STOCHASTIC MODEL BASED TECHNIQUES FOR CLASSIFICATION AND SEGMENTATION OF TEXTURES

ALIREZA KHOTANZAD, Purdue University

Abstract

This study is concerned with the classification and segmentation of textures. The main emphasis is on the statistical structure of the texture rather than its visual structure. The texture is viewed as a two-dimensional stochastic spatial process, i.e. a random field. Hence, spatial interaction models which capture the spatial structure of random fields are used for texture modelling. A class of such models known as simultaneous autoregressive (SAR) models is used to represent the texture. The estimated parameters of a SAR model fitted to an image are suggested as features for its texture. Results of supervised classification experiments using these features indicate that they are powerful. However these features are rotation variant, i.e. change if the texture is rotated with respect to the camera. In order to get a rotation invariant feature set, a new class of spatial interaction models called circular autoregressive (CAR) is developed. In a CAR model, any pixel value is written as a finite sum of intensity values at locations on a circular neighborhood around it and a noise sequence. The parameters of this model are rotation invariant. The estimated CAR parameters plus other rotation invariant functions defined on SAR parameters are used as textural features. These features have direct interpretation in terms of visual attributes of texture. The strong discriminating power of the proposed features is shown by using them in supervised classification experiments involving differently oriented test and training samples from each class. A new technique is presented which segments an image into regions of similar texture when no prior information about the textures is available. The image is scanned by a small window and a number of textural features extracted from each window. Different pairs of features corresponding to various windows are plotted. Some of these 2-d plots contain distinct clusters corresponding to different texture types. Only those two features having the highest clustering power are used. Segmentation is done by defining decision boundaries on the clusters and mapping them back into the spatial domain. Several experimental results are presented.

Degree

Ph.D.

Subject Area

Electrical engineering

COinS