Invariant image recognition using a multi-network neural model

A new model which permits visual patterns to be invariant to affine transforms (translations, rotations, and dimensions) is presented. A training multilayer fully connected network of ADALINE neurons is proposed as a preprocessing step for invariant image extraction. A second neural network has been trained by the popular backpropagation algorithm for recovering the real image without distortions. First, the sample invariants are obtained by the preprocessing network. In the second step, the general invariant that includes all the sample invariants is computed. Afterward, the reordered sample invariants are input to a multilayer neural network trained by the backpropagation algorithm. The original image, without distortions, is obtained in the output of this system. Several test images have been computed, and evaluation of the results shows that in the case of images with intrinsic perceptual similarity, the learning procedure leads to a global invariant extraction that requires less computational effort in comparison with an arbitrary training selection. After the training process, this system is able to extract the generalized invariant image from an arbitrary picture recovering the input image without distortions.<>