Low Rank Based Discriminative Least Squares Regression with Sparse Autoencoder Processing for Image Classification

Least squares regression is one of the commonly used classification approaches among supervised learning. However, there are several common issues among least square regression at present. One is that it often overlooks the relationship between each sample; the second is that the used zero-one label matrix usually limits its performance in image classification. To overcome the aforementioned problems, we propose a novel method, i.e., low rank based discriminative least square regression with sparse autoencoder processing (LRDLSR_SAE) in this work. This method can utilize the low rank constraint to decrease the distance among samples in the same class. Furthermore, an error term is implemented to relax the zero-one label matrix for exploring a better discriminate transformation matrix. Besides this, we also apply the sparse autoencoder to process raw samples to obtain more suitable features as new input data for classification. These implementations can improve the classification performance compared to other approaches, which is also confirmed by extensive experimental results in this study.