Power infrastructure inspection using an improved YOLOv5s model for UAV-based image target detection

Abstract

An improved YOLOv5s(You Only Look Once version 5s) object detection method is proposed to address the issues of target occlusion, frequent missed small targets, and low recognition accuracy in unmanned aerial vehicle (UAV) detection images of power infrastructure sites. GrabCut(Interactive Foreground Extraction using Iterated Graph Cuts) and template matching are combined to achieve pixel-level foreground and background fusion data enhancement, enriching multi-scene training samples. CBAM(Convolutional Block Attention Module)is an integrated attention module with specified structural parameters into the backbone and neck of YOLOv5s, and replace the original CIoU(Complete Intersection over Union) loss with SIoU(Scalar Intersection over Union) loss to optimize bounding box regression. The experimental results show that the inference speed of this method remains at 50 FPS(Frames Per Second), which can accurately identify various safety hazards in power construction sites and provide reliable technical support for the intelligent detection of drones.