Automated road defect detection has attracted growing attention due to its importance for intelligent transportation and infrastructure maintenance. Traditional techniques, such as edge detection [1], wavelet-based analysis [2], and texture descriptors [3], are highly sensitive to noise, lighting variation, and road texture diversity. In contrast, deep learning-based methods [4,5] offer superior robustness and accuracy. Zhang et al. [6] developed a CNN-based pipeline for crack detection that significantly outperformed handcrafted approaches. Shi et al. [7] proposed a structure forest model to handle the complexity and topological variation of cracks. More recently, Park et al. [8] introduced an adaptive pixel neighborhood segmentation method, which improved detection under noisy backgrounds.

YOLO-based detectors have emerged as a strong baseline for real-time road defect detection. For instance, Jocher et al. [9] released YOLOv5, which offers a good balance between speed and performance. Several studies [10–14] have adapted YOLO variants to road scenarios, incorporating tailored preprocessing or architectural changes to handle challenges such as occlusion, low resolution, and class imbalance. However, small-scale and unevenly distributed defects (e.g., micro-cracks or edge disintegration) remain under-detected due to limited feature expressiveness in conventional backbones.

Attention mechanisms have become integral in enhancing CNNs’ and transformers’ ability to model long-range dependencies and focus on task-relevant features. Channel attention modules, such as SE-Net [15], ECA-Net [16], and the Coordinated Attention (CA) module [17], improve feature channel calibration, enhancing performance across classification and detection tasks. Spatial attention, as used in CBAM [18] and SAM [19], enables focus on key spatial regions, which is particularly useful in defect localization. The combination of spatial and channel attention has also been extended into multi-branch fusion networks and deformable attention [20]. As shown in Table 1, MGD-YOLO achieves the highest mAP with a lightweight architecture compared to recent methods.

In the context of road defect detection, attention modules help reduce background interference and emphasize texture-disruptive regions. For instance, Liu et al. [29] proposed a graph-based attention fusion method to integrate multiple defect cues. Wang et al. [30] used dual attention paths to handle noise and occlusion. Inspired by these advances, our work incorporates a Multi-Scale Dilated Attention (MSDA) module, which enables the model to simultaneously focus on semantic patterns at various receptive field sizes, effectively enhancing sensitivity to subtle structural anomalies.

Deploying detection models in real-world infrastructure applications often requires low-latency and lightweight architectures. Depthwise Separable Convolutions, first introduced in MobileNet [31,32], have become a standard tool for reducing parameter count and computation, followed by enhancements like inverted residual blocks in MobileNetV2 [33] and re-parameterization in RepVGG [34]. For real-time edge deployment, recent frameworks like YOLOv7-Tiny [22,35,36] and YOLO-NAS [37] attempt to balance accuracy and efficiency through backbone redesign and NAS-based optimization.

On the feature fusion side, models such as FPN [38], PANet [39], and BiFPN [40] improve multi-scale prediction by enhancing the flow of semantic information across network layers. However, simple concatenation or summation can introduce redundancy and reduce spatial precision. To address this, attention-guided fusion modules [41–44] and context-aware decoders [45–47] have been proposed. Our method builds upon this line of work by designing a Visual Global Attention Upsampling (VGAU) module that leverages high-level semantics as global context guidance to refine low-level feature maps during upsampling, thereby enhancing the localization of small or low-contrast defects.

In object detection tasks, high-level feature maps typically encapsulate rich semantic information but suffer from limited spatial resolution, making it challenging to accurately localize fine-grained targets. Conversely, low-level feature maps preserve high-resolution spatial details yet lack semantic abstraction. Bridging this semantic-resolution gap is critical for improving detection performance across varied object scales, particularly in complex scenarios such as road defect detection. Previous studies have attempted to address this challenge through hierarchical feature fusion [38,39], yet simple aggregation often introduces redundant information and fails to resolve semantic inconsistencies between layers.

