Deep Sparse Representation Learning for Multi-class Image Classification

This paper presents a novel deep sparse representation learning for multi-class image classification (DSRLMCC). In our proposed DSRLMCC, we use dictionary learning for sparse representation to train the deep convolutional layers to work as coding layers. The dictionary-learning algorithm uses input training data to learn an exhaustive dictionary and sparse representation. The deep sparse coding layer enforces locality constraints for activated dictionary bases to achieve high convergence. With the second deep learning layer, fine-grained components are learned, which, in turn, are shared by all atoms in the input dictionary; thus, a low-level representation of the dictionary atoms can be learned that is more informative and discriminatory. Comparing the proposed model with several prominent dictionary learning strategies and deep learning models, we found that the proposed method outperforms them. We have executed the proposed method on three prominent datasets, and the results are satisfactory.