HDR Cookbook – The Three Rules of Noise Reduction

One of the main problems of any HDR workflow that involves tone-mapping is noise. The most widely used tone-mapping operators are based on local contrast enhancement. Unfortunately, they also “enhance” the noise in your images. Thus, noise reduction (NR) is imperative. But what is the best way of doing this, and at which point of your workflow should you apply NR? In this recipe, I will show you the best way of reducing noise. I will present three simple rules that help you optimize your own workflow, and I will show you the results of a series of experiments in order to prove my point.

The Three Rules of Noise Reduction

Probably everybody who creates HDR images will agree that you need to apply proper NR techniques to your HDR images. But what is the best point in your workflow to do so? Which software should you use? And how should you apply this software? There are different answers to each of these questions, and it is not obvious which one is correct.

Let us start with the When. In the following, I will briefly list the steps in a fairly standard HDR workflow, and I will indicate the possible points at which you may apply noise reduction.

• Take your source exposures in RAW format
• Develop 16-bit TIFF images from your RAW files using a RAW converter like Adobe Camera Raw (ACR)
• NR opportunity no. 1: ACR (and other RAW converters) offer NR capabilities. You could reduce noise at this early stage.
• NR opportunity no. 2: You could also apply a more sophisticated noise reduction software one step after that – on the TIFF files ACR generates
• Load your TIFF files into Photomatix
• NR opportunity no. 3: Photomatix (and other HDR software) offers noise reduction capabilities. In Photomatix’ Processing Options dialog that opens immediately before you can create the 32-bit HDR, there is a Reduce noise option.
• Create the 32-bit HDR image
• Tone-map the image back to a 16-bit image
• Load your tone-mapped image into Photoshop for post-processing
• NR opportunity no. 4: You could apply an arbitrary noise reduction plugin right after you loaded the tone-mapped image and before you do any other post-processing work to get a clean start. Alternatively, this can be done using a stand-alone NR software before you load the tone-mapped image into Photoshop.
• Apply image enhancement software like Topaz Adjust
• NR opportunity no. 5: You could apply noise reduction after applying Topaz Adjust to get rid of any kind of noise that is produced by this kind of software.
• Apply arbitrary enhancement (contrast, saturation etc.)
• Apply sharpening
• It is common sense that sharpening should be the last thing you do. This is definitely too late for noise reduction.

That leaves us with 5 possible points in your workflow. You could come up with other opportunities, but let us stick with those 5 for the purpose of this discussion. Wherever you apply NR, it is always a trade-off between two things: Reducing the noise and retaining the details to get a clear and sharp final image. Therefore, you should not apply NR too heavily. However, you should also not apply it too often to the same image. For sure you will agree with that statement, but there is a good chance that you (yes, I am talking about you) are applying NR three times (or more) in your workflow: When you use your RAW converter, when you merge the HDR, and finally in Photoshop. This mix of at least three different NR technologies inevitably leads to images that have fewer details and more blur than necessary. This leads us to our

NR rule no. 1: Apply NR only once if at all possible!

When you look at the different technologies that come into play in our little workflow above, you will notice that they are very different in nature. ACR has a very simple NR interface and not a very convincing quality. Photomatix gives you an even more simplistic interface with merely a single Strength slider and without any visual feedback. Dedicated NR technologies, on the other hand, have become very sophisticated and powerful. They specialize on differentiating between the noise and the details in the image, removing the former and retaining the latter. There are a number of such products on the market and opinions on which one is the best vary. But one thing is common sense: All of these products are highly superior to the NR capabilities offered by ACR, Photomatix or anything built into Photoshop (at least up to CS4) or other image editing tools. There is a special technique called image stacking that is supported in Photoshop since version CS3. It allows you to stack equal exposures of the same scene on top of each other such that the random noise averages out. This is a very effective way of reducing noise that is used e.g. in astronomy photography. However, you would need multiple independent exposures for each exposure value to apply this technique, which is usually not the case in a standard HDR workflow. Thus, our next NR rule is

NR rule no. 2: Use dedicated NR technology!

Under these two rules, the possible points of applying NR in our workflow are reduced to 3 (opportunities 2, 4, and 5). But there is still the question as to where the dedicated NR software should be used. There are two philosophies with respect to this, and both have their followers:

1. Apply NR as early as possible to avoid that subsequent post-processing steps amplify the noise making it much harder to remove it without losing details.
2. Apply NR at the end in order to really kill all the noise in the image and avoid reintroducing noise through products like Topaz Adjust.

One can debate this question to great lengths, and people are doing this in diverse discussion groups all over the Internet. Personally, I tended more towards second philosophy. But I have learned that the opposite is true. Let me jump ahead here and present my NR rule no. 3:

NR rule no. 3: Apply NR as early as possible!

