Robust double relaxed regression for image classification

Abstract

In this paper, a Robust Double Relaxed Regression (RDRR) is proposed for image classification. The core idea of RDRR is to alleviate the contradiction between the increase of model flexibility and the aggravation of the overfitting problem through noise reduction. Three strategies are used in RDRR to improve flexibility and avoid overfitting: 1) The ϵ-drag technology is introduced to relax the regression target; 2) The graph regularization term containing two matrices is introduced to further improve the flexibility; 3) The noise in the data is separated by using the idea of sparse matrix decomposition and autoencoder to alleviate the overfitting problem caused by (1) and (2). To verify the effectiveness of the proposed method, comparative experiments are designed on many different fields and types of datasets. The projection matrix obtained by RDRR demonstrated robust and superior classification accuracy compared to the current state-of-the-art methods.