A structured neural network invariant to cyclic shifts and rotations

Abstract

Shift and rotation invariant pattern recognition is usually performed by first extracting invariant features from the images and second classifying them. This poses the problem of not only finding suitable features but also a suitable classifier. Here a class of structured invariant neural network architectures (SINN) is presented that performs adaptive invariant feature extraction and classification simultaneously. The special characteristic of the pyramidal feedforward architecture of the SINN is sparse connectivity and the use of shared weight vectors. This guarantees the invariance of the network output with respect to cyclic shifts and rotations of the input image. In experiments the recognition ability of the SINN is shown on a database of textile images. Without any preprocessing of the images and without the need to choose an appropriate classifier the SINN achieves similar or even better results than standard pattern recognition methods.