To overcome these limitations, we incorporate an attention-based strategy into the feature fusion process of YOLOv5 by introducing the Multi-Scale Dilated Attention (MSDA) module. Attention mechanisms have shown significant effectiveness across various domains, including object detection [18], semantic segmentation [48], and natural language processing [49,50], due to their ability to dynamically emphasize task-relevant features while suppressing irrelevant noise. Classic modules such as Squeeze-and-Excitation (SE) [15], CBAM [18], and ECA [16] have demonstrated the benefit of channel-wise and spatial recalibration. However, these approaches often lack flexibility in modeling variable context scales. In contrast, MSDA introduces dilated self-attention across multiple receptive fields, enabling the network to capture both local details and global semantic dependencies in a unified framework.

As shown in Fig. 3, given an input feature map F∈RH×W×C, the MSDA module first splits it into multiple attention heads along the channel dimension. Each head applies a window-based self-attention mechanism with a specific dilation rate r∈{1,2,3} to expand its receptive field. This enables each head to capture context at different spatial scales. For a query position x, the output of head i is computed as: (1)Ai(x)=∑y∈𝒩ri(x)Softmax(ϕq(x)⊤⋅ϕk(y))⋅ϕv(y)

where 𝒩ri(x) denotes the neighborhood region centered at x under dilation ri, and ϕq, ϕk, ϕv are learnable linear projections for queries, keys, and values, respectively. The outputs of all attention heads are then concatenated and passed through a lightweight multi-layer perceptron (MLP) to produce the refined feature map: (2)F′=MLP(Concat(A1,A2,A3))+F

This design enables MSDA to aggregate multi-scale semantic cues while maintaining efficiency through dilated sampling, thus reducing information redundancy without introducing additional heavy computation.

The combination of dilated sampling and dynamic attention aggregation not only enhances the representational capacity at multiple scales but also reduces information redundancy compared to naive multi-branch fusion, leading to improved feature discriminability with lower computational cost.

We integrate MSDA immediately after the C3 module within the YOLOv5 backbone, as depicted in Fig. 4. This placement ensures that enriched multi-scale semantic features are incorporated before the upsampling stage. By doing so, low-level features retain detailed structural information, while high-level features provide contextual guidance, forming a more consistent and discriminative representation for defect detection.

Theoretically, the multi-scale dilated attention promotes a more stable feature learning process by ensuring that both localized anomalies (e.g., small cracks) and broader contextual cues (e.g., surface material transitions) are simultaneously emphasized during forward and backward propagation. This stabilization effect improves model convergence behavior and robustness during training.

Empirically, this configuration allows the model to better focus on subtle texture changes and irregular defect boundaries that are critical in road inspection tasks. Consequently, the MSDA-enhanced feature maps significantly contribute to improving the model’s detection accuracy, particularly in identifying small, scattered, or visually ambiguous road defects.

While convolutional neural networks (CNNs) have achieved remarkable success in object detection due to their hierarchical feature representation and end-to-end trainability [51,52], they often suffer from loss of fine-grained spatial details during deep feature extraction. High-level features, although semantically rich, are typically downsampled and lose precise localization cues, which is particularly detrimental for detecting small or low-contrast objects like road cracks or surface repairs. Conversely, low-level features retain spatial resolution but lack semantic context, making it challenging to distinguish defects from background textures.

To address this issue, various U-shaped architectures [53,54] have explored the integration of decoder paths to restore fine details. However, these designs often involve complex multi-stage decoders and impose high computational overhead, limiting their suitability for real-time applications on resource-constrained platforms.

To enable efficient and context-aware upsampling, we draw inspiration from the Global Attention Upsampling (GAU) module [55,56] and propose an improved variant tailored for road defect detection, termed Visual Global Attention Upsampling (VGAU). Our VGAU module leverages high-level global semantic context to recalibrate low-level feature responses, enhancing localization precision without introducing significant computational complexity. Unlike traditional decoder structures that independently process low-level features, VGAU introduces top-down semantic guidance during upsampling, which improves the discrimination ability of low-level feature activations while stabilizing feature propagation across the network. This facilitates a smoother gradient flow and more robust convergence during training.

