Segmentation of tomographic data by hierarchical watershed transform

The aim of the proposed watershed based image segmentation technique is to split images into spatially homogeneous regions, which can be further processed by different image analysis tools. The advantage of such approach, in comparison to pixel oriented processing, is its lower sensitivity to superimposed noise due to averaging of regions properties over their area. The watershed segmentation technique is based on interpretation of an image as a topographic relief and on simulation of flow of water along steepest descent paths called downstreams. Thus, for each local minimum of the image, a drainage region is defined, which, if computed for a gradient image, represents an area with approximately constant properties. The segmentation technique is further extended for multi-scale image analysis by means of Gaussian smoothing. The aim of smoothing is to suppress image details that are smaller than standard deviation of the Gaussian. However, smoothing results not only in the desired increase of region size, but it also affects position of region boundaries, at least for larger standard deviations of the Gaussian filter. Therefore a new technique is proposed, based on region hierarchies, which enables to transfer region contours with precise position from the levels with low smoothing to levels with higher smoothing. Thus, segmentation of an image into large regions, but with exact contours, is obtained.