A road defect detection algorithm incorporating partially transformer and multiple aggregate trail attention mechanisms

Abstract Road infrastructure, fundamental to daily life, inevitably sustains damage over time. Timely and precise identification and remediation of road defects are critical to prolong the lifespan of roads and ensure driving safety. Given the limitations of the widely-used You Look Only Once (YOLO) algorithm, including its insufficient receptive field and suboptimal detection accuracy, this paper introduces a novel road defect detection method. First, we propose a new attention mechanism, aggregate multiple coordinate attention, that effectively retains and concatenates channel information while preserving localization data, thereby enhancing the focus on intrinsic features. Second, we design a cross stage partial—partially transformer block (CSP_PTB) that combines CNNs and transformers to yield richer and more varied feature representations. Finally, we develop a novel neck structure, the re-calibrated feature pyramid network (Re-Calibration FPN), which selectively combines boundary and semantic information for finer object contour delineation and positional recalibration. Experimental results show that the S version of the algorithm in this paper achieves a detection accuracy of 73.2% on the road defect dataset, which is 4.2% higher than the YOLOv8 algorithm. Additionally, with an FPS of 80, it meets the requirements for real-time detection, achieving a good balance between detection speed and detection accuracy. Additionally, it exhibits excellent generalizability and robustness on the UAV asphalt pavement distress and PASCAL VOC 2007 datasets.