Effect of noise in moment invariant neural network aircraft classification

Abstract

An image may be reduced to a small number of moment invariants such that these are independent of the shift, scale, and rotation of an object in the image. However, noise interferes with the ability to provide invariance. The authors examine the effects of noise on the invariance provided by the moment invariants. They then show that rotation invariance is maintained in low levels of noise. They then show that a neural network may be used to provide robustness against noise. The moment invariants, computed for different levels of noise, are used to train a neural network to identify two aircraft. A split inversion algorithm is used because it is much faster than back propagation. The resulting network provides accurate classification in high levels of noise.<>