with reduced computational demand, underscoring the efficiency gains derived from the proposed grouped and multi-branch fusion strategies. Typical vehicular or UAV-based platforms operate at acquisition rates of 10FPS -30 FPS, leaving ample computational headroom (approximately 20 ms per frame) for auxiliary tasks such as sensor fusion or data logging. In addition, the model has a moderate number of parameters (3.3M), which is well within the storage range of edge devices. In summary, we think that the increased computational cost in exchange for improved detection performance on irregular road cracks is worthwhile. Concluding remarks In this paper, we propose a novel road damage detection network, called DMSFNet, to address the challenges of recognizingobjects with large-scale variations and slender, curved cracks in complex urban environment. Based on YOLOv11, weintroduce two key modular innovations to enhance multi-scale and geometric feature representation. First, we proposea Group Fusion Block (GFB), which incorporates a multi-scale hybrid convolution architecture and a grouped featureaggregation strategy to improve cross-scale feature fusion and expand receptive fields. To accurately capture slender andirregular crack structures, we design a Multi-Branch Feature Fusion (MBFF) module with three branches to adaptively alignthe convolutional kernel along curved contours, enabling precise localization of thin cracks and strengthening contextualrepresentation. Comprehensive experiments demonstrate that DMSFNet achieves state-of-the-art performance in detectingmulti-scale damages and geometrically complex cracks. In the future, we will focus on improving road damage detection inlow-light conditions, further reducing computational overhead, and enhancing the robustness of detection algorithms acrossdiverse environmental conditions. Author contributions Hong Zhou and Yiyang Qin wrote the main manuscript text and Validated the method proposed in the article. Jiahuan Ren and Yan Zhang revised the manuscript and analysed the experiment results. All authors reviewed the manuscript. Funding This work is supported by the Jiangsu Provincial Higher Education Natural Science Research Project (Z413B25331) and TrustedFirmware and Intelligent Software Jiangsu Key Laboratory of Higher Education Open Fund Project(JSET-KLTSIS-2025-004). Data availability Publicly available in a repository: RDD2020 are available at the following URL: https://data.mendeley.com/datasets/5ty2wb6gvg/1 SVRDD are available at the following URL: https://zenodo.org/records/10100129 USRDD are available at the following URL: https://doi.org/10.6084/m9.figshare.21431547. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Hong Zhou, Yiyang Qin, Jiahuan Ren, Yan Zhang and Shangbing Gao. References Associated Data Data Availability Statement Publicly available in a repository: RDD2020 are available at the following URL: https://data.mendeley.com/datasets/5ty2wb6gvg/1 SVRDD are available at the following URL: https://zenodo.org/records/10100129 USRDD are available at the following URL: https://doi.org/10.6084/m9.figshare.21431547.