改进R-LoFTR++的智能巡检特征匹配算法  

Improved intelligent inspection feature matching algorithm of R-LoFTR++

作　　者：舒军[1] 王江舸 杨莉[2] 舒心怡 无 SHU Jun;WANG Jiangge;YANG Li;SHU Xinyi;无(Hubei Provincial Key Laboratory of Efficient Solar Utilization and Energy Storage Operation Control,Wuhan 430068,China;School of Computer Science,Hubei Second Normal University,Wuhan 430205,China;Faculty of Science,The University of Melbourne,Melbourne VIC3010,Australia)

机构地区：[1]湖北工业大学太阳能高效利用及储能运行控制湖北省重点实验室,武汉430068 [2]湖北第二师范学院计算机学院,武汉430205 [3]墨尔本大学理学院,墨尔本VIC3010

出　　处：《重庆理工大学学报(自然科学)》2025年第2期86-96,共11页Journal of Chongqing University of Technology:Natural Science

基　　金：国家自然科学基金青年项目(62306107)。

摘　　要：在应用增强现实技术的变电站巡检工作中,背景纹理复杂难以提取特征点,大视角变化情况下匹配正确率低,针对以上问题提出改进特征匹配算法R-LoFTR++。引入高斯滤波进行预处理,有效减少图像噪声并平滑部分纹理,降低了匹配的复杂度。设计了特征方向描述子模块,增强了网络对图像关键点的方向敏感度,提高了大视角差图像间的匹配率。集成MAGSAC++算法,优化匹配过程、剔除误匹配点,提升了匹配的正确率。实验结果表明,R-LoFTR++算法在变电站真实数据集上的匹配效果都优于参与对比的其他特征匹配算法。在MegaDepth相同特定场景子集的实验中,R-LoFTR++在户外姿态评估实验中AUC指标相比于原网络提升了约0.92%~1.63%。With the deepening of smart grid construction,the traditional substation inspection model is undergoing a critical transformation towards digitalization and intelligence.We investigate two core issues in the augmented reality-based substation equipment inspection system:(1)In complex industrial scenarios,the special materials on the surfaces of power equipment and repetitive metal structures in the background cause strong reflections and low texture characteristics,leading to significant feature point loss in traditional feature matching algorithms;(2)Large perspective differences caused by inspection personnel moving between devices result in an increase in the mismatch rate of matching points in existing deep learning models.These existing problems severely limit the deep application of AR technology in the power industry.Therefore,an improved feature matching algorithm,R-LoFTR++,is proposed based on the Transformer and neural network combined algorithm LoFTR.First,a set of dynamically learnable Gaussian filters are utilized,building a mixed Gaussian kernel space withσ=1.2~2.4.It achieves a dynamic balance between noise suppression and feature retention,effectively reducing image noise and smoothing some textures,markedly lowering the complexity of subsequent feature matching,thereby improving matching efficiency.Next,for large-angle power equipment images collected,a feature orientation descriptor module is designed based on positional encoding information.The positional encoding information in large-angle images is corrected with a rotation factor,allowing the module to output consistent feature representations when the image’s perspective and scale change.This enhances the algorithm’s sensitivity to the orientation of image keypoints,reducing errors in image matching with large perspective differences and significantly improving the matching rate.Under certain degrees of perspective changes,the directional information of feature points is accurately described and matched,ensuring the stability and accur

关 键 词：特征匹配 LoFTR算法 增强现实技术 电气检测 

分 类 号：TM93[电气工程—电力电子与电力传动] TP391[自动化与计算机技术—计算机应用技术]