RCYOLO: An Efficient Small Target Detector for Crack Detection in Tubular Topological Road Structures Based on Unmanned Aerial Vehicles

Abstract

Unmanned Aerial Vehicles (UAVs) combined with target detection algorithms can enhance the detection of road cracks. In response to the challenges presented by complex crack shapes and textures, small sizes, and highly integrated backgrounds, this article developed a UAV-based road crack target detection algorithm using road crack you only look once (RCYOLO). RCYOLO was composed of a C2f-DySnakeConv (C2f-DSConv) module located in the ninth layer and a simulated attention mechanism (SimAM) module situated above the Spatial Pyramid Pooling-Fast (SPPF), along with a dyhead attention detection head that integrated three types of attention mechanisms. Initially, the C2f-DSConv was proposed to effectively extract tubular features of cracks. Subsequently, the SimAM addressed the issue of target-background fusion, enhancing feature recognition of the targets while suppressing background interference. Finally, the dyhead strategy incorporated three types of attention mechanisms, effectively resolving the issue of small target omissions. Our results showed that on the custom UAV dataset road crack image, which included close-range and long-range images, RCYOLO outperformed the baseline network YOLOv8 by 5.9% in mAP@0.5, 6.5% in recall, and 9.8% in precision. On the public dataset Detection of Objects in Aerial Images, mAP@0.5 also exceeded YOLOv8 by 5.8%, indicating that RCYOLO performed well in other remote sensing image tasks, making this target detection algorithm more suitable for high-altitude photography of crack targets than other mainstream algorithms.