摘要
已有的高阶算法中,构建相似模型时仅使用少量超边构建稀疏相似模型,同时高阶相似模型仅考虑使用单阶的高阶相似关系.为解决这两个问题,以规范割算法为基础,采用直推式学习技术,从标准化和非标准化拉氏矩阵两个角度分别构建全相似高阶模型和全相似多阶相似模型.根据规范割算法构建直推式学习框架,然后展示该框架如何在算法中训练全相似关系.研究结果显示,在所提出的算法中超边之间的全相似关系能以一个简洁的形式应用.以此为基础,将多阶全相似关系进行融合,提出融合多阶信息的全相似多阶相似模型.将构建的全相似高阶相似模型和全相似多阶相似模型应用到规范割算法框架中,提出全相似高阶规范割算法和全相似多阶规范割算法.在两种高阶相似模型中,全相似张量采用稀疏张量逆的形式,并且该逆矩阵可以转换为规范割框架中稀疏张量特征分解问题.将所提出的算法应用于运动分割,并与现有的高阶算法进行对比,实验结果显示,所提出的算法具有一定的优势.
In the higher-order algorithms,only a small number of super edges are used to construct the sparse affinity model,and the higher-order affinity model only considers the single-order higher-order similarity relationship.In order to solve these two problems,based on the normalized cut algorithm and by using the transductive learning technique,a full higher-order affinity model and a full multi-order affinity model are constructed from the two perspectives of normalized and non-normalized Laplacian matrixes.A Transductive learning framework is constructed based on the normalized cut algorithm,and it is showed how the framework trains the full similarity in the algorithm.The final result shows that the full similarity between the super edges in the algorithm can be applied as a simple form.Based on this,the multi-order full similarity relation is merged,and a fully similar multi-order affinity model with multi-order information is proposed.The constructed full higher-order affinity model and the full multi-order affinity model are applied to the framework of the normalized cut algorithm,and a full affinity higher-order normalized cut algorithm and a fully affinity multi-order normalized cut algorithm are proposed.In the two higher-order affinity models,the full affinity tensor is in the form of sparse tensor inverse and the inverse matrix can be transformed into the sparse tensor feature decomposition problem in the normalized cut.In the experimental part,the proposed algorithm is applied to motion segmentation and compared with the existing higher-order algorithms.The results show the advantages of the propesed algorithm.
作者
张敬茂
沈艳霞
ZHANG Jing-mao;SHEN Yan-xia(Engineering Research Center of IoT Technology and Application of MOE,Jiangnan University,Wuxi 214122,China)
出处
《控制与决策》
EI
CSCD
北大核心
2020年第4期852-860,共9页
Control and Decision
基金
国家自然科学基金项目(61573167,61572237)
江苏省研究生科研与实践创新计划项目(KYCX17_1454).
关键词
全相似高阶模型
全相似多阶模型
直推式学习
规范割
运动分割
full higher order affinity model
full multi-orders affinity model
transductive inference
normalized cut
motion segmentation
