Proximity measure image based region merging for texture segmentation through Gabor filtering and watershed transform

In this paper, an unsupervised texture segmentation scheme is proposed, based on region merging which is carried out on the proximity measure image. A bank of Gabor filters are first applied to the texture image to be segmented, and the proximity measure image, as the feature image, is then constructed by fusing all individual proximity measure images in which pixel intensity reflects the statistical similarity of local histograms in a given neighborhood. Based on the feature image, the task of texture segmentation is thus casted as a problem of region merging and edge detection. To efficiently classify different textures, precisely detect and locate the boundaries of distinct textures, the watershed transform is applied to the feature image for initial segmentation. A region merging procedure is then realized to deal with the over-segmentations resulted from the watershed transform, by iteratively grouping adjacent regions. To demonstrate the effectiveness of the proposed scheme, experiments are carried out on both texture composites and real-world images. Results show that the proposed scheme performs well, in terms of both segmentation accuracy and precision in locating boundaries between distinct textures.