Zero-shot Image Classification Based On Attribute Adaptation

In zero-shot learning,because of the non-intersection of training classes and testing classes,the data distributions of training samples and test samples are different.Therefore,the attribute classifier learned in training images is not applicable to testing images,and the domain shift problem arises.In response to the above issues.The main content of this article is expressed as follows:Firstly,the problem of attribute adaptation is solved from the perspective of classifier adaptation,and a multi-source domain attribute adaptation model based on adaptive multi-kernel alignment learning is proposed.Firstly,the construction of multi-source domain is performed according to the class-class correlation,and the weighted source domain is constructed using the association probability between attribute and source domain.Secondly,the attribute relationship between the source domain and the target domain is used to perform the attribute-oriented similar feature selection.The feature selection narrows the distribution difference between the two domains.Then,a adaptive multi-kernel attribute classifier based on kernel alignment is constructed by using pre-training attribute classifiers,multi-kernel learning,the maximum mean difference between the two domains,and the centered kernel alignment.Finally,the attribute adaptation model is applied to the zero-shot image classification through the direct attribute prediction model.Secondly,the problem of attribute adaptation is solved from the perspective of feature representation adaptation,and a multi-source domain attribute adaptation model based on deep feature transfer is proposed.Firstly,the construction of multi-source domain is performed according to the class-class correlation.Secondly,the image preprocessing is performed on the training images and testing images.Thirdly,the deep adaptation network is used to extract the transfer features from the source and target domains.The deep transfer features of source domain are used to learn attribute adaptation models,and the features of the target domain are weighted for zero-shot classification.Then,the relationship between classes and attributes is mined according to the class-attribute relevance.Finally,the attribute adaptation model and the multi-source decision fusion algorithm are introduced into the indirect attribute prediction model to complete zero-shot image classification.In this paper,simulation experiments and comparisons are carried out on the scene recognition dataset(OSR),animal dataset(AWA),a-Yahoo attribute dataset,andattribute discovery Shoes dataset(Shoes).Experimental results verify the better classification performance of proposed models in zero-shot image classification.