Noise Reduction
Noise reduction remains one of the most important features in a video processor. Noise can be introduced during many stages of the movie-making process. It can be a result of film grain or sensor noise at the time of recording, during the film editing and digital compositing, or even compression noise at the time of encoding.
Although there has been some recent concern about “cheating” with drivers, the real issue is that noise reduction continues to be a difficult algorithmic problem. A professional level noise reduction algorithm such as that found in an Algolith Mosquito can cost as much as $3000 for a device. In the consumer area, HQV tends to offer the best noise reduction capabilities thanks to its true per-pixel processing.
One of the best overviews of noise reduction can be found at
http://www.hqv.com/technology/index1/noise_reduction
It comes from HQV.com, but it is very accurate in describing the issues at hand. I know, because I did the storyboard for that video. In general, the more noise reduction you apply, the more detail is lost.
NVIDIA PureVideo uses an adaptive temporal filter. This works by averaging pixels over several frames. When the noise reduction is turned all the way up, this produces some smearing of the image (which can be seen in the images that were distributed by AMD). At lower noise filtration, this effect is diminished, and by default NVIDIA does not enable noise reduction. ATI Avivo does not allow end-users to adjust the level of noise reduction.
The difference between AMD, NVIDIA, and HQV’s noise reduction techniques is how well the algorithms are at determining motion. Adaptive spatial-temporal noise filter works by identifying the pixels that are in motion -- these should be exempt from temporal noise reduction. The remaining pictures can then be filtered. HQV detects motion on the per-pixel level, allowing it excellent discrimination between motion and static elements of an image. NVIDIA detects motion on a region-based level which allows it to distinguish between motion and static elements of an image, but with less precision. AMD’s solution also features a spatial-temporal noise reduction algorithm, but it does not allow you to adjust the intensity of the effect.
![Autumn 2007 Video Processing Face Off [ No noise reduction @ 1920 x 1080 ] > View Full-Size in another window.](images/19-s.png) No noise reduction
|
|
![Autumn 2007 Video Processing Face Off [ ATI Avivo HD @ 1920 x 1080 ] > View Full-Size in another window.](images/20-s.png) ATI Avivo HD
|
|
![Autumn 2007 Video Processing Face Off [ NVIDIA PureVideo HD @ 1920 x 1080 ] > View Full-Size in another window.](images/21-s.png) NVIDIA PureVideo HD
|
|
![Autumn 2007 Video Processing Face Off [ Silicon Optix HQV Med Noise Reduction @ 1920 x 1080 ] > View Full-Size in another window.](images/22-s.png) Silicon Optix HQV Med Noise Reduction
|
|
![Autumn 2007 Video Processing Face Off [ Silicon Optix HQV Max Noise Reduction @ 1920 x 1080 ] > View Full-Size in another window.](images/23-s.png) Silicon Optix HQV Max Noise Reduction
|
|
![Autumn 2007 Video Processing Face Off [ Image comparison. To see full resolution image, please select click for hi-res version. @ 1080 x 198 ] > View Full-Size in another window.](images/24-s.png) Image comparison. To see full resolution image, please select click for hi-res version.
|
|
The Realta HQV is the king of noise reduction in this round-up with NVIDIA coming in second. AMD appears to offer similar quality to NVIDIA, however the Radeon is displaying an image with less than full 1080p resolution. Any time an image is downscaled, noise gets filtered out – this prevents fair comparisons from being made as we are now comparing kumquats and oranges.
In this test clip, there is a split screen with a static right side of the screen. This is intended to provide a reference image. As a static image, very little noise filtration should occur. In general, filtering noise from static images results in a loss of picture detail unless you are dealing with sophisticated wavelet-based algorithms such as that found in the digital photography world (Noise Ninja, Neat Image, etc.)
![Autumn 2007 Video Processing Face Off [ Effect of Noise Reduction on a still image @ 1080 x 198 ] > View Full-Size in another window.](images/25-s.png) Effect of Noise Reduction on a still image
|
|
Once again, direct comparison with AMD images are impossible. Since the image is downscaled, an equal sized crop will provide the appearance of a sharper image. NVIDIA’s noise reduction (when set to 51%) preserves most of the detail in the image. It does appear to be a smidge blurrier than the still image. The HQV noise reduction algorithm set in its default “medium” intensity provides a gentle reduction in noise, even in these static areas, without any significant loss of resolution. When set to “high” intensity, some details are beginning to disappear. Even so, it’s still sharper than NVIDIA’s solution (look at the deep purple veins in the lower right petal) while offering a creamy, smooth reduction of noise.
There is no question that Silicon Optix’s HQV noise reduction remains one of the best in the industry. We hope to take a closer look at products such as the Algolith Mosquito or Flea HDMI in the future.