Today’s noise reduction software is capable of incredible results. Images in the past can be cleaned significantly with modern Danoise algorithms. But what is the real advantage of these devices than capturing more light in the first place?
Today, I answer the answer to that question by measuring the performance of the algorithm and capturing the light by measuring the performance of the algorithm. I’m going to focus exclusively Dxo’s pureraw 4 software (Photography life reviewed here) Both due to its popularity and due to its high performance. I have also tested more traditional noise reduction algorithms that do not depend on machine learning.
Modern noise decrease performance
Let’s look at an example of an image with additional noise – a rejected picture taken on my Nikon D500 20,000 ISOs:
Wow, noise! Above, I have indicated the crop that I am using to show you what it looks.
Frankly, without any reject, the result is frightening. I tried to reject it using a non-machine-learning algorithm in Rawtherapee, and also with machine learning algorithms in Dxo Pureraw 4:
I think the results speak for myself. Traditional noise reduction algorithm does not perform the DXO, topaz and still the machine learning algorithm used by Adobe -with today’s machine learning algorithms. He said, you still do not get the right quality in the processed image due to high level noise in the original.
Decreased noise, or capturing more light?
In such tests I rarely see, compared to capturing more light in the area. How to do Today’s algorithms compare to collect more light only?
In other words, if I could take the same shot accurately, but with the shutter speed twice or four times long, how would the best noise cut algorith be compared? We all said that “ISO 6400 is now like ISO 800” and various such claims. Well, I have done such a test using a strong tripod, a cable release and a trial theme of a bill:
In fact, to see the effects of the algorithm decreased noise, I have used a tight crop:
I took the gradual photos of this scene at 1/800, 1/400, 1/200, 1/100, 1/50, and 1/25 second shutter speed. As a result, an additional stop of light was captured each time. Conversely, I reduced my ISO every time. Here are the results:
In terms of expanding expansion and image quality, where do modern noise reduction stand on the algorithm list? To measure it, we need an objective, mathematical standard of measuring image equality.
There are several algorithms to measure deviations from an ideal or “ground truth” image. After testing a half a dozen image equality measures, I found that the image was great in measuring quality loss due to noise: the so -called UIQ or “Universal Image Quality Index”.
Zhou Wang and Alan C. According to Bowl, which Published this algorithm in 2002It measures “loss of correlation, luminance deformation and contrast deformation”, as I have come to know, broadly corresponds to the appearance and perception of expansion.
I used this UIQ algorithm to measure noise in different types of images – with some noise decrease, some simply took over the first place with more light/a low ISO. How many stops are you getting effectively with today’s best noise? These results are:
The score of one is a perfect score. The “original” labeled image is applied with a decrease in noise in ISO 6400 and 1/800 seconds. My ideal image is taken on 1/25 seconds and base ISO 200, which is five stops more light than the original photo. (I have labeled the “five stop” in the above graphic, and according to the definition, it gets a correct score of 1.)
You can see that in this comparison, there is no doubt – a machine learning noise decrease algorithm such as the traditional noise reduction in DXO Pureraw 4 is a clear step on algorithms. Such traditional algorithms score the same as one-stop improvement, while DXO Pureraw 4 is somewhere between one and two stops.
Here is described how this an example looks in the image, compared to the picture taken in ISO 1600 (two stops better than the original ISO 6400 shot):
Here, you can see that the result of DXO feels great. There is not a very clear noise. However, there is also less details – the image with two more stops of light is clearly fine details on the parrot’s face. This is why the UIQ index scored two photos equally – and if anything, gives two more stops to the photo with stops.
I also want to show comparison against traditional noise reduction, such as found in Rothrapi or Darkateable:
The DXO image clearly seems better to me. But something else holds my eye: The collapsed lines on the parrot’s face have been converted into embodied lines by DXO! This suggests that the machine learning algorithms invent slightly expansion through a microscopic level. You can see it very clearly than below (vs. “ideal” image was taken in Aadhaar ISO and 1/25 seconds):
This suggests that in a way, DXO Pureraw 4 and perhaps other machine-learning algorithms are less like denozers and more like “re-draining algorithms”. They use a network trained on millions of images to decide whether to launch the details. Comparing, the traditional Danoizing algorithm did not do the same thing in comparison.
Discussion
There is no doubt that the DeepPrimexds algorithm of DXO Pureraw 4 performs an excellent work. This can give you decent images, even though you give it the noise taken in ISO 20,000, and today some pictures are saving which were not in the past.
At the same time, such algorithms are not an option to get more light – when you Be able to do Get more light, that is. I do not buy in the idea that today’s best noise decrease you get more stops of improving 3, 4, 5, or high-like images. Instead, it provides around two-stop corrections in the relative performance of an unidated photo, and traditional noise reduction about a stop of correction relative to algorithms.
In addition, DXO Pureraw 4 can add a small amount of projection to a fine scale, effectively estimate the extremely fine details to get the result – which is not everyone in which not everyone is comfortable in which everyone is comfortable. , Which includes himself.
Finally, the machine-learning Danoizing ISO makes the most difference in the 6400+ range. Modern sensors do ISO 3200 and below that very well, and noise in such images can be cleaned with a traditional algorithm without major issues. And, in my experience, I find the best images on these lower ISO values ​​anyway, as strong light gives better color and expansion.
Therefore, while DXO Pureraw 4 and other machine learning noise reduction can definitely improve noise images than traditional algorithms, it still pays to customize your camera settings if you want the best image quality if you want the best image quality Are. It is better to capture more light than using software to make for extremely high ISOs. And no software can create a high-like photo as this base was taken in ISO.
Note: I want to thank DXO for granting licenses to use this software for test purposes.
