Invariant pattern recognition regardless of translation, scaling, and rotation of objects is a very difficult problem. A third order neural network can extract invariant features of each class as well as classify objects in different classes with a priori relationship between input active pixels. In order for the effective invariant feature extraction from two dimensional images in noisy environment, noise suppression and edge detection as preprocessing are necessary. Computer simulation shows that morphological third order neural networks have the capability of recognizing invariantly different classes of objects transformed in position, size, orientation and corrupted by noise.
For better recognition of objects having very similar shapes, partial global features that are extracted from the input image by local feature planes are used as inputs to third-order neural networks.
