Image 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. The coarse measurements for detecting these properties cannot be obtained by simple geometric averaging, because they would often average over properties of neighboring segments. A multiscale method that overcomes these difficulties will be presented. The algorithm consists of an adaptive process in which pixels are recursively aggregated into progressively larger -scale aggregates, based on coherence of accumulated properties such as: intensity, texture, direction, boundary direction and curvature, shapes of sub-elements, etc. This bottom-up aggregation is coupled with top-down transfers of aggregation directives, increasingly based at coarser levels on learnt information. The algorithm costs a very low number of operations per pixel. Numerous experimental
results demonstrates its high performance and versatility. In addition, the algorithm produces a novel description of the image as a hierarchy of segments. Several applications under development and a related work on shape recognition will be mentioned.
This is a joint work with Ronen Basri, Meirav Galun and Eitan Sharon.
Presentation Files (Zip Archive)