What’s your “image rehab” routine?

[derangedhermit]

I ran the sample images provided through PNGGauntlet and only got 2%-6% decrease in file size.

[DaleDe]

Look here: http://css-ig.net/png-test-corpus

Dale

[GrannyGrump]

Well, I decided to drop in on a Saturday visit, and see there is lots of discussion going on.

First, if it matters to anyone, I am posting some more images. I forgot one factor in the equation of image comparisons -- the 16-shades images.

The same two images as in the post above, 1200px wide. Saved with the following settings:
16-shades indexed, saved to png. These will not display properly on many (all?) ADE-based reader, long-standing bug.
16-indexed,convert to 256-indexed, saved to png.
16-indexed, saved to 256 default greyscale png.

Reducing images to 16-shades is not useful for jpeg format.
A brief comparison test saving the same image to jpeg got me these results:
326 kb -- 16-shades indexed palette
319 kb -- 256-shades indexed palette
319 kb -- 16 million colors
306 kb -- 16-shades converted to 256-greyscale
299 kb -- 16 million colors converted to 256-greyscale

So if you must use jpeg, it would seem you get best result by going from 16-million-colors to 256-greyscale. (this information is probably on Wikipedia or somewhere, but I don't get out as much as I should ) But I also see that jpeg with more than about 5% compression always dirties up the file, spitting lots of gray pixels into the white, so all our "death-to-specks" efforts are gone for naught.

I am also attaching the source images, originally downloaded in jp2 format, saved to 24-bit png, EDIT - NOPE, had to use the original jp2 files, png too large and won't upload ---] and the full-size cleaned images in 24-bit png. Maybe someone will see something useful to suggest.

@derangedhermit, I wonder if the aliasing you see might not be an inherent result of "cleaning" --- brightening up the background from dark sepia to white is going to lose some of the paler shades. If I use a blur or "softening" tool, the entire image starts to look too soft. I wonder how other folks cure that. I do see that all these images look pretty good in my editor --- Edit --AND IN SIGIL --, and as noted in my earlier post, rather poor in ADE, with aliasing and pixilation.

@Tex --- My PSP only offers the option to use "Web-Safe-Palette", which is an ugly palette indeed; it changes lots of the pixels to "nearest color", and mostly "nearest" is not nearly "near" enough . The poor picture looks like it is suffering from smallpox.

The other related options are "Interlaced or Non-interlaced", "Transparency", and "Gamma". I have never messed with the gamma option, don't know if that would apply for use in ebooks? I guess I will give scriptpng another shot, but I don't know if anything will change much.

[derangedhermit]

I tried a "foolproof' package to root and flash my Nexus 7. It didn't work, and now my laptop will not "see" my Nexus at all, so I have some troubleshooting to do before I can view these images on i. We fools are so clever.

And now: "Starting with version 10 (released in September 2011), Adobe has removed support of JPEG 2000 from its Photoshop Elements software."

I guess I'll try ImageMagick, and see if it can convert these to 100% jpg.

Since "16 shades" apparently doesn't help on jpg, and doesn't work on some readers, and is far from ideal for other displays, it seems to me that approach can be, at least tentatively, set aside.

I personally want to continue to investigate the use of transparency with palette and grayscale pngs, since I think it could potentially be very nice in some cases.

My meager understanding of png says gamma is best left uninvoked.

My meager understanding of png says non-interlaced is preferable, if/when it matters at all.

I would add "dithering" to the options list, since I have a slider for that. Dithering, IMHO, should also be turned off.

I am surprised that using a web-safe palette is not an option that can be turned off in PSP. In PSE 10, at least for jpgs, it can be turned off.

I just now considered buying PSP X6 to play with these, but it seems to not have much market share, and also I came across this security advisory from Secunia. So maybe later...

[derangedhermit]

Originally Posted by GrannyGrump

@derangedhermit, I wonder if the aliasing you see might not be an inherent result of "cleaning" --- brightening up the background from dark sepia to white is going to lose some of the paler shades. If I use a blur or "softening" tool, the entire image starts to look too soft. I wonder how other folks cure that. I do see that all these images look pretty good in my editor --- Edit --AND IN SIGIL --, and as noted in my earlier post, rather poor in ADE, with aliasing and pixilation.

The degradation of the diagonal strokes comes from downsizing the images.
- The "clean" full size images have very nice lines, at all angles, even under magnification.
- Expanding the levels to as suggested on those images doesn't cause a problem (I tried it, but see for yourself).
- Even the images @ 1200px width show more aliasing. (Looked at yours, and I tried it in PSE also. Similar result, resizing from the original size).
Will look for solution to suggest.

[derangedhermit]

On the provided images (according to Color Quantizer):

021-fullsize-cleaned has only 221 colors
021 256-grayscale has 244 colors (!)
021-256-indexed has only 31 colors

236-fullsize-cleaned has only 247 colors
238-256-grayscale has 246 colors
238-256-indexed has only 31 colors

All of these seem strange to me. I expected the fullsize to have >256 colors and both the grayscale and indexed to have 255 (I suppose 240+ is close enough, but it seems strange).

I can't get either imagemagick or graphicsmagick to read the .jp2 files.

[derangedhermit]

256 shades are sufficient, but 16 is lacking (for displays that can display more shades and have sufficient contrast to make it worthwhile). Within 256 shades, there is probably enough room for useful transparency for these images.

For LCD screens, with sharply delineated pixels, some smoothing/ reduced sharpening on downsizing is needed to reduce aliasing. The need and amount can be seen and demonstrated using the "Color Quantizer" program, among others. The CQ program, for example, does resizing using one of many algorithms, with sharpness sliders.

E-ink screens need all the contrast you can give them, and, due to the softer pixel edges, can stand more sharpening than LCDs.

