Noise tolerant moments for neural network classification

Regular moment invariants face two limitations. First, images with symmetry in the x and/or y directions and symmetry at centroid give zero values for odd orders of central moments. Secondly, they are very sensitive to noise, especially the higher order moments. This paper presents a single solution to solve the symmetrical problem and reduce the noise sensitivity of these moments. The solution involves a new set of moment-based features that uses a reference point other than the image centroid. The reference centre is selected such that the new moment features are invariant to translation, scaling and rotation. The derivation of the new moments and their invariance are shown before experimenting them with some symmetrical alphabets. Next, they are shown to be less sensitive under the presence of Gaussian and random noise as compared to the usual regular moment invariants. Noise corrupted English alphabets are classified with a neural network to further verify the advantage of using the new moment features.