Deep Sparse Informative Transfer SoftMax for Cross-Domain Image Classification

In many real applications, it is often encountered that the models trained on source domain cannot fit the related target images very well, due to the variants and changes of the imaging background, lighting of environment, viewpoints and so forth. Therefore cross-domain image classification becomes a very interesting research problem. Lots of research efforts have been conducted on this problem, where many of them focus on exploring the cross-domain image features. Recently transfer learning based methods become the main stream. In this paper, we present a novel transfer SoftMax model called Sparse Informative Transfer SoftMax (SITS) to deal with the problem of cross-domain image classification. SITS is a flexible classification framework. Specifically, the principle eigenvectors of the target domain feature space are introduced into our objective function, hence the informative features of the target domain are exploited in the process of the model training. The sparse regularization for feature selection and the SoftMax classification are also employed in our framework. On this basis, we developed Deep SITS network to efficiently learn informative transfer model and enhance the transferable ability of deep neural network. Extensive experiments are conducted on several commonly used benchmarks. The experimental results show that comparing with the state-of-the-art methods, our method achieves the best performance.