Pattern Recognition Using Shift Invariant Fourier-Mellin Descriptors And A Back-Propagation Net
Claude Lejeune, Yunlong Sheng, Henri H. Arsenault · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 1990
Fourier-Mellin filters are used to generate invariant feature descriptors. The position of the maximum correlation output for each filter is used to create a distance vector. This vector is invariant under translation, rotation, change of scale and intensity of the input object. This method is applied to seven objects and the resulting, vectors are fed to a neural network for recognition. Of the seven objects, five are used to train the network. A three-layer feed-forward network, trained with the back-propagation algorithm is employed. Results show that distance vectors are suitable inputs for recognition by a neural network. The network learns the associations and recognizes the objects.