A novel affine invariant feature extraction for optical recognition

In this paper, we propose a novel method for extracting the affine invariant features of images, named the new polar normalized histogram (NPNH). The feature of an image is extracted from a polar histogram bins originating from centroid of the mass to all other points in it with 5 bins for r and 24 bins for Theta.However, the traditional normalization is rotation variant since it normalizes the image only on two directions: vertical and horizontal. Thus the normalization of the image with different divergences on two directions is different from the normalization of its rotation. The most intuitive way to overcome the difficulty is normalizing the images on all directions. After new normalization, the number in each bin of polar histogram is counted and it is lined row by row to form a vector. Then, the Fourier spectrum of the vector, called Fourier descriptor, is computed. Finally, experimental results of Optical Character Recognition (OCR) are presented and show that the NPNH is a simple, affine invariant and powerful distance in object recognition.