Abstract: The road crack detection task presents various defect types and complex anomaly regions. The current object detection algorithms suffer from redundant feature processing in channel and spatial dimensions, blind fusion of stage information, and other issues, leading to low network efficiency and loss of crucial information. This paper introduces a dual-refinement gate-controlled adaptive fusion network (DR-DETR) that achieves dual-refinement processing of features in channel and spatial dimensions in latent space, addressing previous issues and enhancing road crack detection accuracy. This paper constructs a dual-refinement information distillation mechanism in channel and spatial dimensions, distills redundant features separately in channel and spatial dimensions, reduces redundant processing of features by the network, and achieves efficient representation of crucial features. Aiming at the coarse-grained fusion of stage features, a feature-gated fine-grained adaptive fusion module (FGAF-Fusion) is proposed, which utilizes enhanced stripe-wise dilated convolutions to capture global information, then employs contrastive perceptual attention for inter-channel interaction and fusion, while utilizing a gate-adaptive fusion mechanism to select critical semantic information of small targets. This paper designs the Res-DCNv3 module, utilizing the flexibility of DCNv3 deformable convolutions to accurately extract diverse features of road cracks with varying morphologies. Experimental results on the RDD2022 public dataset demonstrate that the proposed DR-DETR achieves mAP0.50 and mAP0.50:0.95 of 51.7% and 24.9%, respectively, representing improvements of 4.2 and 3.3 percentage points over RT-DETR. In road crack object detection tasks, the proposed DR-DETR can effectively detect various types of road defects, demonstrating highly competitive detection results and good robustness.

Key words: defect detection, dual refinement, gate-controlled adaptive fusion, RT-DETR, deformable , convolution

摘要： 道路裂缝存在缺陷类型多样、异常区域复杂等特点，当前的目标检测算法在通道和空间维度存在冗余的特征处理、阶段信息的盲目式融合等问题，导致网络效率低、关键信息丢失。提出了一种双重细化门控自适应融合网络（DR-DETR），在潜在空间中实现特征在通道和空间维度上的双重精细化处理，解决上述问题的同时，也提高了道路裂缝的检测精度。构建一种通道-空间双重细化的信息蒸馏机制，分别对通道和空间维度内冗余的特征进行信息蒸馏，减少网络对特征的冗余处理，实现关键性特征的高效表征；针对阶段性特征的粗粒度融合，提出了特征信息门控自适应融合模块（FGAF-Fusion），利用增强型条带大核卷积获取全局信息，借助对比感知注意力实现通道间的交互和融合，同时利用门控自适应融合机制筛选出关键性的小目标语义信息；设计Res-DCNv3模块，利用DCNv3可变形卷积的灵活性来精确提取形态各异的道路裂缝特征。在RDD2022公开数据集中的实验结果显示，提出的DR-DETR在mAP0.50和mAP0.50:0.95分别达到了51.7%和24.9%，相较于RT-DETR分别提升了4.2和3.3个百分点。在道路裂缝目标检测任务中，提出的DR-DETR可以有效检测出不同类型的道路缺陷，展现出极具竞争性的检测结果和良好的鲁棒性。

关键词: 缺陷检测, 双重细化, 门控自适应融合, RT-DETR, 可变形卷积