基于注意力强化与补全的RGB-T图像绝缘子检测  

RGB-T image insulator detection based on attentional reinforcement and complementation

作　　者：刘建戈 陈欢 戴欣 曹莉 张鹏宇 李庆武[2] LIU Jian’ge;CHEN Huan;DAI Xin;CAO Li;ZHANG Pengyu;LI Qingwu(State Grid Jiangsu Electric Power Co.,Ltd.,Huaian Branch,Huaian 223000,China;College of Information Science and Engineering,Hohai University,Changzhou 213200,China)

机构地区：[1]国网江苏省电力有限公司淮安供电分公司,江苏淮安223000 [2]河海大学信息科学与工程学院,江苏常州213200

出　　处：《应用科技》2025年第1期213-220,共8页Applied Science and Technology

基　　金：国网江苏省电力有限公司科技项目(J2023061).

摘　　要：针对在复杂环境下绝缘子难以检测的问题,提出了一种基于注意力强化与补全的RGB-T图像绝缘子检测方法。首先,使用双流编码器来提取RGB-T图像的特征信息,结合多模态注意力与补全模块进行特征信息融合,促进跨模态特征信息的深度互补融合;其次,使用全局上下文模块捕获图像多尺度特征,结合特征增强聚合模块,促进图像多尺度特征融合,丰富图像的语义信息;最后,构建特征尺度复原模块对边缘信息进行优化,得到边缘清晰的绝缘子检测结果。实验结果表明,本文提出的方法在检测准确率方面相比其他方法有显著提升。Aiming at the problem that it is difficult to detect insulators in complex environments,an RGB-T image insulator detection method based on attention reinforcement and complementation is proposed.First,a dual-stream encoder is used to extract the feature information of RGB-T images.Applying a multimodal attention and complementation module to feature information fusion,it promotes the deep complementary fusion of cross-modal feature information.Second,a global context module captures the image multiscale features,and promotes fusion of the image multiscale features by combination with the feature enhancement and aggregation module,and thereby enriches the semantic information of the image.At last,a feature scale recovery module is constructed to optimise the edge information and get the insulator detection result with clear edges.The method proposed in this paper achieves a significan improvement in detection accuracy compared to other methods.

关 键 词：注意力强化 双流编码 特征增强聚合 多尺度特征融合 特征尺度复原 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TM769[自动化与计算机技术—控制科学与工程]