光学精密工程2025，Vol.33Issue(8)：1274-1288,15.DOI:10.37188/OPE.20253308.1274
电力巡检中的偏振图像特征融合
Polarized image feature fusion in power inspection
摘要
Abstract
To enhance target detection accuracy in smart grid monitoring,especially under the challenging conditions posed by complex outdoor lighting,a comprehensive framework for fusing polarized and intensi-ty images was proposed.To improve the accuracy of detecting potential hazards in smart grid monitoring systems under complex lighting conditions,this paper proposed a detection method based on the fusion of polarization and light intensity dual-modal information.This framework addressed the inherent difficulties of image analysis in diverse lighting scenarios,ensuring robust and accurate monitoring.Firstly,a dual-path feature fusion network was designed,which used dense convolutional modules to extract features from polarized intensity images and polarization degree images separately,thereby enhancing the retention capability of shallow information.Simultaneously,by constructing feature dependencies in both spatial and channel dimensions,new feature maps were selectively generated,solving the feature aggregation prob-lem in feature fusion.Finally,a multi-scale adaptive structural similarity loss function was introduced,and a weighted algorithm was designed to optimize the quality of reference image generation,enhancing the structural fidelity and target saliency of the fused images,and further improving their quality.Experimen-tal results demonstrate that,compared to state-of-the-art image fusion algorithms,the proposed method shows significant performance improvements across multiple evaluation metrics,compared to intensity im-ages(S0).These improvements are not only statistically significant but also visually apparent,as the fused images produced by our method are clearer,more detailed,and more informative.Ablation experiments further validate the effectiveness and practicality of the network modules and loss function.In a custom tar-get detection dataset,the fused images generated by this method achieve a recognition accuracy of 91.5%,with an mAP@0.5 score of 0.916,These results showcase the superior performance of our meth-od in objective evaluations and highlight its significant contribution to enhancing the detection accuracy of subsequent target detection networks in smart grid monitoring.关键词
电力巡检/机器视觉/偏振成像/图像融合/目标检测Key words
power inspection/machine vision/polarization imaging/image fusion/object detection分类
计算机与自动化引用本文复制引用
倪梦瑶,彭元龙,胡尚,闫龙川,郑锦坤,曹丹华..电力巡检中的偏振图像特征融合[J].光学精密工程,2025,33(8):1274-1288,15.基金项目
国家电网公司科技项目(No.5700-202325308A-1-1-ZN) （No.5700-202325308A-1-1-ZN）
