Symmetrical patterns in natural images

Abstract

One of the approaches to image analysis is based on the use of visual alphabets comprising the compact patterns characteristic of the images of a particular class. Elements of these alphabets can be either obtained from a theoretical model describing the basic image structures or extracted directly from the training sets. Symmetric patterns characteristic of the images of outdoor scenes were investigated in this study. The choice of this pattern class is driven by the fact that the objects of interest usually have at least local symmetry. Furthermore, selection of only symmetric patterns allows most fragments of random textures that in most cases do not have symmetry to be dismissed. The image set used in this study comprised 829 monochromatic photographs showing natural objects and textures at different scales. Fragments with diameters of 13 pixels having 1 of 9 symmetry types were selected from this set. The obtained sets of symmetric fragments were processed using a density-based spatial clustering of applications with noise (DBSCAN) cluster analysis algorithm, which allows clusters of arbitrary shape to be analyzed. Fifteen pattern types were obtained; five of them have not previously been used as elements of visual alphabets. Most of the obtained patterns have reflection symmetry, which agrees with the results of psychophysical studies, according to which reflection symmetry is identified faster than other symmetry types. The obtained data can be used for the formation of elements of visual alphabets as well as for synthesizing the stimuli for the experiments in the field of vision physiology.