Object recognition by a self-organizing neural network which grows adaptively

Abstract

We describe a new type or neural network for object recognition which we call a Cresceptron. The term “Cresceptron” was coined from Latin cresco (grow) and perceptio (perception). The primary objective of the Cresceptron framework is to automatically handle manually intractable tasks: such as constructing a network that can recognize many objects from real world images. The Cresceptron uses a hierarchical structure, and the network adaptively and incrementally grows through learning. For recognition, the network is made largely translationally invariant by using the same neuron at all the positions of each neural plane. Scale invariance is achieved through a multi-resolution representation with the framework of visual attention. Limited orientational invariance is obtained by variation tolerance. Complete orientational invariance is not sought here since the recognition should report also thc orientation. It is interesting to note that psychophysical studies have demonstrated that the human vision system does not have perfect invariance in either translation, scale, or orientation.