As illustrated in Fig. 5, given a high-level feature map Fh∈RH×W×C and a corresponding low-level feature map Fl∈R2H×2W×C′, we first compute a global semantic vector via global average pooling: (3)g=1H⋅W∑i=1H∑j=1WFh(i,j)

This global descriptor g∈RC is then transformed through a lightweight channel-wise excitation function composed of a 1×1 convolution, batch normalization, and a non-linear SiLU activation: (4)w=σ(BN(W1⋅g))

Meanwhile, the low-level feature map Fl is compressed using a 3×3 convolution to reduce channel dimensionality. The attention-guided modulation is then applied by reweighting Fl with w: (5)F~l=w⊙Conv3×3(Fl)

The recalibrated low-level features F~l are then fused with the upsampled high-level features via summation: (6)Fout=Upsample(Fh)+F~l

Theoretically, VGAU enhances training stability by aligning low-level feature distributions with high-level semantic priors, which reduces feature noise and suppresses gradient vanishing phenomena in the decoder pathway. The use of channel-wise attention ensures that the network dynamically emphasizes important semantic clues while filtering irrelevant background patterns, thereby accelerating convergence and improving generalization performance.

As shown in Fig. 6, VGAU is embedded in the upsampling pathway of the MGD-YOLO architecture, where it operates in conjunction with the MSDA and DSC modules. This integration enables the network to preserve both the fine-grained spatial cues and semantic richness required for robust road defect detection.

Compared to the original GAU, our VGAU incorporates two major improvements: (1) the use of the SiLU activation for smoother gradient propagation and non-linearity, and (2) an additional upsampling operation post-fusion to enhance resolution alignment. Overall, VGAU not only improves multi-scale feature fusion efficiency but also enhances model training dynamics by promoting consistent semantic flow across layers, ensuring higher stability and robustness in real-world deployment. These enhancements allow the module to operate efficiently across multi-scale representations, ultimately contributing to improved precision and recall in our detection results.

Traditional convolutional operations, as adopted in the original YOLOv5 architecture, compute correlations between input features and convolution kernels across both spatial and channel dimensions simultaneously. While effective in capturing local patterns, this approach introduces significant computational overhead, especially when dealing with high-dimensional feature maps. The number of parameters and the computational complexity of a standard convolutional layer are given by: (7)Parametersstandard=Dout×K×K×Din (8)FLOPsstandard=H′×W′×Dout×K×K×Dinwhere Din and Dout denote the number of input and output channels, K is the kernel size, and H^′, W^′ are the spatial dimensions of the output feature map.

In the context of real-time road defect detection, such computational demands pose serious limitations, especially for deployment on edge devices with constrained resources. To alleviate this issue and accelerate inference, we replace all standard convolution layers in the network with Depthwise Separable Convolution (DSC) modules, a lightweight alternative initially proposed in MobileNet [31]. As shown in Fig. 7, DSC factorizes the convolution operation into two independent steps: depthwise convolution and pointwise convolution.

1) Depthwise Convolution. This step applies a spatial convolution independently to each input channel. For a kernel size K×K, the number of parameters and computational cost are reduced to: (9)Parametersdepthwise=Din×K×K (10)FLOPsdepthwise=H′×W′×Din×K×K

2) Pointwise Convolution. A 1×1 convolution is applied across channels to linearly combine the outputs from the depthwise step. Its parameter count and computation are: (11)Parameterspointwise=Din×Dout (12)FLOPspointwise=H′×W′×Din×Dout

3) Total Complexity. By combining both operations, the total cost of a DSC layer becomes: (13)ParametersDSC=Din⋅K2+Din⋅Dout (14)FLOPsDSC=H′⋅W′⋅(Din⋅K2+Din⋅Dout)

Compared to standard convolution, this results in an approximate reduction factor of: (15)1Dout+1K2(assuming Din=Dout)

This structural decomposition enables the network to maintain its feature learning capacity while drastically reducing both parameter count and floating-point operations. Such efficiency gains are particularly valuable in road defect detection, where real-time performance and lightweight deployment are crucial.

Moreover, the use of DSC enhances the model’s generalization ability by limiting overfitting from redundant parameterization and encourages efficient representation learning. In our MGD-YOLO architecture, DSC replaces all conventional CBL (Convolution + BatchNorm + LeakyReLU) blocks, further improving inference speed and making the model well-suited for deployment on embedded or mobile platforms for large-scale road condition monitoring.

