DxO PRIME denoising technology in DxO PhotoLab

 

Image noise can best be described as a random variation of pixels (impacting both luminance and color), that is not present in the scene being photographed. Image noise occurs mostly in low-light conditions, and mainly has two causes:

1. The first cause is called ‘shot noise’. This occurs when a sensor measures light received by counting photons. As photons are unconnected, counting a small number of them during a given time (which is the case in low-light conditions or with very short exposure times, leads to some variations from one measure to the next.
If we look at a uniform scene, we can expect each pixel to receive the same number of photons (A). But as the flow of photons is random, it doesn’t affect each pixel in the same way (B). Using a comparison of a rain shower: A is what we might expect, but B is what happens.

2. The second cause of image noise is ‘dark noise’. Even when there’s no light, there still are some random variations observed on pixels. This randomness is related to a mix of causes varying randomly across time (like dark leakage currents, thermal noise, etc.) or randomly across pixels (fixed-pattern noise, dead or hot pixels, etc.).

Denoising is a technique that attempts to remove noise while keeping information (textures, colors, etc.) that are present in the photographed scene.

ISO 12800 | F4 | 1/1250 s | 400mm

But in very low-light conditions, even the best denoising algorithms tend to remove details.

ISO 12800 | F4 | 1/1250 s | 400mm

A useful property from pictures to make a denoising algorithm is image auto-similarities:

  • For each small part of a picture (also called a patch), there is a high probability that many other parts of the same picture have a similar content.
  • For these patches with similar content, the only observed difference between them is noise (that creates random variations).
→ Useful information has auto-similarities, but noise doesn’t; exploiting this property is the key feature of many denoising algorithms.
  • For example, a way to denoise a given part of a picture is to average patches with similar content (the “similar patches”).

Denoising is easier when we know the noise model – this means knowing the plausible range of random variations

  • This noise model is built-in to DxO software thanks to camera calibration which is measured depending on the camera model and ISO value.
  • Denoising is better when it’s applied on a RAW image, because every processing step that’s applied on a picture before denoising complicates the noise structure, making it rougher and hence more difficult to remove.

Selecting similar patches seems easy when the noise level is low (cf. left picture below), but it’s much trickier when there’s a lot of noise in the image (cf. right picture below)

PRIME is a denoising engine for RAW images that outperforms its competitors, thanks to three principles:

1. Selection of similar patches is made robust to noise by using two-step denoising technique:

  • A first denoising step is applied to the picture. The image resulting from this denoising is called ‘Oracle’ (C)
  • A second denoising step is then applied using the original image but estimating patch similarities using the data from ‘Oracle’ (C).

Selecting similar patches in a more relevant way is the key to improving the preservation of details and textures.

2. Choosing a large neighborhood for finding similar patches.
This gives more chances of finding relevant similar patches, and thus accurately denoises (but this makes the algorithm slower, nothing comes for free…).

3. Combining information from similar patches is not just a case of simple averaging (which would assume a locally constant image model).
A more relevant – and thus more complex – image model is learned to better preserve useful information from the picture.

ISO 12800 | F4 | 1/1250 s | 400mm

DxO PRIME denoising technology

DxO PhotoLab 3

The Most Advanced Photo Editing Software

nik collection icons

Full features unlocked

From $ 69 | $ 49.99

FROM $ 129 | $ 99.99