FAST ORTHOGONAL NEURAL NETWORK FOR ROTATIONTRANSLATION- AND SCALE-INVARIANT IMAGE RECOGNITION

摘要

In this article a novel method of image recognition invariantrnunder rotation, translation and scaling is presented. Thernproposed method is based on a fast orthogonal neural networkrnwhich, due to its structural analogy to fast algorithmrnfor Fourier amplitude spectrum computation, enables tornclassify images irrespectively of their translations. Thisrnproperty, in conjunction with log-polar representation ofrnimage amplitude spectrum, is subsequently applied to obtainrnalso the rotation and scale invariance. The proposedrnclassifier is compared to a multilayer perceptron and tornk-nearest neighbors method, showing its superiority in arnseries of tests performed on specially constructed imagerndatabases.