Research Of Transfer Learning In Image Classification Based On Sparse Coding

In computer vision fields,image representation is important for processing and understanding images.As a powerful tool for finding high level semantics in visual images,sparse coding can represent images using only a small number of active coefficients,and it shows good spatial locality and de-redundancy.However,when training and testing samples are sampled from different distributions,the dictionary learned from training samples cannot effectively encode test samples,and this coding discrepancy may seriously degrades the performance of classification.At the same time,the existing sparse coding transfer learning algorithms also have under-adaption problem between different domains.For these problems,in this paper from the aspects of dictionary and coding difference in the sparse coding model,and combining feature correlations between different domains,we respectively construct a feature transfer classification model based on dictionary alignment and joint distribution adaption strategies in sparse coding.The main research contents of this paper are as follows:Firstly,for dictionary discrepancy problem,a transfer sparse coding based on dictionary domain adaption(TSC-DDA)algorithm is proposed.First,two sets of sparse coding models are constructed in the source domain and target domain;Next,the dictionary alignment strategies is introduced into sparse coding model in order to make source dictionary and target dictionary more closer to obtain domain adaptive dictionary.Meanwhile,the regularization term is used to replace dictionary constraint term,so it can convert to an unconstrained optimization problem and it can be easily solved.Then,through dictionary mapping relation,the difference of coding features between source and target domains in shared space is minimized.Finally,the proposed feature transfer algorithm is combined with support vector machine(SVM)to build a sparse feature transfer classification model.Secondly,for under-adaption problem,a transfer sparse coding based on joint distribution adaption and graph regularization(TSC-JDAG)algorithm is proposed.First,the potential geometric features in source and target domains are extracted by constructing a graph structures;Next,joint distribution adaption is introduced to sparse coding model in order to reduce the marginal and conditional distribution discrepancy between source and target domains,and weight factor is used to adjust their contribution on learning target task.Then,by introducing graph regularization term into sparse coding model,geometric features are embedded in sparse features to further reduce distribution discrepancy between different domains.Finally,the proposed feature transfer algorithm is combined with SVM to build a sparse feature transfer classification model.Experimental results on multiple group cross-domain image classification task show that compared with other comparative methods,the proposed algorithms can effectively solve the cross-domain image classification problems based on sparse coding and obtain better classification accuracy.