Determination of the grey level ranks for the segmentation of textured images

In many cases, an image is made of various types of texture. Such diversity can be usefully exploited to get a suitable decomposition of the image into typical classes and then, identify its constitutive elements. Among the approaches used to segment textured images, those based on the statistical analysis of the neighborhood of each pixel seem to be most efficient. In this paper, this kind of method is implemented, which consists in coding the pixels surrounding each point of the image by taking into account the path traveling through them and their grey levels. The rank vectors namely the codes obtained for all the possible paths, are then classified using the K-means algorithm. Considering the 24rank vectors, this method is tested on different images of the Brodatz album and compared with Laws filters and GLCM. It yields a satisfactory reproduction of the image contours and a classification ratio exceeding 98 %. The 24 rank based-method is also applied to meteorological images collected by Meteosat over Europe and North Africa during December 1994. It is found that these images can be segmented into eight typical classes assignable to the soils, the seas and the clouds observed in the regions under study.