A neural architecture for visual recognition of general stationary objects by machines

A machine vision architecture of neural network modules for learning and recognizing objects is described. The modules make up a location channel and a classification channel. The location channel rapidly finds objects in the field of view using feedforward and feedback paths and centers the classification channel on a portion of the field of view. The classification channel contains successive stages of modules capable of learning and classifying the objects in the selected part of the field of view. Learning is done by presenting examples to an unsupervised classifier module that associates high-dimensional features with a few self-defined categories. A supervised classifier module then associates the self-defined categories with externally defined names. An ongoing study with a software testbed using images of human cytology specimens in their natural backgrounds is briefly summarized