Although more recent models in the YOLO series—such as YOLOv8 [23], YOLOv9 [57], and YOLOv10 [58]—offer improvements in detection accuracy and architectural novelty, we choose YOLOv5 as the baseline framework for MGD-YOLO primarily due to its maturity, stability, and deployability in real-world scenarios. As our target application emphasizes real-time defect detection on vehicle-mounted edge devices, lightweight design and efficient inference are of paramount importance. YOLOv5 strikes a practical balance between accuracy and computational cost, with a modular architecture that facilitates easy customization and integration of new components such as MSDA, DSC, and VGAU. In contrast, newer versions often increase model complexity and hardware requirements, which may hinder their deployment in resource-constrained environments. Furthermore, YOLOv5 remains a widely accepted baseline in many road defect detection benchmarks, enabling consistent and fair comparisons with prior work. By enhancing YOLOv5 with carefully designed modules, we demonstrate that significant performance gains can be achieved without sacrificing speed or portability, making the model more suitable for intelligent transportation systems and edge computing platforms.

To comprehensively evaluate the effectiveness of the proposed MGD-YOLO framework, we conducted experiments on three publicly available road defect detection datasets: TD-RD, CNRDD, and CRDDC’22. These datasets collectively include various types of road surfaces–such as cement and asphalt–and cover three representative categories of surface anomalies: cracks, repairs, and potholes. All images were uniformly resized to a resolution of 640×640 pixels to ensure consistency during model training and inference. Fig. 8 provides representative samples from these datasets.

Specifically, TD-RD contains 1532 annotated images collected from three cities in China, CNRDD includes 4218 images from multiple provinces, and CRDDC’22 consists of 9301 images gathered across five countries. We did not adopt the RDD2022 dataset because of its significant label imbalance and annotation inconsistencies, which could introduce noise into model training.

Each dataset was split into training, validation, and test sets using a 60:20:20 ratio. All annotations were provided in YOLO format or converted accordingly, and we used the LabelImg tool for any necessary modifications or corrections. The specific road defect types included in each dataset are summarized in Table 2.

All experiments were conducted on a Windows 10 workstation equipped with an Intel Core i9-10900K CPU and an NVIDIA A100 80GB. The MGD-YOLO model was implemented in PyTorch and trained for 200 epochs with a batch size of 16. All settings aligned with the TD-RD. To enhance generalization, we employed standard data augmentation techniques including mosaic augmentation, random scaling, and horizontal flipping. We also fixed a random seed to ensure reproducibility of results and repeated each experiment three times to report averaged performance.

To thoroughly assess the effectiveness and efficiency of our proposed MGD-YOLO (TD-YOLOv10) framework, we conducted comparative experiments against a wide range of state-of-the-art object detectors, including both CNN-based models (e.g., YOLOv5/6/7/8/9/10 series, PP-PicoDet) and Transformer-based architectures (e.g., YOLOS, RT-DERT, Lite-DERT). The evaluation was performed on three representative road defect detection datasets: TD-RD [59], CNRDD, and CRDDC’22.

Table 3 summarizes the results in terms of mean average precision (mAP), precision (Pre), computational cost (FLOPs), and inference speed (FPS). The best results are highlighted in bold, the second best in red, and the third best in blue.

As seen in the table, our MGD-YOLO consistently achieves top-tier performance across all benchmarks. Specifically, on the TD-RD dataset, it delivers the highest mAP of 87.9%, significantly outperforming models such as YOLOv9-s (85.2%) and RT-DERT (87.7%). In terms of inference speed, MGD-YOLO runs at 240 FPS, which is competitive with lightweight models like YOLOv8-n and YOLOv10-n, while maintaining superior detection accuracy.

Across the CNRDD and CRDDC’22 datasets, MGD-YOLO also demonstrates strong generalization ability, achieving the highest or second-best scores in both mAP and precision. Notably, on CRDDC’22, it reaches a precision of 53.8%, narrowly trailing the best model in that metric while outperforming all others in speed and FLOPs efficiency.

These results highlight the capability of MGD-YOLO to strike a fine balance between detection accuracy, inference efficiency, and deployment readiness–making it well-suited for real-time road defect detection in both cloud-based and edge-based environments.

To further assess the interpretability and effectiveness of MGD-YOLO, we conducted qualitative visualizations including feature space distribution via t-SNE and attention heatmaps from the VGAU module.

