Sonntag, 20. Dezember 2020

Photoshop filters for noise, color and sharpness

Free and commercial solutions compared for scanned negative film

The challenge
 
I have scanned a series of old negatives with my Epson Perfection V600 Photo. Even when the scans are of high quality, the scans have a few problems, which are
A. Film grain
B. Poor color
C. Focus problems
Now I like to compare free and commercial solutions to see whether the money could be worth the investment in terms of time saving and unparalleled results. There is also a D: Dust spots on the film. A topic handled at first when scanning, but I moved this to the end of this article.

High quality scans mean for me scanning at 3200 dpi, in 16bit color and checking that there is no clipping in the histogram. The old family pictures were often made with no-name film types and a SLR camera and occasionally with a compact camera. This means that the input material already has quality issues. Those films suffer from film grain and also from a lack of sharpness. With the same camera, the Ilford black and white negatives are much sharper than the B-type color negatives. Also non-professional lenses have limited sharpness at high magnification. In addition, the scanner lens also has limits in sharpness (but as I know from scanning with the Canon Coolscan 5000, giving comparable results, I don't  think that my Epson V600 scanner is the limiting factor). The poor color is also primarily caused by the film type, as well as light conditions and light direction. Sometimes it is very challenging to get a pleasing color.

This review is for November 2020 (developments progressing quickly), with focus on scanned negatives with film grain and poor focus. I use standard settings or simple sliders to compare the methods and filters. Some of the software considered here is capable of doing more, but for my comparison not taken into account (If a company provides a demo version I should easily be convinced by its power and buy it, right?). The goal of the image editing tasks is to stay as much true to the original picture, preferring a rest-noise over loss of contour lines, and natural colors over artificial light, HDR and too much saturation.

For problems A, B and C, I subsequently describe the methods that I compare, after which I present my results. I compare methods that can be achieved within Photoshop (PS) without extra plugins, with external plugins. The Plugins from the Nik Collection have been provided in a free edition by Google a few years ago, which still can be found on the web. I use this free version in my comparison, since the filters have not changed very much since. I also compare trial versions of plugins of well and less well-known parties on the market. Topaz AI filters use artificial Intelligence and filter parts of the image differently. The examples on the website look very promising.

A. Reduction of film grain noise
 
The PS (Photoshop) own Reduce Noise filter is not very good, because it does reduce detail more than it reduces the noise. The filter needs to be part of a multi-step recipe.

1. PS Reduce Noise Recipe 1
Filter > Noise > Reduce Noise: Only remove color noise: 100%. Apply Dust and scratches with 2 pixels and threshold 16 (On lower resolution scans 4p and thr4).

2. PS Reduce Noise Recipe 2
Copy background layer. Filter > Noise > Reduce Noise: Remove noise 10, color noise: 100%. Apply Surface blur with 5 pixels and threshold 10. Set layer opacity to 33%.

3. PS with Camera raw filter
Copy background layer. Filter > Camera raw filter: * noise reduction +65, d40, c0 and * color noise reduction +50, d50, s50. Also set * Texture -75 and * Clarity +5. Apply.
Set layer opacity to 33%. I think this performs slightly better than the recipes above.

4. PS use channel information to reduce film grain
Sometimes the film grain is more prominent in one or two color channels and the third looks nice and smooth. In this case this method can perform well. Choose the channel with the least noise. Select all (ctrl+A) and copy, select rgb channels again. Paste as new layer. Set blending mode to luminosity and opacity to 50%. Paste again, Set blending mode to multiply and the opacity to 30%.

5. Dfine - from Nik Collection

6. Topaz DeNoise AI, price 80$

7. Noise Ninja (V2.4.2) - This was formerly a standalone filter and is now integrated in PictureCode Photo Ninja, price 129$

8. Boundary Noise Reduction 2 from Colormancer (I used the freeware edition for 8bit images)

9. Neat Image, price 40$


Example 1:


Example 2:


 
Noise reduction

Details retained

Notes
Verdict
(10/10)
Original - - - -
PS Noise Reduction v1
some reduction mostly retained
6/10
PS Noise Reduction v2 some reduction a bit smoother
6/10
PS Camera Raw sophisticated noise reduction
looses sharpness
7/10
PS Channels
occasionally very good, often little effect yes changes colors 6/10
Dfine satisfying loses sharpness
7/10
DeNoise AI removes noise by replacing it with something else almost no loss of sharpness - takes long generate previews, which makes trying different settings unattractive
- generates artifacts
- sharpens too
5/10
Noise Ninja good yes

8/10
Boundary Noise Reduction satisfying mostly retained
6/10
Neat Image good mostly retained creates unwanted haze 5/10

Conclusion:
with Photoshop you can do noise reduction quite well. With a layer copy and a mask you can regain sharpness for parts of the image, and have a filter applied for regions with much noise. The best noise reduction was expected from DeNoise AI, but delivered the lowest quality. The AI does not recognize film grain and 'guesses' textures. The older version of Dfine does perform equally well as it competitors. Most satisfying results were achieved with Noise Ninja, but the performance is only slightly better. Conclusion: I prefer Noise Ninja and setting opacity of the layer a bit lower to regain details, or use a grey pencil to mask out parts with detail like faces.

