Research On Image Classification Based On Zero-shot Learning

Object classification is the key technology of Airborne Opto-Electronic Platform and the important issue of computer vision.With the development of deep learning,good performance of object classification can be obtained training deep networks with large amount of data.However,with the increase of data,not all classes have adequate and high-quality samples in real case.Transferring the learned model from one dataset to another is an interest research for more and more researchers.Zero-shot learning deals with the problem of predicting class labels for unseen instances.It can be considered as a special case of transfer learning where the source and target domains have different label spaces.Traditional embedding methods directly employ the projection functions learned in source domain to target domain without any adaptation,which causes projection domain shift during testing.This article proposes three kinds of zero-shot object classification.It relates the domain shift and domain adaption by proposing a bi-directional embedding method.So,an explicit expression is acquired.Besides,in this article,we develop an adaptive bi-directional embedding method to solve the problem.An adaptive feature space is learned with bi-directional embedding by exploiting the information of the target feature distribution.It reduces the distribution differences between the shifted and the actual target features to alleviate the projection domain shift.In addition,this article proposes an aligned multi-view embedding exploring the complementarity of multiple semantic information to improve the performance of the multi-view embedding in other way.Existing multi-view zero-shot learning haven't considered the inconsistent physical meaning of the multi-view information,which is the drawback of existing multi-view embedding methods.This article aligns the multi-view information making the multi-view embedding more efficient.Extensive experiments are carried out and the results show that our approach outperforms the popular embedding methods on benchmark datasets.