Dictionary learning in convolutional sparse representation

Abstract

In this paper, we present a new convolution form based on dictionary learning and study its sparse version based on Convolution Dictionary Learning (CDL). An effective algorithm for learning sparse convolution features is proposed by combining Alternating Direction Multiplier Method (ADMM) and Fast Iterative Shrinkage Threshold Algorithm (FISTA). Through numerical experiments, we show that the proposed algorithm can not only lead to faster convergence speed, but also produce better sparse features.