Reduced Complexity Architecture for Integral Image Generation

Integral image plays an important role in AdaBoost algorithm which uses Haar-like features. The calculation of integral image needs many accesses to memory and most of the required addresses are not sequential. In addition, its calculation is compute-intensive. In this paper we propose an approach for generating integral image to cope with nonesequential addresses by affine transforming input image and using a pipeline architecture to compute results in an improved way. This approach needs the lowest clock pulses for integral image generation and in addition, its architecture is improved and less complex.