Multilayer dynamic link networks for establishing image point correspondences and visual object recognition

The major tasks for automatic object recognition are segmentation of the image and solving the correspondence problem, i.e. reliably finding the points in the image that belong to points in a given model. Once these correspondences are found, the local similarities can be used to assign one model out of a set of known ones to the image. This work defines a suitable representation for models and images based on a multiresolution transform with Gabor wavelets. The properties of such transforms are discussed in detail. Then a neural network with dynamic links and short-term activity correlations is presented that estimates these correspondences in several layers coarse-to-fine. It is formalized into a nonlinear dynamical system. Simulations show its capabilities that extend earlier systems by background invariance and faster convergence. Finally, the central procedures of the network are put into an algorithmic form, which allows fast implementation on conventional hardware and uses the correspondences for the successful recognition of human faces out of a gallery of 83 independent of their hairstyle. This demonstrates the potential for the recognition of objects independently of the background, which was not possible with earlier systems.