Research On Fast Sparse Coder And Its Applications

Sparse coding algorithm learns to get a group of over complete bases, taking advantage of the redundancy of natural images. The over complete bases are usually called dictionary. The purpose of sparse representation is representing the signals by choosing a set of bases from the dictionary, and learning to obtain corresponding coefficients. In spite of the widely use in signal processing, sparse representation loses some details especially in case of some operations on complex texture. In addition, the high computation complexity hinders its real time application. The key of the paper is to explore the research on fast sparse encoder and its application in image denoising.Based on sparse coding, the paper improves two aspects of it.(1) The paper develops a Fast approximately Sparse Coding Network (FSCN), combining the traditional neural network and sparse coding. Neural network is so fast while predicting the feature of input, and sparse coding is very effective. Our fast sparse encoder is applied to image denoisng, with the both of advantages. So sparse coding is much faster than before. The results of FSCN are acceptable and even better in image denoising.(2) Deep learning is an emerging topic in artificial intelligence. Thanks to deep learning, a deep sparse encoder (DSC) is stdudied in this paper. We apply it to digital handwriting recognizing, and image denoising.