面向电力设备红外-可见光图像配准的自适应监督重训算法  

Adaptive Algorithm with Supervision and Retraining for Power Equipment Infrared and Visible Images Registration

作　　者：范澜珊 刘云鹏 刘一瑾 赵涛 裴少通 闫泽玉 FAN Lanshan;LIU Yunpeng;LIU Yijin;ZHAO Tao;PEI Shaotong;YAN Zeyu(Hebei Provincial Key Laboratory of Power Transmission Equipment Security Defense,North China Electric Power University,Baoding 071003,China)

机构地区：[1]河北省输变电设备安全防御重点实验室(华北电力大学),保定071003

出　　处：《高电压技术》2025年第4期1785-1800,共16页High Voltage Engineering

基　　金：国家电网公司科技项目(5700-202219436A-2-0-ZN)。

摘　　要：为实现不同光学模态信息优势互补,以助力电力设备故障检测与定位任务,该文采用可见光图像增强红外图像的纹理信息。针对现有红外-可见光图像配准技术难以精确对齐电力设备局部精细化结构的问题,首次提出自适应监督重训配准算法(adaptive registration algorithm with supervision and retraining,ARSR),主要包括双阶各向异性高斯方向导数机制(dual order anisotropic Gaussian directional derivative,Dual-AGDD)以及双视图匹配参数重训框架(double-view matching parameter retraining,DVMPR)。首先,提出Dual-AGDD完成特征点筛选与定向。1阶AGDD进行自适应电力设备局部细化角点检测,2阶AGDD构建高斯特征三角形确定特征点主方向,采用局部强度不变性方法构建特征描述子。接着,提出DVMPR框架对图像透视尺度与视野旋转进行制约校正。最后,基于3σ原则改进支持向量回归,对误匹配点进行剔除,完成异源数据配准。试验结果显示,对不同旋转和尺度差异、不同环境的电力设备异源图像进行配准时,该文算法的平均定位误差为2.65,平均配准精确率为98.57%,具有较强的图像旋转、尺度不变性和环境鲁棒性,显著优于现有CAO-C2F、SuperPoint-SuperGlue等配准算法,可提高电力设备精细化结构异源图像配准精度。In order to realize the complementary advantages of different optical modalities and to help power equipment fault detection and location tasks,this paper uses visible images to enhance the texture information of infrared images.The existing infrared-visible image registration technology is difficult to accurately align local fine structures of power equipment,thus,to address this issue,an adaptive registration algorithm with supervision and retraining(ARSR)is proposed for the first time.The algorithm mainly includes a dual order anisotropic Gaussian directional derivative(Dual-AGDD)mechanism and a double-view matching parameter retraining(DVMPR)framework.Firstly,Dual-AGDD is proposed to filter and orient feature points.The 1st-AGDD performs adaptive local refinement corner detection for power equipment,and the 2nd-AGDD constructs a Gaussian feature triangle to determine the main direction of the feature point.The local intensity invariance method is adopted to construct feature descriptors.Then,the DVMPR framework is proposed to correct and constrain the perspective scale and view rotation of images.Finally,based on the 3σprinciple,support vector regression(SVR)is improved to eliminate the mismatched points and complete the registration of heterogeneous data.The experimental results show that,when heterogeneous images of power equipment with different rotations and scale differences or in different environments are registered,the average RMSE of the algorithm in this paper is 2.65,and the average precision is 98.57%.The results exhibit strong rotation invariance,scale invariance,and environmental robustness,and are significantly better than those of existing registration algorithms such as CAO-C2F and SuperPoint-SuperGlue algorithms.The algorithm can improve the registration accuracy of heterogeneous images of power equipment with fine structures.

关 键 词：电力设备 红外-可见光图像配准 各向异性高斯方向导数 自适应监督重训 角点检测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]