Fast implementation of Gaussian filter by parallel processing of binominal filter

Abstract

Some of image processing techniques including noise reduction and feature extraction are realized by convolving filters designed for various purposes. A Gaussian filter is a smoothing filter, and is used in various applications such as feature point extraction. However, since coefficients of a Gaussian filter are real numbers, which requires a computational burden especially for large deviation filters. This paper describes an approximation of Gaussian filters using multi-layer convolutions of the basic binomial filter, which is implemented only by an addition and a shift operation. This study also aims at fast implementation with a parallel computing of the binomial filters on GPU (Graphical Processing Units) under CUDA (Compute Unified Device Architecture) platform introduced by NVIDIA corporation.