电力学报2024，Vol.39Issue(4)：281-291,11.DOI:10.13357/j.dlxb.2024.031
基于多尺度特征融合的变电站低照度表计图像增强网络
Low-illumination Image Enhancement Network of Substation Meter Based on Multi-scale Feature Fusion
摘要
Abstract
Aimed at the problems such as insufficient brightness,weak contrast,excessive noise and lack of detail caused by low-illumination image of substation meter,a low-illumination image enhancement network based on multi-scale feature fusion(MFF-Net)is proposed.The network model consists of three sub-networks,namely shallow feature extraction network,deep feature fusion network and brightness adjustment network.Firstly,the low-illumination image is extracted by the shallow feature extraction module composed of InceptionV3 net-work,the details of image edge and texture are initially extracted.Secondly,a deep feature fusion network com-posed of up and down sampling structure is designed,and a multi-task Transformer module is introduced to ad-just feature image details,color difference and noise removal.Meanwhile,a cross-layer attention enhancement module with integrated attention mechanism is designed to adaptively fuse the feature relationships between dif-ferent feature layers.Finally,through the brightness adjustment network,the output image brightness is more uniform,and human eye sensory effect is improved.L1 loss is taken from the enhanced images of different scales,and the total loss function is obtained by summing to measure the training loss of the network.The abla-tion comparative experimental results show that compared with other algorithms,the proposed network can re-store details while improving image brightness and reducing noise.The rPSNR value on LOL and MIT-Adobe FiveK test sets is increased by 3.57%,4.32%,and sSSIM value is increased by 1.58%and 1.49%,respectively.The image enhancement network of the actual substation low-illumination meter shows strong generalization,robustness and accuracy,which provides the basis for the subsequent meter recognition reading.关键词
变电站/变电站表计/低照度图像增强/多尺度特征融合/注意力机制/亮度调整Key words
substation/substation meter/low-illumination image enhancement/multi-scale feature fusion/atten-tion mechanism/brightness adjustment分类
信息技术与安全科学引用本文复制引用
胡成扬,王旭红,樊绍胜,刘星宇..基于多尺度特征融合的变电站低照度表计图像增强网络[J].电力学报,2024,39(4):281-291,11.基金项目
国家自然科学基金项目(62271087). （62271087）
