Computer Vision Winter Workshop 2006, Ondˇ rej Chum, Vojtˇ ech Franc (eds.) Telˇ c, Czech Republic, February 6–8 Czech Pattern Recognition Society Texture segmentation through salient texture patches Lech Szumilas 1 , Branislav Miˇ cuˇ s´ ık 1 , and Allan Hanbury 1 1 Vienna University of Technology Pattern Recognition and Image Processing Group Favoritenstr. 9/1832, A-1040 Vienna AUSTRIA lech@prip.tuwien.ac.at, micusik@prip.tuwien.ac.at, hanbury@prip.tuwien.ac.at Abstract A novel approach to automatic, hierarchical texture de- tection in single images as a step towards image understand- ing is presented. First, the proposed method searches for al- ternating color patterns through hierarchical clustering of color pairs from adjacent, symmetrical image segments to localize salient regions in terms of color and texture. Sec- ond, the salient regions are fed as seeds to an image seg- mentation method based on min-cut/max-flow in the graph to localize the texture boundaries more accurately. The final result is a hierarchy of potential textured regions in the im- age useful for further object/texture recognition step. This work gives a proof of concept that the stable salient texture regions supported by a semi-automatic segmentation algo- rithm may provide fully automatic image segmentation into uniform color and/or texture regions. The results are pre- sented on some images of natural scenes from the Berkeley database. 1 Introduction Texture detection and classification play an important role in many image analysis tasks. Detection of texture boundaries is crucial for general image segmentation algorithms, while texture classification can provide extremely useful informa- tion for object recognition methods. The term “texture” typically describes the presence of some regularity in a continuous image region, which may manifest itself as a spatially repeating color pattern or shape, but it is not defined how regular it must be. Several segmen- tation algorithms, either automatic or semi-automatic, exist which are capable of dividing an image into a uniform color and/or texture regions [4, 11, 8, 9, 7, 10, 12]. The primary problem these methods face is related to texture discrimi- nation. The texture similarity cannot be precisely defined as it depends on the particular application – if we want to segment out the whole tiger from an image, then our tex- ture similarity criteria will have to be different than if we want to segment out its legs, tail and head separately. This leads us to conclusion that the definition of a texture is task- dependent, which means that detection of texture should be knowledge based. Current segmentation methods do not al- low the selection of a knowledge database for texture detec- tion. In this work we attempt to combine a separate texture (a) (b) (c) Figure 1: Automatic texture detection from single images. (a) An input image. (b) A mask provided by the proposed salient texture patches detection method. (c) Final result after the seed segmenta- tion method using the mask from (b). detection method with the seed segmentation method [8] to achieve a fully automated image segmentation, see Figure 1, further useful for knowledge based texture classification. The method Feature Co-occurrence Texture Detector (FCTD) proposed in this paper detects textures at various “regularity levels” and produces a hierarchy of textures. The advantage of such an approach is the ability to detect less or more regular patterns automatically and then to make a knowledge based selection. Since the texture classification is not yet implemented, we devised a simple method to se- lect the most regular textures from the hierarchy of detected ones, and use it as a texture marker for an image segmenta- tion algorithm. This work is therefore a proof of concept of whether the combination of both methods works well. A very common pattern found in textures is the alterna- tion of two or more colors or luminance levels like for exam- ple the patterns covering tiger and zebra skins or the ripples on water. Therefore FCTD at the moment uses only color features to do a coarse texture detection. We assume that the texture consists of spatially mixed patches of uniform color, called texture elements. The general idea is to segment the image into small segments with a relatively uniform color and then find alternating color patterns among these seg- ments. The pattern we are looking for is a group of similar 1