Invariant Object Recognition Using Neural Network Ensemble on the CM

ABSI'RACI' This paper concerns machine recognition of objects from their images, where the recognition is invariant to scale, translation, and rotation. A neural network used for recognizing input objects is four layer backpropagation network and a cluster of interconnected units spanning four layers of each network forms a functional block called a column. The 90" rotation invariance has been obtained b a specific interconr nection scheme between the units in the irst and second layers in the network. The scale invariance has been achieved by superimposing many columns of different spatial resolutions at a specific position on the input visual field. The translation invariance has been obtained by overlapping two adjacent columns of the same scale. The generalization ability of object recognition system has been improved by using neural network ensemble supported by a consensus voting scheme. The neural network ensemble has been implemented on the Connection Machine, in such a way that (i) all training sam les are presented simultaneously, and (ii) multiple networ ! ! s are implemented simultaneously. A set of rigorous experimentations using 2-D key images has been performed to demonstrate the usefulness of neural network ensemble for invariant object recognition in terms of computation time, convergence characteristics, rotation invariance, size invariance, translation invariance, and the effect of noise.