Comparison with Transformer-Based Detectors. In addition to YOLO-based baselines, we compared MGD-YOLO against lightweight transformer-based detectors such as RT-DETR-R18 and Lite-DETR. To ensure fairness, we selected configurations with similar FLOPs and parameter scales to MGD-YOLO. As shown in Table 4, MGD-YOLO consistently outperforms these models on the TD-RD dataset, demonstrating both higher accuracy and better inference speed.

Feature Embedding Visualization. We utilized t-distributed Stochastic Neighbor Embedding (t-SNE) to project high-dimensional features extracted from the penultimate layer of different models onto a 2D space. As shown in Fig. 9, MGD-YOLO produces more compact and well-separated clusters for each defect class, indicating superior discriminative capability in feature learning compared to the baseline.

Attention Map Visualization. We further visualized the attention responses from the VGAU module to understand how the model focuses on defect regions. Qualitative results in Fig. 9 further illustrate that MGD-YOLO yields more complete and accurate detections, particularly for small and ambiguous defects. As illustrated in Fig. 10, MGD-YOLO demonstrates stronger spatial localization capability by highlighting regions with fine-grained details, such as hairline cracks and boundary edges of potholes, which are often missed by other models. As shown in Fig. 11, our model outperforms other baselines.

These qualitative results corroborate our quantitative findings and demonstrate that MGD-YOLO not only improves detection accuracy but also enhances feature representation and localization precision.

To further validate the individual contributions of each module in the proposed MGD-YOLO framework, we conducted a series of ablation experiments focusing on the three core components: Multi-Scale Dilated Attention (MSDA), Depthwise Separable Convolution (DSC), and the Visual Global Attention Upsampling (VGAU) module. Table 5 summarizes the detection performance under various module configurations on the road defect dataset.

The experimental results highlight the effectiveness of each module. Specifically, the incorporation of DSC improves precision significantly, suggesting its utility in enhancing feature representation efficiency. MSDA proves beneficial for increasing mAP, although it results in a slight drop in recall when used in isolation. On the other hand, VGAU introduces a strong gain in precision (+4.0%) and contributes to better localization of fine-grained defect regions.

When all three modules are integrated into MGD-YOLO, the model achieves the highest overall performance, with an 6.9% increase in precision, a 1.6% improvement in recall, and a 6.6% gain in mAP50 over the original YOLOv5 baseline.

The training dynamics, illustrated in Fig. 12, show that MGD-YOLO converges faster and more stably than its counterparts, while also achieving higher final accuracy.

In real-world applications, deployment efficiency on different hardware platforms is a critical factor for road defect detection systems. To evaluate the practical deployability of MGD-YOLO, we conducted inference speed and memory usage tests on three representative hardware environments: NVIDIA A100 GPU (datacenter server grade), NVIDIA RTX 4090 GPU (consumer high-end grade), and NVIDIA Jetson Xavier NX (embedded edge device).

On the A100 GPU, MGD-YOLO achieved an average inference speed of 240 FPS with a peak memory usage of 3.2 GB. On the RTX 4090, the model achieved 215 FPS while maintaining a memory usage of 2.7 GB. On the Jetson Xavier NX, after TensorRT optimization and model pruning, MGD-YOLO maintained a real-time performance of approximately 45 FPS with a memory footprint of 1.8 GB.

These results demonstrate that MGD-YOLO strikes a favorable trade-off between detection accuracy and computational efficiency, enabling deployment across a wide spectrum of hardware platforms–from high-performance servers to resource-constrained edge devices. Notably, the integration of Depthwise Separable Convolution (DSC) and Visual Global Attention Upsampling (VGAU) significantly contributes to the reduction of model size and inference latency without sacrificing accuracy.

Therefore, MGD-YOLO offers a flexible and scalable solution for intelligent road maintenance applications, supporting both cloud-based large-scale monitoring and decentralized on-vehicle real-time inspection systems. Future work will further optimize model quantization and pruning strategies to enhance deployment efficiency on ultra-low-power embedded systems.

Although MGD-YOLO demonstrates strong overall detection performance, some misclassification cases were observed, particularly between visually similar road defect types. To better understand these errors, we conducted a qualitative analysis on the TD-RD, CNRDD, and CRDDC’22 datasets.

