Research On Unsupervised Domain Adaptation Methods In Image Recognition

Traditional image recognition methods assume that training and test ing samples follow the same distribution,however this assumption is often not valid in many practical applications.If the distribution divergence between training samples and testing samples is very large,the performance of the classifiers trained by traditional image recognition methods would be reduced.Domain adaptation is a means to deal with the problem of domain migration in image recognition.The goal of domain adaptation is to overcome the distribution divergence between training samples and testing samples,so as to realize the migration of classifier.Domain adaptation has been widely investigated in the field of image recognition.Firstly,a new algorithm based on kernel subspace alignment(KSA)is proposed.Using the same kernel function,the source samples and the target samples are mapped into high dimensional spaces respectively,and then the source subspace and target subspace are obtained by using principal component analysis(PCA)respectively on the high dimensional source samples and target samples.In the end,an alignment matrix is learned to align the source subspace and the target subspace.As the result,the distribution divergence between the source samples and target samples is minimized.Secondly,a kernel space alignment algorithm(SKSA)which uses the label information in the source domain is proposed.The alignment matrix searched by SKSA can align the source subspace and target subspace and make the samples of same class in the source domain get together,which can improve the final classification accuracy.Finally,based on linear discriminant analysis,we propose a kernel space alignment algorithm(LKSA)which constrains the samples of different classes in the source domain.The alignment matrix searched by LKSA can not only align the source subspace and target subspace,but also make the samples of same class in the source domain get together and samples of different classes in the source domain separate from each other.In the end,LKSA improves the performance of classifier.In order to verify the effectiveness of the methods mentioned above,extensive experiments are performed under the visual object recognition databases(O ffice+Caltech-256),handwriting recognition databases(USPS+MNIST)and face recognition database(PIE).Experimental results il ustrate that the proposed methods are superior to other competitive domain adaptation methods.