FRACTIONAL CENTRAL MOMENT METHOD FOR MOVEMENT-INVARIANT OBJECT CLASSIFICATION

Within the context of moment methods for movement-invariant feature vectors the authors derive a new 'low-level' moment method capable of retaining scale and translation properties demonstrated by the alternative central moment low-level moment method. The new low-level moment method, denoted fractional central moments (FCM), provides a path for expressing the high-level moment method of pseudo-Zernike moments in terms of low-level moments, thus defining a set of feature vectors providing invariance to translation, scale and rotation of objects contained within the image space. The FCM representation provides more moment method terms per order than alternative low-level moment methods, thus it is shown to demonstrate greater image encoding/descriptive properties at a given maximum moment method order. The authors quantify differences between central and fractional central moment methods using discriminant analysis as applied to a specific data set proposed for the purpose of investigations described in a sequel paper quantifying neural network generalisation ability.