Regularized denoising latent subspace based linear regression for image classification

This paper proposes a novel method, called Regularized Denoising Latent Subspace based Linear Regression (RDLSLR), for noisy image classification. RDLSLR model divides the traditional subspace learning model into two steps. The first step is adding a denoising latent space between the vision space and label space to obtain clean data by an undercomplete autoencoder and the second step is using another transformation matrix to learn regression target by clean data. In order to further optimize the distribution of data in subspace, an additional Laplacian Regularization is introduced to label space with the help of manifold learning. In addition, epsilon-dragging technique is used in the label space to make the RDLSLR model more discriminative. In the RDLSLR model, data denoising, local structure, and label relaxation are considered at the same time. A joint optimization model is constructed, and an efficient iterative algorithm is designed to solve the proposed model. In order to verify the effectiveness of the RDLSLR model, several experiments involving the face, biometric, object, and deep feature recognition have been conducted. The experimental results show that the proposed RDLSLR model is achieved compared with many state-of-the-art methods.