红外技术2026，Vol.48Issue(5)：620-629,10.
多源图像融合与改进YOLOv8的电气设备小目标故障识别
Multi-source Image Fusion and Improved Fault Recognition of Small Targets in Electrical Equipment by Improved YOLOv8
摘要
Abstract
To solve the problem of the image target being small and susceptible to false detection caused by complex backgrounds and shadow occlusion interference in the inspection of electrical equipment,this study proposes an image fault detection algorithm for electrical equipment based on improved YOLOv8 based on image fusion of the target image.To reduce false detections and missed detections of small target faults in complex backgrounds,the recursive gated convolution of a High-Order Recursive Network(HorNet)was used to design a convolutional block with customized components(CHC),which replaces the original cross-stage partial fusion with two convolution(C2F)modules in the network,strengthens the spatial information of inter-neighborhood features,and introduces an Efficient Multi-Scale Attention(EMA)mechanism in the backbone network to capture richer detailed information and improve the feature extraction ability of the model.The Wise-Intersection over Union(WIoU)loss function was used to replace the original network loss function,effectively addressing the problems of missed detections,false detections,and overlapping target suppression.Experimental results show that,compared with the original YOLOv8n network,the improved network achieved a 19.8%increase in accuracy,a 10.8%increase in recall,and an 11.4%increase in mAP50.The proposed model demonstrates stronger feature extraction capability and higher detection accuracy for small targets and is suitable for fault detection in fused multi-source electrical equipment images.关键词
多源图像/故障识别/YOLOv8/递归门卷积/图像融合/EMAKey words
multi-source image/fault identification/YOLOv8/recursive gate convolution/image fusion/EMA分类
信息技术与安全科学引用本文复制引用
舒凡,肖志云,李绪,赵岩..多源图像融合与改进YOLOv8的电气设备小目标故障识别[J].红外技术,2026,48(5):620-629,10.基金项目
内蒙古自治区科技计划项目(2021GG0345) （2021GG0345）
内蒙古自治区自然科学基金项目(2021MS06020). （2021MS06020）
