Word length reduction for the integral image

The integral image is an image containing accumulated sums of pixel values taken from an input image. It is an important concept for multi-scale image processing algorithms, for it provides a very economic way to compute the sum of pixel values in any rectangular input image region. Unfortunately, the integral image requires a large binary word length to represent the accumulated sums. This is an issue for platforms having limited memory, power, and bandwidth like in mobile devices. Our paper deals with two methods for word length reduction, involving computation through the overflow and rounding with error diffusion. We show by experiment that, based on a word length reduced integral image, the Viola and Jones face detector for a VGA resolution can work on a 16-bit CPU (i.s.o. 27 bits, which becomes 32 bits on byte-oriented CPUs), enabling face detection on a wider range of platforms.