CUDA Based Computation of Quadratic Image Filters

Image processing applications usually requires nonlinear methods due to the nonlinear characteristics of images. Quadratic image filter which is a class of nonlinear image filters are widely used in practice such as noise elimination edge detection and image enhancement. On the other hand, second order products of the pixels make quadratic image filters computationally expensive to implement when compared to linear convolution. In the last decade, CUDA accelerated computing has been widely used in image processing applications to reduce computation times. In this study, an efficient method for the CUDA acceleration of the quadratic image filter has been implemented. For this purpose, alternative algorithms were examined comparatively since the performance of the GPU is sensitive to memory utilization. Because quadratic filter has a large number of coefficients and quadratic terms, the algorithm which utilizes the shared memory for storing image blocks provided the best throughput among the examined methods. Comparative results that were obtained using various images in different sizes show significant accelerations over sequential implementation.