融合分层大核卷积与高效注意力的轻量化RT-DETR道路缺陷检测

Lightweight Road Defect Detection Based on RT-DETR Fusing Hierarchical Large Kernel Convolution and Efficient Attention

摘要: 当前道路缺陷检测算法难以兼顾高精度与实时性要求，且传统卷积神经网络因局部感受野限制，对细长裂缝等特征的捕捉能力有限。为此，本文提出一种基于改进RT-DETR的轻量化道路缺陷检测算法。该方法以端到端的RT-DETR为基线，在消除后处理延迟的同时增强全局特征建模。具体改进包括：1) 采用轻量级StarNet替换原ResNet骨干，结合大核深度可分离卷积与线性门控机制，在显著降低参数量和计算成本的前提下保持多尺度特征提取能力；2) 为缓解下采样导致的细微裂缝特征模糊与丢失，使用分层大核卷积(HLKConv)替换标准下采样层；3) 在骨干网络高层引入高效多尺度注意力模块(EMA)，通过跨通道与跨空间交互建模，提升模型对多尺度、不规则缺陷的感知与聚焦能力。在RDD2022数据集上的实验结果表明，改进后模型的mAP@0.5达到85.79%，较原始RT-DETR-R18提升0.67%，而参数量和计算量分别大幅降低55.9%和56.5%。相较于其他主流算法，本方法在维持高检测精度的同时，显著降低了硬件资源需求，更适用于计算资源受限的边缘设备进行实时道路巡检。

Abstract: Current road defect detection algorithms struggle to balance the requirements of high accuracy and real-time performance. Furthermore, traditional Convolutional Neural Networks (CNNs) are constrained by local receptive fields, limiting their ability to capture features of slender cracks. To address these issues, this paper proposes a lightweight road defect detection algorithm based on an improved RT-DETR. Leveraging the end-to-end RT-DETR as a baseline, the proposed method enhances global feature modeling while eliminating post-processing latency. The specific improvements include: 1) Replacing the original ResNet backbone with the lightweight StarNet, which combines large-kernel depth-wise separable convolutions with a linear gating mechanism to maintain multi-scale feature extraction capabilities while significantly reducing parameter count and computational costs; 2) Utilizing Hierarchical Large-Kernel Convolution (HLKConv) to replace standard downsampling layers, thereby mitigating the blurring and loss of fine crack features caused by downsampling; and 3) Incorporating the Efficient Multi-Scale Attention (EMA) module into the high-level layers of the backbone to enhance the model’s perception and focus on multi-scale and irregular defects through cross-channel and cross-spatial interaction modeling. Experimental results on the RDD2022 dataset demonstrate that the improved model achieves an mAP@0.5 of 85.79%, an increase of 0.67% over the original RT-DETR-R18, while the parameter count and computational volume are substantially reduced by 55.9% and 56.5%, respectively. Compared with other mainstream algorithms, this method significantly lowers hardware resource requirements while maintaining high detection accuracy, making it highly suitable for real-time road inspection on resource-constrained edge devices.