Automatic generation of morphological opening-closing sequences for texture segmentation

Texture segmentation is an important task in image processing. The objective is to assign the same value to those pixels in an image which belong to the same texture. This is usually called pixel-classification. There exist two fundamental approaches to solve this task. The first and most general approach we call classification oriented by which we mean that the actual labeling is done by some general classifier, i.e. maximum-likelihood. As these classifiers require some feature vector as input, there has to be such a vector for each pixel. By this we not only get some segmentation but we can also identify every texture with respect to the database used to train the classifier. The problem with this approach is that usually the computational costs will be high for each image to be classified, as they are determined primarily by the costs of calculating the feature vector for each pixel. These costs may be reduced by using pruning algorithms, but if a single feature relies on the calculation of several image transformations like in [2] or in [4], the performance gain may be small. In this paper we consider the case that we know which texture combination will appear in an image. A different approach may be used for this problem which we call transformation oriented. By this we mean to find a transformation adapted to a particular texture combination allowing a segmentation of the image by applying some threshold algorithm (i.e. as in [3]) on the transformed image. Our approach therefore consists of three parts: (1) Building a database of texture descriptions, (2) automatic configuration of an appropriate segmentation algorithm for a given texture combination (this is not a particular image, nor any image at all), and (3) segmentation of a particular image. The intention is that the expensive part is Step 1 and has to be done only once for each texture. Step 2 can be performed quickly and has to be done only once for each texture combination in question. Step 3 finally is very cheap and considered to be used with several images containing the same texture combination. This paper is organized as follows: In the next section we give an outline of our approach. This consists of an idealized texture model with an appropriate segmentation strategy. To realize step 1 described above we need a texture description, which is exact for the idealized model and approximative for real textures. It is presented in section 2.1. For step 2 we give two alternative configuration algorithms shown in section 4. step 3 is finally described in section 5. Finally we discuss the results in section 6.