Why is that? Well, current NR technology is very good at recognizing the noise patterns produced by today’s cameras. If you process an image before applying NR, you change those patterns into something that the NR software cannot recognize that easily. Hence, It will have a tougher job of removing it, and in the attempt it will remove more of the details as well.

The consequence of NR rule no. 3 is that only NR opportunity 2 stays in the race. In the following, I will prove my points by presenting a series of experiments and showing you the results in a visual comparison.

The Experiments

I am approaching this comparison as scientifically as possible. In the following, I describe 6 experiments I conducted, with the NR step in different positions of the workflow. In order not to overwhelmed you with the details, I have omitted them here. You will find all the details in the appendix at the very end of this recipe.

In summary, I have conducted the following tests:

• Test #1: NR is applied only once using dedicated software as early as possible.
• Test #2: NR is applied only once after the tone-mapping and before any other processing steps.
• Test #3: NR is applied only once right before the sharpening at the end of the workflow.
• Test #4: NR is only applied once, in Photomatix.
• Test #5: NR is applied twice – in Photomatix and early in the post-processing workflow
• Test #6: NR is applied twice – in Photomatix and late in the post-processing workflow

Results of the Experiments

In the figure below, you see 100% crops from different regions of the test image for each of the tests (one test per column). Please click on the image with your middle mouse button (mouse wheel). This will open it in full resolution in a new tab. If necessary, zoom in to 100% in your browser window to compare the results.

Conclusions from the Experiments

Upon careful inspection of the results, you will notice that the early NR with a dedicated product (Test# 1) yields the best results. The crops in the left-most column are still detailed but yet most of the noise has been removed.

Applying NR early in the post-processing (Test #2) leaves homogeneous and visible noise (resembling film grain) across the entire image. This grain was added by Topaz Denoise (very mildly) to achieve a more natural look in the denoised image. Topaz Adjust has amplified it to a level where it becomes visible. Note also that Test #2 is much less sharp than Test #1. Applying NR late in the post-processing work (Test #3) produces a cleaner and also somewhat sharper result. The conclusion here is that if you apply NR only in the post-processing, do it after applying any image enhancement tools (such as Topaz Adjust).

In Tests #4, #5 and #6, Photomatix’ NR technology was applied. Using this technology exclusively (Test #4) yields a lot of noise and produces the worst result in the test. This is interesting as our winner (Test #1) applies the NR basically at the same point in the entire workflow – before the tone-mapping. This shows that Photomatix has some homework to do in terms of NR. In Tests #5 and #6, Topaz Denoise was used in addition to Photomatix’ NR tool (deliberately breaking our first rule of noise reduction). The results are comparable to Test #2 but notably more blurry than Test #1.

In summary, the experiments show that using a dedicated NR tool before merging and tone-mapping the HDR is the best solution. It reduced the noise and retains details in a way such that the resulting image does not need any further NR step in the post-processing work. The details are preserved best. If you thought that Photomatix would take care of your noise problems, this test may be an eye-opener for you. My advice is: always turn off NR in Photomatix!

How to Reduce Noise the Right Way

Now you know when to apply NR, but what about the How? Of course, this is highly depending on the product you are using. Different vendors use vastly different technologies and strategies. Some offer camera profiles and enable you to train the software to deal with the noise characteristics of a specific camera model. Since I cannot cover all the products here, I will give you a few basic tips:

1. Don’t underestimate the effect of NR. You may think that NR is something for pixel peepers. But keep in mind that according to our findings above, you will apply it very early in your workflow, leaving a lot of room for other tools and post-processing steps to amplify any noise that was left. So you should invest some time to get the NR step right!
   
2. Learn to master your NR product. Yes, that may seem obvious, but do you really understand all the sliders of your tool? Read the manual carefully and experiment with different images until you know what you are doing.
3. Use presets and automatic settings with care. I know that this is the most convenient way – you just have to press one button and the software does the rest. But in most cases, you will end up with suboptimal results because no software can really analyze the visual appearance of your image correctly. You may start by applying a preset, but you should always fine-tune the parameters to the specific properties of each image. This is where tip #2 comes in again!
4. Reduce the noise in each of your exposures separately. The noise will be different in each of your exposures. Therefore, you should not simply apply the same setting to each one of them. Optimize the settings for the specific exposure and keep in mind that different regions of each exposure will end up in the final image. Thus, there is no need to work extensively on the darkest regions of the darkest exposure. These will not make it to the final image.
5. Pay attention to the shadows. Noise is always most intense in the shadow regions since the signal-to-noise ratio of your image is very low there. That is, the actual pixels and the noise pixels both are dark. Brightening those areas boosts both and makes the noise really come out. A good NR product should let you adjust shadow and highlight areas separately. Work on the shadow areas in the brightest exposure (see tip #4).
6. Do not overdo it. Any NR software has a main strength slider, and the noise will disappear completely if you only push this slider high enough. But that will also remove the fine details that make your image appear sharp. You should work on those details to retain them. That may mean that some of the noise will remain in the image. You can fix this later.
7. Apply NR selectively. If you fail to hit the right compromise between reducing noise and retaining details, go for the details. This will leave noise that is most obvious in homogeneously colored regions (e.g. the sky). Photoshop’s layer masking allows you to blend different versions of your image in post-processing. So you can apply NR again in post-processing to optimize those noisy regions and then blend them with the rest of the image. This breaks rule no 1, but only a little.

