Research On Image Visual Attribute Learning For Cross Domain Classification

The performance of computer vision methods are heavily dependent on the choice of the data representation.Representing data in a more visual and effective way has been an important research direction in computer vision on the background of big data of images.In recently years,visual attribute stands out and has been paid close attention of the rearchers due to its successful applications in image understading,image clssification,image retrival and etc.Though its potential in different fields of computer vision,there are problems that image data with visual attribute are far from enough and visual attribute is diffcult to annotate,limiting the widely application of the visual attribute.In order to solve the problem,applying the visual attribute on other domain data effectively with limited annotated data,this thesis has conducted reasearch on image visual attribute learning and improved its effects on different domains based on domainn adpatation in trasfer learning.The main work can be summarized as follows:Firstly,in terms of four abstract levels of visual attribute dataset,low-level visual features,bag of words,and visual attribute,the thesis gives the details of general visual attribute learning.Following that,the thesis compare the visual attribute and visual features in semantic discrimitive power and classification performace with throughly experiments.The results show that the visual attribute has advantages over visual features in semantic discrimitive power and classification performace.Secondly,Proposing a discrimitive subspace alignment based visual attrubte learning method for the domain shift when learning across domains.The proposed method imcoporates the label imformation into the subspce learning and generate the pseudo lable of source domain based on source domain.This method reduces the effect of domain shift by alignmenting the discrimilative subspace learned on source and target domains.The experiments show that the method can preseve the semantics of visual attribute and improve its semantic discrimitive power and classification performace compared with visual attribute without domain adpatation.Finally,Proposing a latent sparse subspce based visual attribute learning method inspired by the idea of adpting the domains by reducing the discrepancy between domains.The method propose to project the source and target data onto a common latent generalized space which is a union of subspaces of source and target subspace and learn the sparse representation in the latent generalized subspace.By employing the minimum reconstruction error and maximum mean discrepancy(MMD)constraints,the structure of source and target domain are preserved and the discrepancy is reduced between the source and target domains and thus reflected in the sparse representation.