Similarity analysis of images based on information granulation and fuzzy decision

This paper proposes a computational scheme for fuzzy similarity analysis and classification of images that uses first an information granulation procedure followed by a subsequent fuzzy decision procedure. A special new version of the growing unsupervised learning algorithm is introduced in the paper for information granulation. It reduces the original ldquoraw datardquo (the RGB pixels) of the image to a considerably smaller number of information granules (neurons). After that two features are extracted from each image, as follows: the center-of-gravity and the weighted average size of the image. These features are further used as inputs of a special fuzzy inference procedure that computes numerically the similarity degree for a given pair if images. Finally, a sorting procedure with a predefined threshold is used to obtain the classification results for all available images. The proposed similarity and classification scheme is illustrated on the example of 18 images of flowers. It is also discussed in the paper that the appropriate tuning of the parameters of the fuzzy inference procedure is quite important for obtaining plausible, humanlike results Therefore a simple empirical process for selection of these parameters is also suggested in the paper.

I. Introduction

The similarity analysis of large number of images is very important step in the overall procedure for unsupervised classification [1], [2] of different types of pictorial information. This is a very specific and important task, where humans are still better in performance than the currently available computerized systems. Among the different reasons for such results are the obvious complexity and the vagueness (subjective nature) of the problem itself, as well as the various possible criteria that could be used for the similarity analysis. Nevertheless a “good” (i.e. true and plausible) solution to the problem similarity evaluation is a key factor for a success in many applications such as fast visual search through a large amount of image information, for a proper sorting and classification.