Abstract: Abstract: Precise detection of road pavement distress was significant for ensuring traffic safety, extending road service life, and enhancing maintenance decision-making efficiency. To address poor robustness under complex background interference and low recognition accuracy for multi-scale distresses such as minute cracks in existing algorithms, an enhanced road distress detection algorithm based on an improved DEIM framework was proposed. First, a Cross-scale Feature Interaction Enhancement Structure (CCFS) was designed. By optimizing the cross-scale feature interaction mechanism to compensate for fine-grained texture information, the structure enhanced the model's capability to capture minute distress targets. Second, a texture-aware Enhanced Encoder (EAIFI) was designed to strengthen local discriminative capacity and noise robustness in complex environments. Finally, an Efficient Channel Attention (ECA) module was embedded at the front end of the decoder to suppress unstructured environmental interference and directionally enhance key distress responses. Experimental results on the public SVRDD dataset demonstrated that, compared with the baseline DEIM-D-FINE-S, the improved model achieved increases of 2.3, 3.8, and 3.1 percentage points in AP@0.50, AP@0.75, and AP@0.50:0.95, respectively, with only an increase of 0.3?M parameters and 1.8?GFLOPs in computational cost. Compared with the suboptimal DP-DETR algorithm, the improved model also achieved a higher AP@0.50:0.95 by 1.6 percentage points. The algorithm achieved a favorable balance between detection accuracy and computational cost, thus providing a feasible technical solution for automated road distress detection and maintenance.

Key words: Keywords: Road distress detection, DEIM, Multi-scale feature, Fine-grained, Complex background

摘要： 摘 要: 道路病害的精准检测对保障交通安全、延长道路服役寿命和提升养护决策效率具有重要意义。针对现有道路病害检测算法在复杂背景干扰下鲁棒性较差，且对微小裂缝等多尺度病害识别精度偏低问题，提出一种基于改进DEIM框架的道路病害检测算法。首先，设计跨尺度特征交互增强结构CCFS，通过优化跨尺度特征交互机制补偿特征的细粒度纹理，提升模型对微小病害目标的特征捕获能力。其次，设计具有纹理感知的增强编码器EAIFI，强化模型在复杂背景下的局部判别能力与抗噪鲁棒性。最后，在解码器前端嵌入高效通道注意力ECA模块，抑制非结构化环境干扰并定向增强关键病害响应。实验结果表明，在SVRDD公开数据集上，改进模型相较于基线模型DEIM-D-FINE-S，在仅增加0.3M参数量与1.8GFLOPs计算开销的条件下，AP@0.50、AP@0.75及AP@0.50:0.95指标分别提升了2.3、3.8和3.1个百分点，相较于指标次优的DP-DETR算法，其AP@0.50:0.95指标也高出1.6个百分点，该算法在检测精度与计算开销间取得良好平衡，为道路病害检测与自动化养护提供了一种可行的技术方案。

关键词: 关键词: 道路病害检测, DEIM, 多尺度特征, 细粒度, 复杂背景