Improved RT-DETR for Road Defect Detection

Timely detection and repair of road defects are vital for traffic safety and road quality. Existing methods struggle with accuracy, feature extraction, and multi-scale feature fusion, especially for small object detection and redundant feature processing. This paper proposes a road defect detection algorithm based on an improved RT-DETR-R18 model. By enhancing the ResNet18 backbone and designing the PAC module, the model improves small object detection by enabling multi-scale feature extraction and redundant feature suppression. Experimental results show a 2.8% improvement in mAP 50,1.4% in mAP 50-95, 2.3% in precision, and 3.4% in recall. The improved model also reduces GFLOPs from 58.2 to 52.7 and increases FPS by nearly 10. These results demonstrate the method's effectiveness in enhancing detection accuracy, feature expression, and efficiency for road defect detection.