Texture Learning by Fractal Compression

Abstract

This paper proposes a texture learning method based on fractal compression. This type of compression is efficient for extracting self-similarities in an image. Square blocks are studied and similarities between them highlighted. This allows establishing a score and thus a rating of the blocks. Selecting the blocks that best encode the biggest part of the image or a texture leads to a database of the most representative ones. The recognition step consists in labeling blocks and pixels of a new image. The blocks of the new image are matched with the ones of the different texture databases. As an application, we used our method to recognize vegetation and buildings on aerial images.