Statistical dependence of pixel intensities for pattern recognition

Abstract

In this paper, we describe an algorithm for speeding up object recognition by reducing the amount of pixels taken into account when processing images. We show that some statistically stable regions can be found on an image. Taking just one pixel from each region preserves the most of information of the image. We employ linear dependency between pixel intensity values to organize neighbouring pixels in groups. Bayesian classification was chosen to prove suitability. We present the results that show computation speed increase without significant performance losses.