Neural network classification of symmetrical and nonsymmetrical images using new moments with high noise tolerance

The classification of images using regular or geometric moment functions suffers from two major problems. First, odd orders of central moments give zero value for images with symmetry in the x and/or y directions and symmetry at centroid. Secondly, these moments are very sensitive to noise especially for higher order moments. In this paper, a single solution is proposed to solve both these problems. The solution involves the computation of the moments from a reference point other than the image centroid. The new reference centre is selected such that the invariant properties like translation, scaling and rotation are still maintained. In this paper, it is shown that the new proposed moments can solve the symmetrical problem. Next, we show that the new proposed moments are less sensitive to Gaussian and random noise as compared to two different types of regular moments derived by Hu. Extensive experimental study using a neural network classification scheme with these moments as inputs are conducted to verify the proposed method.