Implementation of Gaussian and Box Kernel Based Approximation of Bilateral Filter Using OpenCL

Abstract

A Bilateral filter is basically an edge-preserving and smoothing, non-linear filter. It consists of two kernels, namely spatial and range kernels which can be constant or arbitrary. Algorithms for bilateral filtering with constant time computational complexity are present today, but their execution time is too high for real time applications. Also, hardware latency and throughput sometimes reduce the speed of filtering. In this paper, we introduce a novel algorithm for bilateral filtering in which we combine box spatial kernel with Gauss-Polynomial range kernel. Parallel implementation of the algorithm is done on GPU (AMD Radeon HD 7650M) using OpenCL and an average run time of 15ms is achieved for an image of dimensions 256 x 256. Results of this algorithm is found to be about 15 times faster than the parallel implementation of bilateral filter with Gaussian spatial kernel and Gauss-Polynomial range kernel. We infer that while the PSNR values obtained are very close, there is a significant improvement in run-time when we use the proposed algorithm.