Moment matrices for recognition of spatial pattern in noisy images

We present a method for the detection and classification of a spatial pattern in noise contaminated binary images which is based on performing subspace decomposition on a nonnegative definite matrix of higher order moments of the image. We introduce a method which uses normalized power moments or ascending factorial moments as descriptors. While the set of p-th order factorial moments are in one-to-one correspondence with the set of p-th order power moments, the computation of factorial moments is much more numerically stable than the power moments. Indeed, using factorial moments we are able to implement pattern classifiers with over 30% more moment descriptors. We illustrate these techniques for word classification in binary document images