Problems And Approaches For Cross-domain Recognition

Recently,with the popularization of network,video monitoring and the deflationary force of data storage media,the collection and storage of data is becoming easier and easier.At present,most algorithms usually rely on the assumption that the distribution of data does not change with the environment.However,in real-world applications the data acquisition and generation mechanism change with the demand,such as in some stations,airports and other public places,in order to better monitor and analyze potential dangers risk,video equipment is often installed in the cross position.Since the variations of light,angle and occlusion,the collected data(cross-domain data)have different probability distributions.Compared to tradi-tional data,it is more complicated to deal with the data of cross-domain owing to the complex structure difference.How to effectively use the internal distribution data of cross-domain to accurately seek and analyze the information required by users has become a research hot spot in data analysis,among which the most significant achievements are cross-domain recognition and analysis.By reducing the distribution divergence of cross-domain data,knowledge can be reused between domains to enhance the learning performance of the algorithm.With the advent of the era of big data,the billionaire scale label lack of data in the data and statistical dif-ference problem increasingly prominent,how to carry on the knowledge transfer between data with different distributions/reuse will be inevitable,especially for the cross-domain data anal-ysis and applications.Although many methods have been proposed to solve some problems in cross-domain recognition and have achieved remarkable results,especially in the applications of video analysis and image retrieval.However,due to the complexity of dealing with different distributed data,there are still problems to be explored and solved in cross-domain learning(as shown in figure 1).Since these problems overlap each other,it is very difficult to solve these problems.To this end,this paper proposes a series of cross-domain recognition learning theory and algorithm framework.Research results of this paper can not only further expand to more areas,such as multi-view and multi-modal learning,and can be applied in the prac-tical object recognition(such as person re-identification under cross-camera),video analysis(such as video retrieval under cross-view),speech recognition(such as dialect and mandarin cross recognition),biometrics(such as multiple features/cross feature joint recognition)and multi-model/view image retrieval,and so on.Specifically,the main works of this paper are as follows:(1)Aiming at the shortcoming that traditional linear discriminant analysis method cannot effectively process data with different distribution,a transferrable linear discriminant analysis method based on matrix low-rank transformation is proposed,which further extends the tradi-tional linear discriminant analysis method to transferrable environment.The main idea of this method is based on that the data from the same subspace has a low rank structure.Therefore,under the of framework of big marginal,we use the low-rank transformation of matrix as the main criterion to propose both the local and global low-rank transformations for transforming the data from different domains into a subspace by using transformation matrix.Moreover,we use the difference of rank of matrix of transformed data to make the data sharing the same label have the low-rank structure,which can effectively reduce the distribution divergence of the data from different domain but have the same label.Therefore,the data from different domains but with the same label can be aligned together well,enabling domain knowledge to be reused.(2)This paper proposes a new method of data reconstruction,i.e.,exploit both of the self-reconstruction and cross-reconstruction of projective data to learn a single dictionary.Different from the traditional dictionary learning based cross-domain recognition method that matches the data from different domains by learning two different dictionaries,the method proposed in this paper guarantees both the reconstruction ability of and transferability of dictionary respec-tively by using the self-reconstruction and cross-reconstruction of projective data.In order to improve the discriminative ability of dictionary,this paper proposes a more flexible label con-sistency item.By introducing a non-negative matrix to fit the labels of data from cross-domain as much as possible and learning a non-negative matrix of labels,the label consistency item can avoid the semantic gap caused by the distribution divergence of data,so as to effectively use the label information to learn a discriminative reconstruction coefficient matrix.(3)This paper proposes a new classifier learning method,this method first analyzes the traditional methods that reduces the conditional probability distribution and marginal distribu-tion of the data from cross-domain cannot fully makes the data from different domains overlap together.Therefore,this paper further proposes to train two different classifiers on the new feature representation of data from diferent domains.Moreover,we require that these two classifiers are dynamically approximated during learning by using the strategy of dynamic approximation.We finally use the mean classifier to fuse the classifiers.The theory and ex-periment prove that the strategy of dynamic classifier approximation can learn a classifier with strong transferrable ability.(4)In order to improve the efficiency of cross-domain recognition,a latent elastic network transfer learning method is proposed.In this method,1)all the original data are projected into a latent subspace to reduce the distribution divergence of data from cross-domain;2)the low-rank constraint based matrix elastic network regression technique is introduced to transform the feature of latent data into the label space,and at the same time,the low-rank is used to preserve the class structure of data.Theoretical and experimental results show that this method can overlap the data from different domains but with the same label which effectively reduce the distribution divergence of data and greatly improve the recognition rate of cross-domain.