Invariant object recognition using higher-order neural networks, line-segment spectra and multi-resolution training

A second-order neural network architecture is introduced that achieves invariant recognition with respect to an object's position and orientation in an image. This network does not show the combinatorial growth in network size as image size is increased, which is commonly observed in higher-order architectures. A new concept called an object's line-segment spectrum is introduced. It is argued that the weights of the second-order architecture are determined by these line-segments. Training time then becomes a function of object size rather than image size. The network is tested on the 26 capital letters of the alphabet. It is shown that in this application a multi-resolution training approach leads to reduced training time and improved performance.