A hybrid learning network for shift, orientation, and scaling invariant pattern recognition

A three-layer neural network is presented as a generic approach for visual pattern recognition invariant with respect to the geometric appearance such as translation, orientation and scale of the patterns. The invariant recognition is achieved by representing the geometric variations internally in the network by nodes in the input and middle layers, which are laterally connected and trained by a hybrid algorithm combining both competitive and Hebbian learning. As the result of the hybrid learning, each pattern will be represented by a particular subset of middle-layer nodes all specialized to respond to the same pattern but with different geometric appearances. The nodes in the output layer are then trained by competitive learning to recognize the different pattern internally represented by the middle-layer nodes, independent of their location, orientation and size. The proposed algorithm is generic and robust and can be applied to various practical recognition problems. Moreover, the network is relatively simple and biologically plausible and can serve as a computational model to account for the invariant object recognition in the biological visual system.