引导滤波算法的CUDA加速实现

Speed-up Implementation of Guided Filtering Approach Based on CUDA

针对引导滤波算法运算速度慢、无法实时处理的问题,提出基于统一计算设备架构(CUDA:Compute Unified Device Architecture)实现引导滤波算法的加速。利用CUDA并行编程实现图像邻域窗口像素值求和,进而获得图像邻域均值;通过利用寄存器和纹理存储器,同时优化算法步骤,获得引导滤波关键参数,进而实现对算法的整体优化。实验结果表明,与基于CPU实现引导滤波算法相比,基于CUDA并行处理可在很大程度上提高运算速度,基本达到了实时处理的要求。

For the shortcoming of guided filtering,such as slow operational speed and non-real time processing,the algorithm is speeded up based on CUDA（ Compute Unified Device Architecture）. In the proposed method,the sum of neighbor pixels＇ value is calculated based on CUDA parallel programming,and the mean value is calculated. The key parameters of guided filtering are obtained by taking advantage of texture memory and registers and algorithm optimizing. The whole optimum of approach is achieved. Experimental results show that,compared with CPU-based guided filtering algorithm,the proposed CUDA-based algorithm is greatly speeded up and basically meets the requirement of real-time processing.