Image description by Hierarchical Prioritised Fuzzy Systems

The inherently hierarchical problem of evaluating the complexity of an image interpretation is of relevance in both computer science and cognitive psychology. In this paper a new method of rule generation for the hierarchical prioritized fuzzy system, HPFS, is proposed, which overcomes the problem of lack of interpretability of most of the traditional fuzzy systems in modelling image. A hierarchical structure of different fuzzy systems is presented in this work based on prioritising, through the use of a relevance measure of a fuzzy system. For this hierarchical structure we propose a new algorithm to be used in two learning phases: structure building and parametric identification. This new fuzzy modelling technique automatically generates and tunes the sets of fuzzy rules in the hierarchical prioritized fuzzy structure. The learning strategy performs the division of the learning data among the various levels of the hierarchical structure. The effectiveness of the proposed method is tested on cross image recognition.

I. Introduction

In image interpretation, computer vision and structural recognition, the management of imperfect information and of imprecision constitutes a key point. Fuzzy modelling is one of the currently used techniques that exhibits nice features to represent spatial imprecision at different levels, imprecision in knowledge and knowledge representation by fuzzy sets, and which provides powerful tools for fusion, decision-making and reasoning.