线性复杂度双边滤波快速算法研究

The bilateral filter serves as the foundation for many algorithms in the image processing and computer vision community. However, the speed of these algorithms is affected heavily by the bilateral filter. In order to accelerate the computing speed, people proposed many acceleration algorithms. Nevertheless, these algorithms lack of detailed theoretical analysis and efficient implementation. Our research will establish a unified bilateral filtering framework, design new fast algorithms, construct a compiler to translate the fast algorithm into efficient code and widen the application spectrum of bilateral filtering. Specifically, our research will 1) accelerate the speed of the fast bilateral filtering algorithm without scarifying its approximation accuracy; 2) implement a compiler that translates the fast filtering algorithm into hardware-efficiency code automatically; 3) investigate the possible application of the fast bilateral filtering algorithm in optimization to break up the application constrain of the bilateral filter. Finally yet importantly, our research not only can speed up a set of algorithms building on the bilateral filter, but also can widen the application range of the fast bilateral algorithm via optimization.

双边滤波是构建图像处理与计算机视觉中一大类算法的基础，但其过高的滤波复杂度却限制了这些基于双边滤波构建起来的算法的运算速度。为了提高计算效率，人们提出了各种加速方案。然而相关工作却一直缺乏基础理论支撑与高效代码实现，本课题将建立统一的双边滤波快速算法框架，并以此为基础设计新的快速算法，构建由快速算法到高效代码的转换系统，拓展双边滤波的应用范围。具体来说，本项目包含三个方面：1、算法设计：在不降低逼近精度的条件下，进一步提升线性复杂度双边滤波算法的计算速度；2、算法实现：实现双边滤波快速算法到各个并行计算平台代码的转换系统，得到双边滤波快速算法在当前硬件上的高效实现，降低运算时间。3、算法应用：研究双边滤波快速算法在优化问题上的求解方法，突破双边滤波算法只能处理图像的限制。项目预期将会加速计算一大类基于双边滤波的算法，并可通过优化的途径从根本上扩展双边滤波的应用范围。

本课题重点解决双边滤波计算效率问题，建立了统一的双边滤波快速算法框架，并以此为基础了设计新的加速算法，构建了由快速算法到高效代码的转换系统，拓展双边滤波的应用范围。我们构建了统一双边滤波快速算法框架，并以此为基础设计新的快速算法，构建了由快速算法到高效代码的转换系统，拓展了双边滤波的应用范围。相关成果共发表论文7篇，申请专利2个，培养硕博士研究生12名。