The Guide (really long)

Quick-start guide to NeatImage:

NeatImage ("NI", from here on) is a photo noise removal tool. Its modest price ($30 USD) belies its capabilities. An unlimited-time Demo version (slightly limited features, slower performance) can be downloaded for free from www.neatimage.com. The remainder of this article was written using NI Demo 1.1.1.6. At the time of writing, the NI User's manual is version 1.09 (dated July 2, 2002). Print it out and keep it for reference. After the next few pages of this article, the NI manual may be more useful to you than it is right now.

O.K., so you've downloaded the demo version (or purchased the full version), you've read section 3 (pages 8-9) of the user's guide on the test image (where you open their test image and apply a pre-created noise profile). The one-page “how-to” example they walk you through doesn’t REALLY show you how to create noise profiles and remove noise from images effectively, without over-doing it.

So, how do you use NI on YOUR images? This article is designed to take you to the next step, and keep you from having to read the entire 35-page user's manual. Consider this article as the "condensed" version of the NI User's Guide. Also included are one user’s tips, observations, and “gotchas”.

We pick up at section 4 of the user's guide (page 10)...

First, open your image. NI can open .TIF, .BMP, and .JPG files. You CAN paste into NI from the clipboard, but make sure that your graphics card is set up for 24-bit color. If you open a 24-bit file and are using a 16-bit screen, things going to the clipboard go at 16-bit (screen depth).

Next, build a device noise profile (the tab at the top that looks like a camera). The manual says "it is always better to build a new device noise profile for each image, because it better matches the noise characteristics of the image". You CAN re-use them, but only if the pictures are of basically the same subject, with the same lighting settings, the same JPEG compression, the same camera settings (sharpness, ISO, white balance, etc). Changes in ANY of those can cause the LAST noise profile you created to be a bad choice for THIS picture.

The first decision to make is the color space. NI says (skipping a few words here and there), "... you should use YCrCb JPEG for color photographic images, or the YCrCb Symmetric space for gray scale images. If you intend to ONLY filter ONE of the Red, Blue, or Green channels, then you will need to use the RGB color space". If you get half-way through the process and then want to change color spaces, you have to start all over. So, use YCrCb JPEG for color, YCrCb Symmetric for black-and-white, and you better know why if you think you want to use RGB. 'Nuff Said.

The Device and Device mode boxes are informational only, and can be used to help you with saved profiles (if you want to either use a generic profile as a rough starting point, or if you want to save this profile and later use the same profile on multiple images).

Now, select a nice, large (preferably gray, but use what you can) area of the image that has no details in it, so that NI can see what the overall noise profile is (drag the mouse over the image). Use clear sky, a blank wall, etc. It needs to be an area that contains no details; just noise on a solid patch. The absolute minimum is 60x60 pixels, and 100x100 up to 300x300 is recommended -- the bigger the area, the better. While you're moving the mouse around, the X,Y coordinates in the lower-left corner of the screen tell where you are. Once you click and DRAG over a selection area, the X,Y values become the WIDTH and HEIGHT of the selection. This will help you with knowing whether you've got a sufficiently large area.

If you're having trouble seeing the details in a light or dark area, you can use the "negative" button (or select it from the menu) to flip the image temporarily as needed (only in memory, the original is not changed).

Once you select your “device profile noise area”, the text in the box next to the blue ruler will say "selection can be analyzed", indicating that this is an acceptable area. Click on the blue ruler to get the noise profile from the selected area. Now the base profile is loaded, and the other three tabs are enabled (next to the camera). If, rather than "selection can be analyzed" it says "signal clipping", you've most likely selected an area that is too bright. Try another area.

That gives us what NI calls the "Device Noise Profile". This is the rough, overall noise profile. Now we need to fine tune it. Click on the tab whose icon looks like an equalizer. If the "Enable" check box is un-checked, check it.

Noise is quite often different at different brightness levels (the dark shadows may have more or less than the mid-tones, which may have more or less than the bright areas). This is what fine-tuning does -- it tells NI that certain brightness levels, in certain channels (you can fine-tune R, G, and B separately here), the noise is more or less than in the rough, overall DEVICE noise profile we selected in the first step. Each of the channels is broken into 9 bands, with the left-most being the darkest, and the rightmost being the brightest (see the gradient at the bottom of the equalizer).

Now select another (different) area (just like you did in building the device noise profile). For fine-tuning, you MUST use an area of at least 30x30 pixels. This is enough to catch what NI calls "high-frequency" noise. Think of high-frequency noise as "pixel-to-pixel" noise. I'm not sure what inter-pixel distance they use, but "high-frequency" means "fine grain". If you use an area that's at least 60x60 pixels, NI can capture information on both high- and "medium-frequency" noise -- noise from patch-to-patch (as opposed to pixel-to-pixel). Finally, areas of 100x100 or more can capture noise info on all three -- high-, medium-, and low-frequency (area-to-area; blotches it you prefer). The largest selection you can make is 300x300.

