Image segmentation is one of the most basic tasks for both computer and biological vision systems, and is a prerequisite for higher-level processes from motion detection to object recognition, and can also be used for picture denoising, compactification, miniaturization, etc. Segmentation is difficult because objects may differ from their background by any of a variety of properties that can be observed in some, but often not all scales. Our algorithm of Segmentation by Weighted Aggregation (SWA) consists of an adaptive process in which pixels are recursively aggregated into increasingly larger scale aggregates of coherent properties: The larger the scale the longer the list of internal properties that are accumulated for each aggregate and used to calculate its coherence with neighboring aggregates. The list may include: average intensity and its standard deviation; averages of standard deviations at various finer levels; length, width, orientation and other shape moments, and statistics of these properties at finer levels; boundary directions; and so on. This collection of properties easily captures coherent textures, fiber continuation and more.
Since the computation is recursive and mostly done at coarser levels, the algorithm costs only several dozen operations per pixel, and is highly parallelizable.
Experimental results demonstrate a dramatic improvement over current state-of-the-art methods. Moreover, the hierarchical segmentation produced in this way is in a form directly usable by learning/recognition on systems, since each segment emerges equipped with a vector of numbers representing textures, standardized shapes, sub-segments with their own vectors of numbers and other identifying features.
Back to Multiscale Geometric Analysis: Theory, Tools, and Applications
This is a joint work with Eitan Sharon, Meirav Galun and Ronen Basri.