Research on Remote Detection and Width Measurement of Dangerous Rock Cracks Based on Improved YOLOv10s and Edge-based Minimum Width

In response to the current problem of relying on manual field observation for detecting and measuring dangerous rock cracks, this paper proposes an enhanced YOLOv10s model for rock crack recognition and the Edge-based Minimum Width (EMW) for width measurement. The model was trained using a self-made dataset designed explicitly for rock crack detection. We integrated the improved Convolutional Block Attention Module (CBAM) and Ghost Convolution (GhostConv) to form C2fGBA, maintaining the performance while reducing the cost. Meanwhile, Switchable Atrous Convolution (SAC) is used to dynamically adjust the receptive field and enhance the perception at different scales. To address the issue of unbalanced sample quality in the dataset, we introduce Wise Intersection over Union (WIoU), which ensures high-quality samples contribute more during training and significantly improve the convergence speed. Additionally, we propose a one-stage width measurement algorithm that simplifies the process and enhances stability compared to traditional skeletonization-based methods. The experimental results demonstrate that our improved model achieves a mAP@0.5 of 78.5%, with a 3% increase comparing to the original model, and decreases of parameters and FLOPs by 2.9% and 8.2%, respectively. Other state-of-the-art algorithms, such as Faster R-CNN and SSD, achieved mAP@0.5 scores of 64.1% and 63.6%, respectively. The EMW shows stronger robustness to round-shape noise, with an average relative error of 13.4%, compared to 28.4% and 27.2% by two skeletonization-based methods. This paper provides a comprehensive solution for dangerous rock crack detection and width measurement using machine vision, offering valuable technical references for related detection and survey tasks.