Try to select some fine-tuning patches that are bright, some that are medium, and some that are dark. When you select a patch, the numbers above one or more sliders will turn red. This means that IF you use that patch to fine-tune the profile, that's the brightness level that will be affected for that channel. If you want to use the selected patch, click the green ruler, and the equalizer slider for the band with the red number will change to match whatever noise profile is in that patch (if you DON'T want to use that patch, just select a new one -- nothing is really "set" until you click on the green ruler).

Note that the numbers on the sliders are positive and negative percentages. These indicate deviations from the "rough" noise profile captured in the first step. So, the first step is to get the rough profile, and then tell NI that at certain brightness levels, in certain color channels, there is more or less than that amount of noise.

When you click on the green ruler to fine-tune the profile according to the current selection, if the selection is good (valid), then the shading around the number above the slider will turn green. If NI thinks it is invalid (perhaps you selected a patch that has details in it, or is clipped), the shading goes to RED. If you've manually adjusted a slider to something other than what NI adjusted it to, it goes to YELLOW. You don't want any RED.

Suppose, for example, you select an area, and it affects the middle slider, and then you select another area, and it affects the same slider, but to a different degree. Which change should you use? NI's options let you set it to take the BIGGEST adjustment for a given band (most aggressive noise reduction), the SMALLEST adjustment (most conservative noise reduction), the average of all adjustments, or the last selection used. Personally, I leave it set on Maximum, and then apply the overall reduction sparingly (see below). Remember, we're still DESCRIBING the noise profile for this picture, not telling NI how hard to scrub it out of the image. I prefer to DESCRIBE it aggressively, and to SCRUB it lightly.

Select as many patches in whatever size you can to fine-tune the profile. Ideally, you'd like to find enough patches that you can use to fine-tune the profile across the entire spectrum, for high-, medium-, and low-frequency noise, but that's often not the case. If you've taken a good ROUGH profile, and have also fine-tuned 3-5 of the 9 bands, you're probably fine. The manual even says that fine-tuning is completely optional, but I think you’re crazy not to take the time to fine-tune the profile a bit.

You can also manually adjust the sliders, even if you can't get a patch for NI to analyze for you. Generally, you should be very conservative with manual adjustments. If you're just trying to get rid of high ISO noise, then focus most on the high-frequency noise (analyze more and smaller patches if you have to).

O.K. Now we've taken a rough "device noise profile", and we've fine-tuned it to this particular image. Now what? Time to SCRUB. The last to tabs are the noise filter (looks like a blue and white eraser) and the sharpening filter (looks like a blue crystal).

Whenever one of these tabs is selected, the operation of clicking or dragging in the image window is different than when selecting the noise profile or when fine-tuning it (the first 2 tabs). When you're in the Noise Filter or Sharpening tabs (the last 2), clicking an area of the image selects a 120x120 pixel area and displays it in the preview pane in the bottom-left corner of the screen. If you don’t see the preview pane, select it from either the toolbar or the menu. This is how we preview the noise reduction without actually running it (NI does the reduction on this little piece of the image on-demand). Select an area (just click, you don't have to drag, but you can drag to move the selection window). To have NI calculate the result, click on the icon above the preview that looks like a pair of eyeglasses (or press F3). If you want it to do so automatically, there's an option to turn on auto-preview, and the time delay (between when the selection is made and when the preview is computed) can also be set. If you want to zoom in, use the up/down arrows. You can’t zoom out any farther than 1:1. Personally, I find zooming in doesn’t usually help much.

The three sliders for the frequency components on the “Noise Filter” tab are indicated with fine, medium, and large "grains". These correspond to high-, middle-, and low-frequency noises. The sliders are deviations from the noise profile -- if you leave them at zero, then NI will try to remove what is in the noise profile. If you want it to completely ignore one of the frequency components, use -100 (that's 100 percent less than the noise profile). If you want it to get extra aggressive at one of the frequency levels, set it to +10 or +20 (110 or 120 percent of the noise profile). Dropping any of the sliders to -100 makes NI run faster, but don’t let execution speed drive the quality of your images.

The manual has some hints and guidelines here. It suggests that rarely should you go over +50%, and that typically +20 to +40 is as far as you'll need to go. Because I tend to use the more aggressive fine-tuning setting of "keep the maximum fine-tuning sample", and because I'm typically trying to remove ISO noise, I tend to set HIGH at 0 to -20, MID at -10 to -30, and LOW at -20 to -40 (I want to remove more high than mid, and more mid than low, but I don't want to be too aggressive on any of them). Clicking on the trash can sets the three sliders back to zero.

