人民珠江2026，Vol.47Issue(4)：70-78,9.DOI:10.3969/j.issn.1001-9235.2026.04.008
基于改进YOLOv10的堤坝渗漏智能检测方法
An Intelligent Method for Detecting Leaks in Dams Based on an Improved YOLOv10 Model
摘要
Abstract
As a key component of water conservancy infrastructure,the structural safety of dams is directly linked to flood control and the safety of people′s lives and property.Seepage is one of the most common hazards affecting dams.If not identified early and effectively controlled,it can easily evolve into seepage-induced failures such as piping and soil flow under hydraulic pressure,leading to structural instability or even dam failure.However,the complexity of the dam environment—such as poor lighting conditions leading to low contrast between targets and the background,the scattered distribution of leakage points,and the difficulty in detecting leaks from small targets—poses significant challenges to the accuracy and effectiveness of target detection results.To address the above problems,this paper proposes RFS-YOLO,an intelligent dam seepage detection method based on improved YOLOv10.Firstly,a Retinex-inspired Illumination Module(RIM)is proposed,which realizes adaptive decomposition of illumination and reflection components based on Retinex theory,and effectively enhances the expression of seepage features in low-light environments.Secondly,by introducing the FourierConv frequency domain convolution module,Fast Fourier Transform(FFT)is utilized to achieve a global receptive field,which solves the limitation of local receptive field in traditional convolution and enhances the modeling capability of cross-regional feature correlation.Finally,a Scale-aware Cross-Scale Pyramid(SCSP)module is designed.Through P2 feature injection and pyramid sparse attention mechanism,it significantly improves the detection accuracy of small-target seepage points while avoiding the computational overhead of the traditional P2 detection layer.Experiments show that while maintaining its lightweight characteristics,RFS-YOLO achieves a mAP@0.5 of 72.0%,which is 3.9%higher than that of the original YOLOv10-n.关键词
堤坝安全/渗漏检测/深度学习/目标检测/YOLOv10Key words
dam safety/seepage detection/deep learning/object detection/YOLOv10分类
建筑与水利引用本文复制引用
黄鑫,余增鑫,王留毅..基于改进YOLOv10的堤坝渗漏智能检测方法[J].人民珠江,2026,47(4):70-78,9.基金项目
国家自然科学基金(52078143) （52078143）
