Two-dimensional object matching using Kohonen maps

This research aims to develop a neural-network-oriented object matching system that can provide translation, rotation and scale invariant recognition even when the object is partially occluded. Unlike classification, the goal in object matching is extended to extract the geometric pose of the object in addition to identity. Most earlier attempts to apply neural networks in object matching assumed a close to noiseless matching environment, or relied on application-specific a priori knowledge to simplify the computation. We consider here the cases where the feature extraction process is imperfect, where the environment cannot be controlled and the resulting scene is noisy with respect to the model. We present a novel and effective paradigm based on modified Kohonen neural networks. The merit of this approach is that it offers a scheme for multiple concurrent searches. These focus into different regions of the feature space and interact co-operatively using a compact framework comprising two Kohonen feature maps. The scheme has been subjected to a wide spectrum of rigorous tests, and results have proven to be reasonably accurate even in the presence of noisy scenes and with a large search space.