针对误匹配剔除的网格加权表征策略学习模型  

Learning model of grid weighted representation strategy for mismatch elimination

作　　者：杨丽洁 黄倩[1,2] 黄燕滢 张云飞[1,2] YANG Lijie;HUANG Qian;HUANG Yanying;ZHANG Yunfei(Key Laboratory of Water Big Data Technology of Ministry of Water Resources, Hohai University, Nanjing 211100, China;School of Computer and Information Engineering, Hohai University, Nanjing 211100, China)

机构地区：[1]水利部水利大数据重点实验室(河海大学),南京211100 [2]河海大学计算机与信息学院,南京211100

出　　处：《中国科技论文》2022年第3期274-280,共7页China Sciencepaper

基　　金：国家重点研发计划项目(2018YFC0407905);江苏省重点研发计划项目(BE2016904);中央高校基本科研业务费专项资金资助项目(B20020188)。

摘　　要：针对多数特征匹配方法在视觉任务中难以保持适用性的问题,提出了一种网格加权表征策略学习模型。基于对匹配对邻域一致性约束的认识,通过加权策略为每个匹配对创建一组匹配表征,包括邻域元素一致性、邻域加权拓扑一致性和外观匹配稳定性。这一组表征包含匹配对的空间、外观信息,可有效剔除误匹配。针对现有方法使用通用的方法来确定邻域,限制了它们在实时应用程序中的使用,考虑通过网格结构快速构建邻域,使每个匹配对确定邻域的时间复杂度与初始匹配对数量无关,提高了算法运行速度。最后将提出的方法在3个公开数据集TUM、KITTI、CPC和工程项目所收集的SFDS数据集上进行验证,并与RT、GMS、LMR方法进行了对比。实验结果表明,所提出的网格加权表征策略学习模型在精度和召回率指标上表现最好,在速度上是LMR方法的2倍。因此,网格加权表征策略学习模型在误匹配剔除的效果和速度上实现了更好的平衡。Aiming at the problem that most feature matching methods are difficult to maintain applicability in visual tasks,a learning model of grid weighted representation strategy was proposed.Based on the true matching neighborhood consistency constraint,a weighted strategy was used to create a set of matching representations for each matching,including neighborhood element consistency,neighborhood weighted topology consistency,and appearance matching stability.The group of representations contained spatial and appearance information,which effectively rejected false matching.For existing methods,general methods were used to determine neighborhoods,which limits their use in real-time applications.To quickly determine neighborhoods,grid structure was used to make the time complexity independent of the putative matching number.The proposed method was verified on three public data sets TUM,KITTI,CPC and SFDS datasets collected by engineering projects,and compared with RT,GMS,and LMR methods.Experimental results show that the proposed learning model of grid weighted representation strategy perform best in both precision and recall metrics,and is 2 times faster than the LMR method.Therefore,the grid weighted representation strategy learning model achieves a better balance between effectiveness and speed of mismatch removal.

关 键 词：模式识别 图像处理 特征匹配 邻域一致性 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]