Dictionary Learning by Alternating Direction Method of Multipliers

• Main Authors: Wolfgang Xaverius Dorojatun Jalma Nuswantara, Jalma
• Other Authors: Tsung-Hui Chang
• Format: Others
• Language: en_US
• Published: 2015
• Online Access: http://ndltd.ncl.edu.tw/handle/45259232160673563018

碩士 === 國立臺灣科技大學 === 電子工程系 === 103 === This thesis investigates the application of the alternating direction method of multipliers (ADMM) to dictionary learning for noise removal problem. Dictionary learning (DL) is the process of acquiring dictionary that can yield sparse representation of desired signal by learning from training sig-nal, instead of using prespeci ed transformation basis. The prior arts of the dictionary learning algorithms include the method of optimal direction (MOD) and K-SVD. These methods have main drawback in computational complexity.By contrast, the ADMM, which we use to transform the complex problem, into simple update steps, yields lower computational complexity. We further proposed on inexact ADMM that can reduce the computational time, for scenarios with large dictionary size. Simulation results show that the proposed methods can successfully train the dictionary and yields promising performance for the image noise removal problem. In particular, the pro-posed ADMM method can be around 30 times faster than K-SVD, while inexact ADMM method can further reduce the computational time around 40 %.