An efficient fusion detector for road defect detection.

As deep learning networks deepen, detecting multi-scale subtle defects is a challenging task in road images with complex background, due to some fine features gradually disappearing, which significantly increases the difficulty of extracting these fine features. To address this problem, an SCB-AF-Detector is proposed, which combines space-to-depth convolution with bottleneck transformer and employs enhanced asymptotic feature pyramid network to fuse features. Firstly, an SCB-Darknet53 backbone network is designed, which integrates SPD-Conv structure and bottleneck transformer to effectively extract the subtle and distant defect features in complex background. And then, asymptotic feature pyramid network is developed, which first fuses the two shallow semantic features of the backbone network, and then fuses the deep semantic features. In this way, the subtle features in the shallow layer can be preserved, and the deep semantic features can be extracted. Finally, experiments are carried out on the Iran Road Disease Dataset (IRRDD), which contains 25,000 road images. The results show that the proposed method achieves 90.8% (Precision), 95% (Recall) and 75.2% (mAP) in the classification and detection of multi-scale subtle defects respectively, which meets the high-precision detection requirements of road defects.