Neural networks for invariant object recognition

A neural network model for invariant object recognition is presented. The model consists of two stages. In the first stage, features are extracted from an image of an object, and the second stage is used to recognize the object. A neural network is used as a classifier in the recognition stage. Consideration is given to rotational, translational, and scale differences. Many techniques for invariant feature extraction are available in practice. They include moment invariants, Fourier transform coefficients, complex log images and their transforms, adalines, etc. The technique of moment invariants for feature extraction is investigated. Two types of learning paradigms are used: backpropagation learning and competitive learning. In backpropagation learning the network learns with training samples, whereas in competitive learning the network learns without training samples. As an illustration, images of various types of aircraft are considered.<
>