We found that hairline cracks are occasionally confused with patch edges or surface texture artifacts, especially under poor lighting or complex backgrounds. For instance, in low-resolution images or heavily textured asphalt surfaces, small cracks may be misidentified as material joints or construction patches. Similarly, certain pothole boundaries with gradual depth transitions were sometimes mistaken for repaired areas with minor surface degradation.

Representative examples of such misclassifications are illustrated in Fig. 13. These cases highlight the inherent difficulty in distinguishing fine-grained defect boundaries based solely on visual appearance, especially when spatial scale and intensity contrast are minimal.

To address these challenges, several potential strategies are considered for future enhancement: (1) introducing multi-scale post-processing techniques to refine defect boundaries at different spatial resolutions, and (2) incorporating complementary sensing modalities such as infrared imagery or 3D surface profiling data to provide additional discriminative cues beyond RGB textures.

We plan to explore these directions in future work, aiming to further boost the detection accuracy and robustness of MGD-YOLO, particularly for small, low-contrast, or visually ambiguous road defects in diverse real-world environments.

To evaluate the robustness of MGD-YOLO under different experimental settings, we conducted a brief sensitivity analysis focusing on two factors: input image resolution and dataset split ratio.

Image Resolution: We varied the input size between 512×512, 640×640 (default), and 768×768 pixels. As shown in Table 6, MGD-YOLO maintained stable performance, with mAP50 fluctuating within 1.2% across resolutions, demonstrating resilience to input scale changes.

Dataset Split Ratio: We tested different training/validation/test splits, specifically 70/15/15 and 60/20/20. As summarized in Table 7, the model exhibited less than 1.0% variation in mAP50, indicating good generalization under different data partitions.

These results verify that MGD-YOLO maintains robust detection performance across varying resolutions and dataset splits, supporting its practical deployment under diverse operational conditions.

This paper presents MGD-YOLO, an improved object detection framework based on YOLOv5, specifically designed for accurate and efficient road defect detection. By incorporating Multi-Scale Dilated Attention (MSDA), Visual Global Attention Upsampling (VGAU), and Depthwise Separable Convolution (DSC), the proposed model significantly enhances feature extraction, contextual reasoning, and computational efficiency. Extensive experiments on three public road defect datasets demonstrate that MGD-YOLO outperforms state-of-the-art models in both detection accuracy and inference speed. The model achieves superior performance in identifying diverse defect types–including cracks, potholes, and repairs—while maintaining a lightweight architecture suitable for real-time applications. Qualitative visualizations and ablation studies further confirm the effectiveness of each proposed component. In addition, overfitting was carefully monitored during training through validation loss tracking, early stopping, and standard data augmentation strategies.

In future work, we aim to further optimize MGD-YOLO by exploring lightweight backbone alternatives and neural architecture search techniques to reduce the model’s complexity without compromising detection accuracy. Specifically, we intend to investigate the integration of efficient transformer-based modules or dynamic convolution operators to further enhance multi-scale feature extraction while maintaining low computational overhead.

We also plan to extend our method to multi-modal data settings, incorporating complementary cues such as thermal or LiDAR information to enhance robustness under adverse environmental conditions. Incorporating heterogeneous sensing modalities will allow MGD-YOLO to better capture subtle surface anomalies and environmental context, improving detection performance under low-visibility conditions such as nighttime, rain, or dust.

Additionally, we will explore domain adaptation strategies to improve generalization across different geographic regions, pavement materials, and lighting variations. We are particularly interested in adopting invariant representation learning techniques and domain adversarial training frameworks to minimize generalization error when transferring the model to new domains with distinct feature distributions.

Furthermore, we plan to conduct systematic sensitivity analyses on data resolution, dataset split ratios, and sensor variations to evaluate the robustness of the model under diverse operational settings, ensuring its reliability and stability during real-world deployment.

Ultimately, our goal is to deploy MGD-YOLO in edge devices and smart transportation systems to enable large-scale, real-time road condition monitoring in the wild. To support practical deployment, we will also benchmark the model’s performance and resource consumption across different hardware platforms, including mobile GPUs and embedded systems, providing comprehensive guidelines for hardware-software co-optimization.