A neural network for calculating adaptive shift and rotation invariant image features

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 structured invariant neural network architecture (SINN) is presented that performs adaptive invariant feature extraction and classification simultaneously. The network is sparsely connected and uses shared weight vectors. As a result features especially well suited for a given application are calculated with a computational complexity of O(N) for N = 2n input elements. Experiments show the recognition ability of the invariant neural network on synthetic and real data.