Alternate formulation for transform learning

Abstract

Dictionary learning has been used to solve inverse problems in imaging and as an unsupervised feature extraction tool in vision. The main disadvantage of dictionary learning for applications in vision is the relatively long feature extraction time during testing; owing to the requirement of solving an iterative optimization problem (l0-minimization). The newly developed analysis framework of transform learning does not suffer from this shortcoming; feature extraction only requires a matrix vector multiplication. This work proposes an alternate formulation for transform learning that improves the accuracy even further. Experiments on benchmark databases show that our proposed transform learning yields results better than dictionary learning, autoencoder (AE) and restricted Boltzmann machine (RBM). The feature extraction time is fast as AE and RBM.