Concrete Final Recommendation

I have not tried all the NR products out there. There are at least three major competitors: Noiseware, Noise Ninja and Topaz Denoise. Personally, I use Topaz Denoise, and I am pleased with the results. It does a very good job of reducing noise and retaining the details if used right. It offers separate controls for shadows and highlights, a way of recovering details, and a debanding technology that avoids color banding effects.

Denoise offers several standard presets that are a good way to start. Additionally, it has a noise estimation technology that adjusts every preset to the specific image. This is very useful. Usually, I go through the RAW presets starting at the lightest to see which one works best. At some point I see all the noise disappear in the preview. This is too much, and I go back to the previous preset. Based on this preset, I optimize the settings.

Please Refer to This Page!

Did you find this tutorial helpful? Did you use it in your work? Then there is a simple way of giving something back to me:

Please refer to this page when presenting your work online. You can simply use the following HTML code in your image description to refer to this site in a way that you think is appropriate:

<a href=”http://farbspiel-photo.com/”>HDR Cookbook</a>

Why should you bother to refer to this page? Well, for you it is a convenient way of revealing information about your work. And you know, the more information you give, the more attention you get. You do not need to write a whole novel because I already did this for you here. For me, the reference is beneficial because it generates some attention for this cookbook.

So, you see that referring to this page is good for both of us – a real win-win situation.

Thanks!

APPENDIX: Details of the Experiments

The setup was the following:

• The Source Images:
• Shot in RAW format with a Nikon D7000
• Settings: 1/25s, 1/100s, 1/400s, f/4,5, ISO800
• Development
• The source images were developed in Adobe Camera Raw (ACR) using neutral settings. That is, no noise reduction or sharpening was applied at this stage.
• Chromatic Aberration was reduced in ACR.
• In each test, the results were saved as 16-bit TIFF images and fed into Photomatix.
• Tone-mapping
• All images were tone-mapped with the Detail Enhancer option using identical settings.
• The result of the tone-mapping was saved as 16-bit TIFF
• Processing in Photoshop
• The tone-mapped image was directly loaded into Photoshop without any further intermediate steps.
• In each test, a Levels layer was applied to correct the contrast and a Saturation layer was applied to correct the colors.
• No other processing was applied except for the Topaz plugins as described in the test descriptions below.
• All Topaz plugins were applied with identical settings in all experiments.

Test #1

Dedicated pre-tone-mapping NR

• Preparation
• The TIFF images produced by ACR were loaded into Photoshop and the Topaz Denoise plugin was applied to each of the three exposures separately with different settings for each on of them.
• The results were saved back to 16-bit TIFF images and fed into Photomatix.
• Photomatix
• NR in Photomatix was turned off
• Photoshop
• Topaz Adjust was applied
• Topaz InFocus was applied (for intelligent sharpening)

Test #2

NR only in post-processing (early)

• Preparation
• No noise reduction was applied to the images produced by ACR. So the three exposures were not subject to any noise reduction prior to being fed into Photomatix
• Photomatix
• NR in Photomatix was turned off
• Photoshop
• Topaz Denoise was applied for early NR
• Topaz Adjust was applied
• Topaz InFocus was applied (for intelligent sharpening)

Test #3

NR only in post-processing (late)

• Preparation
• Same as in Test #2
• Photomatix
• Same as in Test #2
• Photoshop
• Topaz Adjust was applied
• Topaz Denoise was applied for late NR
• Topaz InFocus was applied (for intelligent sharpening)

Test #4

NR in Photomatix only

• Preparation
• Same as in Test #2
• Photomatix
• The three exposures were merged into an HDR using a NR strength of 150% (max. possible value)
• Photoshop
• Topaz Adjust was applied
• Topaz InFocus was applied (for intelligent sharpening)

Test #5

NR in Photomatix and early Post-processing NR

• Preparation
• Same as in Test #2
• Photomatix
• Same as in Test#4
• Photoshop
• Topaz Denoise was applied for early NR
• Topaz Adjust was applied
• Topaz InFocus was applied (for intelligent sharpening)

Test #6

NR in Photomatix and late post-processing NR

• Preparation
• Same as in Test #2
• Photomatix
• Same as in Test#4
• Photoshop
• Topaz Adjust was applied
• Topaz Denoise was applied for late NR
• Topaz InFocus was applied