Update: I somehow missed Number 10 in my comparison: Imagenomic Noiseware, price 80$. I did not test it yet, but I remember from a test many years ago, that it performed comparably well as Noise Ninja.


B. Correct poor colors

1. PS > Image > Levels > Choose Options: Try the different options (and set clipping to 0%!) and choose the best. In two additional steps increase contrast with curves, and increase vibrance and saturation.

2. PS > Filter > Camera raw filter: * dehaze +15 and * clarity +5. Use 'auto' or use the eyedropper to set white balance. Set blacks and whites, set highlights and shadows to taste.

3. Viveza 2 - from Nik Collection

4 Colormancer - Color correction tools  - White Balance and Highlight Recovery are provided in a free edition.

5 Perfectly clear, price 129$

6 Color Perfect - 78$

Example 1:


This example has many part with white tones and it is difficult to catch skin tones. All methods improve white balance, but two stand out. The PS own auto levels and Color Perfect. The latter is giving indeed perfect whites and skin tones, without effort.

Example 2:


In this image the green and red leaves are expected to have certain colors, as well as the sky. They are wrong in the original. Color perfect gives the best colors, for the parts where it does not clip blacks. Not only the leaves and the sky have the right color, also the clothes and the bike and the plaster stones. Second best is PS Levels.


Example 3:


Also a photo with vegetation and sky, here with red bricks and white paint in a building. Color Perfect finds the right colors, but clipping is problematic here. The others do well too here. PS camera raw does the best job in improving highlights and shadows.

 
Realistic color
Highlights and
shadows improved

Notes
Verdict
(10/10)
Original - - - -
PS Levels+ yes no  fastest 8/10
PS Camera Raw +/- yes   8/10
Viveza 2 +/- shadows   7/10
Colormancer +/- highlights clips blacks and highlights 7/10
Perfectly Clear +/- +/- clips blacks and highlights 7/10
Color Perfect
yes +/- clips blacks and highlights 8/10

Conclusion: All methods can do a good job in generating more realistic color. The first two methods from Photoshop, especially the Camera Raw filter, can do as well what the third party filters can do. Where you expect from the additional filters to have a great automatic correction, they fail. One exception is Color Perfect, which yield in my opinion in some cases the best color with a few clicks. However I still need to find out where to reduce clipping, because in some cases CP clips too much. As a solution, one could use a layer copy and set it to color. In this way you overcome the clipping, because luminosity is retained in the original background.


C. Sharpening out of focus blur

1. Photoshop Smart sharpen (8-4-2)
PS > Filter > Sharpen > Smart sharpen. I use the following settings for best results:
Amount 200%, radius 8p, reduce noise 40% (for gaussian blur or lens blur); I set for both shadows and highlights the fade amount to 60% and tonal width to 70%. I repeat this with 4 pixels and 2 pixels. If you have a higher resolution scan, you can start with 16p or 12p and repeat with lower numbers.

2. PS high pass sharpening
This is a not so well known method but works quite well:
duplicate layer, Filter > other > high pass 10p. Set layer blending to hard light, eventually reduce to 80% opacity. (In other cases where blurring is less, you may want to use 'soft light' blending mode and reduce opacity to 25%.)

3. Sharpener pro 3 - from Nik Collection

4. Topaz Sharpen AI, price 80$

5. Focus Magic, price 65$
- Pro: estimates focus blur radius well. Con:Produces halo artifacts
Has a very good estimator of the radius of blur. It takes a great deal out of hand
Recovers details that others can't

Example 1:

In this example I wished to have sharp hairs, sharp eyes and few artifacts.The al sharpen well, some with more noise than others. Sharpen AI looks the sharpest, but severely mistakes parts and does not recognize the eyes as eyes giving almost creepy results. Focus Magic is the only one that can bring out certain details, such as the four segments (cross separation) of the window reflected in the eye (the others show two segments). However, FM is not necessary my favorite, because of the noise and somewhat unnatural look.

Example 2:


In this photo the child is out of focus while the background is sharp. Every method will produce over-sharpening of the background - also revealing exaggeration of artifacts. In this case a mask that separates fore- and background is the only solution to get it right (not applied here). In this example you see few differences in sharpness, but well defined differences in the artifacts and noise levels. The sharpen AI is a real mess: the background gets a painting-like texture, the clothes are really sharp, and the most important part - the face - is as blurry as in the original.

