Crack Detection and Analysis Based on Improved YOLOv8n

This paper proposes YOLOv8sepe, an enhanced crack detection model based on improved YOLOv8 architecture, which integrates Partial Convolution (PConv) and Squeeze-and-Excitation (SE) modules to significantly improve crack detection performance in complex scenarios. The PConv module employs a channel-selective computation strategy to reduce computational complexity while preserving critical feature information, while the SE module enhances sensitivity to fine crack features through channel recalibration. Experimental results on a composite dataset combining RDD2022 and custom-collected samples demonstrate that YOLOv8sepc achieves superior performance with 0.961 mAP50 and 51.7 FPS, outperforming conventional methods including Fast-RCNN ($0.862 \mathrm{mAP50}$), SSD ($0.862 \mathrm{mAP50}$), and baseline YOLOv5 (0.892 mAP50). The model exhibits particular strength in detecting challenging transverse cracks, with detection accuracy improved by 69.2% compared to the baseline. The proposed architecture maintains efficient computational performance, making it suitable for real-time structural health monitoring applications.