Research On Recognition And Retrieval Algorithms Based On Subspace Learning

In the past few years,with the rapid development of Internet technologies and storage technologies,an increasing number of multimedia data have appeared and been used by a large number of users.Among them,the image data as the most common data have obtained a great deal of attention from researchers.And the image recognition technology has emerged and developed rapidly.However,many challenges exist and hinder the further development of image recognition technology,such as small sample problems,occlusion and noise problems.In addition,crossmedia retrieval technology as a product of the rapid development of multimedia technology,has gained wide attention in the era of big data.Nevertheless,a basic problem of semantic gap has not been solved with an efficient method.According to the problems in image recognition and crossmedia retrieval,this paper proposes two methods,and the accuracy and reliability of the proposed methods are analyzed and verified on the main datasets.The main work and innovations of this article summarized as follows:1.To solve the problem of small sample training in image recognition,this paper applies the collaborative representation to small sample training,and proposes a two-step small sample face recognition algorithm(TSCR)based on collaborative representation.The method utilizes unlabeled data to reconstruct annotation data and obtains its corresponding synergistic representation coefficient.At the same time,the sample corresponding to the largest value of the coefficient of cooperation is added to the original labeled data set,until half of the unlabeled data set are assigned to the labeled data set.Finally,the remaining unlabeled data are classified using a collaborative representation.The experimental results show that TSCR is efficient.2.In order to settle the local judgment of image in image recognition,this paper fully considers the relationship between unlabeled data and its neighborhood.A weighted local collaborative representation algorithm(WLCRC)based on sparse subspace is developed.Firstly,WLCRC learns a strong correlation dictionary,then it combines weighted collaborative representation with linear regression algorithm to optimize unlabeled data.As a result,WLCRC considers the local neighbors of data and the semantics of annotation data in the reconstruction process of annotation data,therefore,the approach obtains good recognition rates.3.In view of the problem of semantic gap in cross-media retrieval,this paper proposes a semisupervised modal correlation cross-media inspection method(SMDCR).This method makes full use of the relevance between the annotation data semantics and the unlabeled data,and proposes an adaptive learning projection optimization method,which greatly improves the accuracy of data retrieval.Experiment results on several popular datasets demonstrate its effectiveness.4.This paper proposes a semi-supervised distance consistent cross-media retrieval method(SSDC)on the premise of the relationship between labeled data semantics and unlabeled data.This method uses the existing annotation dataset to construct the virtual unlabeled retrieval set.And the relationship of distance between label data and unlabeled data help to achieve efficient cross-media retrieval.The proposed method obtains good results on several related datasets.