Precision detection of miniature dam cracks with a multi-scale enhanced YOLO framework for UAV inspection

Accurate detection of dam cracks from Unmanned Aerial Vehicle (UAV) imagery is crucial for structural health monitoring. However, prevailing methods face significant challenges in achieving a balance between the precise identification of minuscule cracks and computational efficiency when processing high-resolution images. To overcome these limitations, this paper proposes HiResDC-YOLO, a novel deep learning framework based on an enhanced YOLOv12 architecture. Our main contributions are threefold: First, we introduce an Adaptive Dynamic Transformer (ADyT) module to strengthen nonlinear feature representation and stabilize gradient flow during training. Then, we design a Multi-Scale Enhanced Detection (MSED) head that effectively utilizes shallow, high-resolution features to significantly improve the detection capability for fine cracks. Besides, we develop a Multi-Scale Convolutional Attention (MSCA) module to capture comprehensive contextual information across different scales by integrating deep convolutional layers and a multi-branch fusion mechanism. Furthermore, we propose a High-Resolution Adaptive Inference (HiResInfer) strategy, which utilizes region-guided slicing and feature caching to dramatically accelerate the inference speed on full-resolution images without compromising the integrity of small targets. Extensive experiments on a challenging self-collected UAV dam crack dataset demonstrate that HiResDC-YOLO achieves state-of-the-art performance, surpassing existing methods significantly in terms of precision, recall, and mean Average Precision (mAP), while maintaining high computational efficiency. This work presents a robust and practical solution for real-time dam inspection and engineering safety monitoring. The code and related resources are available at: https://github.com/lijin6/HiResInfer-YOLO.git .