Classification of process data and images by human assisted fuzzy similarity analysis

In this paper an incremental classification scheme for large data sets and images is proposed in the form of a two-stage computation scheme. First, information compression of the original data set or pixels is performed by a modification of the Neural-Gas unsupervised learning algorithms. Then two features are extracted from the obtained compressed information model, namely the center-of-gravity of the model and its size, which are further used in a fuzzy inference procedure for similarity analysis. The tuning of the membership functions parameters in the procedure for fuzzy similarity analysis is also discussed in the paper by using a modified particle swarm optimization algorithm that takes into account the predefined human preferences. Finally, the applicability of the proposed classification scheme is illustrated on a test example of 16 images.

1. INTRODUCTION

Similarity analysis, performed over a large amount of images or large data sets is very important step in the procedure for classification of different types of pictorial or process information. This is a very specific area of activity, where in many cases the experienced human performs better and produces more plausible solutions than the currently available computerized systems. One reason for this is the complexity and the vagueness in the definition of the problem. Obtaining a “better” and “more plausible” solution to the problem of similarity is a key factor for success in many applications such as quick search through a large amount of image or process data information, and its proper sorting and classification. The results of this similarity analysis and classification are often used for a proper fault or medical diagnosis and for discovering different abnormalities in the observed systems.