Image Texture Detection and Extraction based on Entropy and Eigenvectors of growing seeds

Abstract

Texture determination, segmentation and extraction is always a field of research and development at image processing and computer vision science. Since there are no clear and general definitions of texture, all known is that texture depends on local pixels gray levels with respect to their spatial position. There has been quite some efforts put on this topic but nearly all of the presented techniques are either very complicated and requiring too much computations, not covering all textures presented in scene or need supervision and databases for training. At this paper a new approach is presented which is completely unsupervised, rapid and adoptable to increase efficiency and accuracy. It also needs no training sequences or databases and can be utilized to identify objects as texture components. The presented approach acquires texture identification features with respect to image areas entropies and basic neighborhoods eigenvectors. After re-defining these parameters for all possible pixel divisions at the image segments with optimal size with respect to the image geometries, certain or random points are chosen as texture seeds and grown up until new additions vary the segment’s entropy dramatically. After this the set of largest grown and separately located areas are introduced in sequence as discovered texture segments. Furthermore different variations of this method can be developed for different functionality properties.