 
Reduces focus blur

Undesired artifacts

Notes
Verdict
(10/10)
Original - - - -
PS smart sharpen yes looses detail
7/10
PS high pass
yes introduces noise
7/10
Sharpener Pro 3 yes introduces noise slider for correcting focal blur limited at max
7/10
Sharpen AI yes many slow
4/10
Focus Magic yes multiple, especially halo
measures focus blur accurately 7/10

 
Conclusion: All methods have their advantages and disadvantages, while none is really outstanding. The worst result was produced by Topaz Sharpen AI. Though some parts are nice and sharp and from a distance seen it delivers the sharpest result, most parts are totally mismatched in material and goal, resulting in a painting rather than a sharp picture. From the remaining methods, the sharpest result was obtained with Focus Magic which also introduced most artifacts. Nik sharpener does a good job as well, but needs a noise reduction afterwards. The Photoshop internal methods are well balanced in sharpness and introduced noise. The high pass method produces more noise and sometimes too dark borders. A good free method may therefore be to combine both Photoshop methods, as I did in an action HERE*. For sharpening out of focus blur, try to treat different parts of the image differently with masks.


Overall Conclusion
 
In all cases, Photoshop offers techniques that yield results that are almost as good as those provided by third party plugins. A few of the secret algorithms of plugins did do a slightly better job. Concerning the Topaz AI filters, they need to improve a lot to outcompete the other methods. They apparently have not trained their algorithm with scanned film negatives. It still wonders me that even human faces are not detected as such. I guess the AI needs to develop another few years till it gets 'mature' for many types of images.
I regret a little bit that I haven't tested Photo Ninja (I was tempted to use an older version of Noise Ninja that I found on the web). Photo Ninja does not only incorporate the patented and improved noise reduction algorithm, it is a complete package that also corrects color and does sharpening. As winner in one category, I wonder how it performs in the other two.
Last but not least, Photoshop has very recently released a version for 2021 with new exiting AI filters and goes the same way as Topaz AI. What will the future bring?
 
 
 

D: Dust spots on the film

At high magnification such as 3200 dpi, you always have to deal with some dust on the glass and especially on the film material. I use VueScan (60€) to scan the infrared channel (which tracks the dust very well), save it in the output (RGBI), and process this channel in Photoshop. I first crop the borders away (extra dust that does not need to be processed). I then create a good mask from the IR channel and use content aware fill to remove it. It is better than the dust removal algorithm in VueScan (or in Epson Scan). Who likes, can download my action for it HERE*. One has to be careful with flash light pictures, which result in IR channel capture in reflections, e.g. a Christmas tree with candles and reflective decoration (deselect manually after stop). After the content aware fill, I validate the correction at 100% magnification. Sometimes 'content aware' msitakes the content and needs manual correction instead.
This option for IR Channel was the reason for me to buy VueScan. It allows me to do the best automated dust removal possible. But if Epson Scan would be capable of doing this, I would use that. VueScan tends to clip the shadows, even with multi-exposure. Changing the film type setting (e.g. Kodak - Commercial, Fuji NHG, 3M Scotch color At 200) can save the scan, but for problematic cases, only the raw scanner output can do the job (the negative version). The option in the Epson Scan of setting the levels manually is so convenient and gives better color scans than I get in Vuescan. Experienced users may get more out of Vuescan but I consider this lack of level setting a big drawback of Vuescan (7/10). Silverfast (Version AI Studio, price 250$) overcomes most inconveniences but is costly.

Update: Good news: many owners of an Epson Scanner can download the SE edition (Normally 50$) for free:  https://www.silverfast.com/show/bundle-epson/en.html
This basic SE version (in contrast to AI Studio) is not intended to work with the dust IR channel, but it can save the scan in 32 or 64bit HDRi (where the i stands for the infrared channel). Unluckily, this output in Tif or DNG cannot be read directly in Photoshop. Nevertheless, I needed to find out wether I could get the same result as with Vuescan. So what I did is making an optimized scan with Silverfast SE (slightly more control over channels and histograms than Vuescan) and save it as tiff. After that I made an Raw scan and save it as HDRi raw tiff. You need software like Gimp, Infanview or ImageJ to extract the IR channel. (Or you try to find a try to find a photoshop plugin). In my case the Gimp did not do the job, so I tried Irfanview and opened the document. Option > Multipage images > Edit multipage tiff. Here I moved layer 3 to the top and saved under a new file name. In Photoshop I opened both Images (scan 1 and the saved IR channel scan). You can go to channels, create a new channel, and paste the IR image in it. Now you have the same image as you would have with Vuescan, and you can apply my PS action to remove the dust. And yes, working with PS content aware fill is better than how Silverfast deals with the dust correction!


- Action to use the IR Channel from a Scan to remove dust by content aware fill.
- Action for Sharpening with both High Pass and Smart Sharpen

Keine Kommentare:

Kommentar veröffentlichen