Design of Steerable Wavelets Using Learning Techniques

Xinchao Wang

Semester Master Project

Microengineering section, EPFL

January 2010

Abstract

In this project, we present a novel design of steerable wavelets using machine learning techniques. We show that by applying machine learning techniques on the coefficients output from Riesz-Wavelet transform, we can display steerable filters in an application-driven manner. The design of steerable wavelets enjoys a number of merits: 1) from wavelet aspect, we can obtain some customized filters in a query-driven manner, so the performance can be significantly improved; 2) from machine learning and pattern recognition aspect, we can benefit from the elegant properties of Riesz-Wavelet transform, e.g. protestation/translation invariance. We focus on two applications in image processing: denoising and classification. We apply unsupervised learning techniques for denoising while apply both supervised and unsupervised learning for classification. In experiments, we show customized steerable filters and classification rate. Experimental results on two texture datasets confirm the validity of our proposed design.