Feature generation from digital images using pseudo-fractal algorithm and its four modifications

The main aim of the paper is to present the authors' original method of feature generation from digital images and to report on a comparison of five various algorithms, which implemented that method. The algorithms are based on an idea by the same authors', which consists in producing a quantitative description of similarity intensity between various parts of an image in various scales. To develop it the algorithms take advantage of fractal coding based on an Iterated Function System. Therefore, the generated features can rightly be called similarity features. In this paper we show that similarity features, when combined with other well known ones, can improve recognition results in some image classification tasks. After presenting how the algorithm works, we compare their properties and report the classification results obtained in two different pattern recognition experiments. Moreover, the paper contains a discussion of the obtained results, and of possible future applications of the similarity features.