Kuwahara filter is a smoothing filter used in image processing for adaptive noise reduction that has the ability to preserve object edges. Applications for this filter exist in fields such as medical imaging and artistic imaging. However, it has a very high computational cost, especially when filter size is large. In this paper, we propose an efficient algorithm to accelerate this filter regardless of filter size. Our method uses Summed-area Table (SAT) to gain fast computation of mean and variance values in Kuwahara filter. After that, three acceleration methods, namely caching mean and variance values, memory access optimization and flexible-format SATs are proposed to optimize the SAT-based Kuwahara filter. The experiments show that our method achieves the lowest possible big-O time complexity and performs at around the same level regardless of filter size, while producing the exact same output image as the original method. The achieved speedup ratio grows quadratically with the filter size, ranging from 10x to 1000x and even more. As a result, our optimized filter can run at a high frame rate even when the filter is large.

Kuwahara滤波器是在图像处理中用于自适应降噪的平滑滤波器，具有保留对象边缘的能力。该过滤器的应用存在于医学成像和艺术成像等领域。然而，它具有非常高的计算成本，尤其是当滤波器尺寸较大时。在本文中，我们提出了一种有效的算法来加速此滤波器，而与滤波器大小无关。我们的方法使用汇总面积表（SAT）来快速计算Kuwahara滤波器中的均值和方差值。此后，提出了三种加速方法，即缓存均值和方差值，内存访问优化和灵活格式的SAT，以优化基于SAT的Kuwahara滤波器。实验表明，我们的方法实现了尽可能低的big-O时间复杂度，并且无论滤波器大小如何，其性能都大致相同，同时产生了与原始方法完全相同的输出图像。所达到的加速比随滤波器尺寸的平方增长，范围从10倍至1000倍，甚至更大。因此，即使过滤器很大，我们优化的过滤器也可以高帧速率运行。