Abstract

Denoising is an essential part of any image- or video-processing pipeline. Unfortunately, due to time-processing constraints, many pipelines do not consider the use of modern denoisers. These algorithms have only CPU implementations or suboptimal GPU implementations. We propose a new efficient GPU implementation of NL-means and BM3D, and, to our knowledge, the first GPU implementation of the video-denoising algorithm VBM3D. The performance of these implementations enable their use in real-time scenarios.

中文翻译：

---

NL-means，BM3D和VBM3D的GPU加速

摘要

去噪是任何图像或视频处理管道的重要组成部分。不幸的是，由于时间处理的限制，许多管道没有考虑使用现代降噪器。这些算法仅具有CPU实现或次优GPU实现。我们提出了一种新的高效的NL-means和BM3D GPU实现，并且据我们所知，这是视频降噪算法VBM3D的第一个GPU实现。这些实现的性能使其可以在实时方案中使用。