An RPP-YOLOv11 model for road crack detection

Abstract

Accurate and efficient road crack detection serves as a critical component in smart transportation systems and infrastructure maintenance. Existing YOLO series models still exhibit limitations in detecting cracks due to their sensitivity to subtle details, diverse morphological variations, and complex background interference from road surfaces. Building upon YOLOv11, this study enhances road crack detection capabilities through three core innovations: 1) Implementing an early-stage fusion strategy combining visible light and thermal infrared images (RGBT) to improve environmental adaptability; 2) Adopting windmill-shaped convolution modules (PSConv) to replace traditional convolutions, thereby enhancing crack feature extraction while suppressing background noise; 3) Introducing a P6 detection layer to establish a four-scale detection framework (P3-P6), expanding global perception capabilities for large-scale cracks. Experiments on the cross-border road damage dataset RDD2022 demonstrate that the proposed RPP-YOLOv11 model (integrating RGBT multispectral fusion, PSConv convolution modules, and P6 detection layer) achieves 74.90% accuracy, 64.36% recall rate, and 69.04% mAP@0.5 with 42.70% mAP@0.5:0.95. Compared to original YOLOv11 and mainstream benchmarks, this model shows significant improvements in detection precision, robustness, and computational efficiency, providing a reliable technical solution for automated road inspection systems.