Supervised Image Classification with Self-Paced Regularization

Abstract

In this paper, we present a new scheme for image classification that is robust to samples noises. The proposed scheme depicts a novel sparse classification model with self-paced learning mechanism. First, inspired by the outstanding performance of curriculum learning, we integrate the idea of self-paced learning into supervised class-specific dictionary learning to select appropriate training samples. Secondly, we design a novel sparse representation model associated with self-paced learning regularization, which employs locally linear reconstruction to improve the accuracy of the classifier and exploit the manifold structure of data. By using the designed model, a classification scheme integrating self-paced learning is proposed to exploit more discriminative image information. The experimental results on two typical datasets indicate that our constructed model achieves the competitive performance when compared with the state-of-the-art methods.