The Channel components can also be individually adjusted. "Y" is the luminance, and Cr and Cb are the Red and Blue Chrominance channels. The manual says "As human vision is not very sensitive to variations of colors, strong filtration in the Cr and Cb channels does not noticeably distort an image, but effectively reduces color noise". It also indicates that if there is "strong color noise", it is recommended to increase it to +30. Personally, I tend to leave these three at zero.

The last slider on this tab is the "Amount" slider. It defaults to 100, which means "remove 100% of all detected noise". I think this is where much of the "plastic" look comes from, and I typically drop the slider to 50-80% (remember, I was a little aggressive with the fine-tuning of the profile, so being a little more conservative here is helpful). The manual says: "For...images of asphalt, sand, or anything else that contains fine natural noise-like features, it is helpful to reduce the amount of filtration from to 60 to 80%...These values provide a good balance between preserving image details and noise removal". I would also add “hair” and “cloth or fabric” to the list of items with natural noise-like features.

Use the preview panes often and make adjustments to the sliders to get the best balance of noise reduction and loss of detail. Pick areas to preview where there should be little or no noise (blank walls, clear sky, etc), as well as areas where there IS natural variation (clothing, hair, textured surfaces, etc.). Much of the “plastic” look comes from leaving the sliders too high at this stage.

Last, but not least, is the tab for sharpening (the one that looks like a blue crystal). Again, as you start changing the sliders, keep using the preview panes to check the results.

The first item is three check-boxes for the individual channels. If you’re using RGB, the NI manual says you should use all three channels here. If you’re using YCrCb (either the color or B&W one), then you usually don’t need to sharpen the Cr or Cb channels (just Y; leave the other 2 unchecked).

The “Threshold” slider tells NI how strong an image detail has to be in order to be sharpened, and this is a percentage of the noise profile. The manual puts it pretty well here: “…if the threshold is +20%, then sharpening is applied only to those details that are 20% stronger than the noise level…The lower the threshold, the more likely it is that noise will be intensified. At +100%, only strong details are sharpened. The recommended setting is +50%.” Personally, I tend to go a little more conservative, and set this to a higher value, like 80%.

Note that the values for sharpening are preserved both between images, and between times you run NI. It is not stored as part of the noise profile, though they indicate this will be in a future version of NI. Don’t let your LAST image’s settings mess up THIS image’s results.

In the section of the manual on using the Preview Pane, NI suggests previewing a number of areas of the image (and adjusting the settings as needed) to get the best results. Their broad adjustment guidelines are:

1) If the noise/filtration looks too strong/weak, adjust the noise filter sliders for the appropriate channels and/or frequency ranges AND/OR fine-tune using more patches.
2) If the noise is amplified by sharpening, increase the sharpening threshold.
3) If the fine details of the image are not sharpened, decrease the sharpening threshold.

This is pretty broad. Note that proper noise reduction is an iterative process. Once you select the Device Noise Profile (the big patch), and then fine-tune that profile on the equalizer (with multiple little patches), and set the overall noise removal and sharpening levels, you need to preview representative patches of the image to make sure you’ll end up with the results you really want.

Filter the image. Click the Apply button (looks like “Play” on a CD player), or select it from the menu. The processing time is proportional to the size of the image (the number of megapixels. Any channels that have been turned off for optimization will make processing go faster, but be careful completely turning ANYTHING off. Usually, less application is more effective than none.

You can have NI automatically minimize itself during filtration, and automatically restore the window when it is finished. You can also set its thread priority. Usually you want to leave this alone, unless there’s something else CPU-sensitive you need to do at the same time, in which case, you want to drop NI’s thread priority.

Once the image is filtered, save the result. The purchased version can save as .BMP, .TIF, .JPG, or back to the clipboard. The free demo version can only save to .BMP. If you don’t like the results, you can select “Image | Reopen” to start over with the original image at the selected profile. If you’re used to using keyboard shortcuts, don’t instinctively try to use Alt+F, A for “File, Save As”. That will instead select “Filter, Apply”, and the image will be re-filtered.

NI’s options are all pretty straightforward, and can be set as you prefer. Take special note of the Equalizer and Preview options. Also, set your default color space to YCrCb JPEG if you typically work on color images.

The examples they give in section 6 of the manual (pages 24-27) are quite good. Take a look at them. There’s an oddly-placed comment on page 27 that is worthy of note: “Even though NeatImage tries to do its best to clean up such images [containing high ISO and JPEG Compression], please avoid using strong JPEG compression!” The less compression (which appears as noise in the mid- and lower-frequencies) you can use, the better. If all you have is ISO noise, then use RAW or TIFF from the camera if possible, as that avoids one source of noise altogether.

In the “Tips and Tricks” section (Section 8 on Page 30), they point out that in some cases, it is preferable to only filter shadows and other dark areas, leaving the lighter areas untouched. In order to do this, set the sliders for the brightest ranges of all three colors (R/G/B) to the bottom of the ranges in the equalizer. This effectively prevents NI from removing noise in those brightness ranges.