Optical and Sonar Image Classification: Wavelet Packet Transform vs Fourier Transform

Abstract

To develop a noise-insensitive texture classification algorithm for both optical and underwater sidescan sonar images, we study the multichannel texture classification algorithm that uses the wavelet packet transform and Fourier transform. The approach uses a multilevel dominant eigenvector estimation algorithm and statistical distance measures to combine and select frequency channel features of greater discriminatory power. Consistently better performance of the higher level wavelet packet decompositions over those of lower levels suggests that the Fourier transform features, which may be considered as one of the highest possible levels of multichannel decomposition, may contain more texture information for classification than the wavelet transform features. Classification performance comparisons using a set of sixteen Vistex texture images with several level of white noise added and two sets of sidescan sonar images support this conclusion. The new dominant Fourier transform features are also shown to perform much better than the traditional power spectrum method.