Image Segmentation by Texture Using Pyramid Node Linking.

Abstract

In a 'pyramid' of successively reduced-resolution versions of an image, by linking nodes representing image blocks to nodes representing nearby larger blocks that most closely resemble them, we can construct trees (defined by the links) representing homogeneous parts of the input image. In this paper, we apply this approach to segmenting an image on the basis of texture. We start from an initial decomposition of the image into small blocks (e.g., 8 by 8); compute a textural property for each block, yielding an array of property values; build a 'pyramid' of reduced-resolution versions of this array; and apply the node linking process to this pyramid. The resulting trees define a segmentation of the original image into unions of the small blocks. This segmentation is similar to that obtained by minimum-error thresholding of the textural property values. Substantially better results are obtained when this 'bottom-up' block linking process is preceded by a 'top-down' process in which large homogeneous blocks are linked to all of their subblocks; the bottom-up linking is then used only for the blocks that were not linked by the top-down process. (Author)

Ähnliche Arbeiten