Size invariance by dynamic scaling in neural vision systems

Vision systems in complex environments are faced with the problem of analyzing visual information on multiple scales to support segmentation and size invariant object recognition. We propose a system which detects relevant spatial scales in images and constructs an explicit size invariant representation. Further, it provides a means of navigating through scale space in the presence of ambiguous scale information, thus adds a behaving component to image analysis. The architecture is based on biologically realistic neural networks like neural elds and neurons with bandpass receptive eld characteristics.