There are fundamental limits on the contrast that a display can show. Horizontal and vertical lines are one thing, diagonal another. Any diagonal line needs about 3 pixels average thickness or more to do fairly acceptable edges without artifacts. These specific images have many areas where the original image dimensions (e.g. 1800p wide) give good results, but downsizing to 1200 pixels (2 pixels/line) requires care and compromise, and 600 pixels gives an inevitable loss in image quality on any screen (in those areas). In some of the image areas, features are large enough to do better, though. For example, 021 has much area that tolerates shrinking, but 238 has much less.

I believe I now understand that the benefit of the png optimization programs is that one can work, largely without concern, in PNG24A, and then use one of these to produce an output image optimized in size and format (e.g. PNG8 or 8A without unneeded chunks) and close-to-optimally compressed. Many programs, including PSP and PS, do not support the full range of png options, or output optimal pngs, although many browsers and applications display them properly. Tools seem to support PNG8 and 8A (palettized /index color images) more widely than 8-bit grayscale.

I think these, in large, hi-res formats, would benefit from transparency when embedded in a book. I have no experience with transparency, but something like 50% for the white pixels, down to 0% for black, and scaled accordingly according to pixel values in between. I'd give up half the palette shades to implement transparency.

[GrannyGrump]

@derangedhermit -- Thank you so much for spending so much time on this. I do appreciate the good advice you're sharing.

I am going to search for Photoshop plugins, maybe they have something that is wonderful for healing aliasing. Otherwise, I guess I will use Gaussian blur, and see where that gets me.

I have been learning to use the histogram. The process of stretching the histogram is wonderfully easy, but my old version of PSP makes it difficult. The histogram window is really tiny, and I don't have wonderful vision anyway --- you need to move a slider to where the gap ends, and it is almost impossible for me to see that. Plus, the analysis is obtained by mousing over the histogram for a tooltip, and I find it really hard to select a specific shade. I think I will try to find a freeware tool that lets me do this with better visibility. Maybe I will have to go over to the dark side and really start using GIMP. Or maybe I will try out PSE

As far as the number of shades in the palette-- after desaturating and cleaning, that is all the shades that were left. I just confirmed by desaturating the source images again (desaturation=100%, light/dark level not used) :
--- 021 source file originally 91,288 colors, desat = 240 colors
--- 238 source file originally 195,232 colors, desat = 250 colors.

If I use my usual routine of Recoloring, it goes like this:
--- 021 = 91,288; recolored = 65,247; desat = 256
--- 0238 = 185,232; recolored = 91,106; desat = 256

I think cleaning would then remove some of those shades by lightening up some of the gray shades. As I have noted elsewhere, some PSP tools don't work unless you are in 16M colors, so I don't grayscale until the end.

I understand your wish for transparency. I don't know enough about support from various devices and software, but I think it is very spotty. I remember mention of at least one device (Kobo?) that displays png transparency as black! Very ungood. Possibly a more supportable (is that a word?) thing would be tinting the image background to a pastel color or pale gray. But that definitely is missing the cool factor!

In case you still would like a peek at the source files, I am attaching a zip of the two full-size images in JPEG. (I know, I know, saving as jpeg results in more degradation.) Used 1% compression, the closest PSP allows to zero for standard jpeg. The PNGS were quite large, and even though the zip at 17mb was under the upload limit of 20mb, the uploader kept timing out.

[Jellby]

When direct resizing causes visible artifacts (aliasing or moiré patterns), it often helps if the resize is done in several steps. For instance a 50% resize is equivalent to three consecutive 80% resizes (plus one 97.66%, if you want to be exact): 0.5 = 0.8*0.8*0.8*0.9766

This is demonstrated, I hope, with the attached files

021_large, 238_large -> original size, cleaned, cropped, etc.
021_aliased, 238_moire -> direct 50% resize
021_better, 238_better -> multi-step resize, plus levels to remove some of the blurriness (use the burn layer trick again to see that).

I'll leave the color reductions and lossless PNG optimizations for others

[derangedhermit]

The histogram in PSE is hard to see also. I don't know if it can be enlarged.

The key question on the # of shades is why 31 shades for 256-indexed? It's fine if you know why, I just wanted to point it out.

[GrannyGrump]

derangedhermit said:

The key question on the # of shades is why 31 shades for 256-indexed? It's fine if you know why, I just wanted to point it out.

I didn't know, but I do now. That happens if I use Median Cut algorithm to go to 256-indexed. I must have clicked that by mistake, I usually don't use Median Cut, but use Octree, which gives a larger file but better color extrapolation, I think. If I use Octree, the number stays at 221, same as original file.

@Jellby--
The improvement to the aliasing is impressive! Thank you so very much. I never knew that, in fact I understood just the opposite. A long time back on the PSP forum, they told me if I got settings wrong (contrast, recoloring, etc etc) to UNDO and then use the desired settings, and not to repeat the adjustment a second time with new settings to correct the error. I just *assumed* that applied to re-sizing as well. I am so happy to find that was incorrect. (I really do need to get out more, and start getting more educated about image-editing.)

PS -- here is "21-better" compared to "21-BetterWithBlur". Adds a Gaussian blur of lowest possible setting of .5, which does improve it even a bit more, but increases the file size by 1/3. From 297kb to 399kb.

[Jellby]

Originally Posted by GrannyGrump

PS -- here is "21-better" compared to "21-BetterWithBlur". Adds a Gaussian blur of lowest possible setting of .5, which does improve it even a bit more, but increases the file size by 1/3. From 297kb to 399kb.

Personally, I find the blur worse that the remaning aliasing (which could, maybe, be avoided by scaling in even more steps).

By the way, I didn't mention it, but you could also try different resampling methods when scaling. I believe I used one called "Sinc (lanczos3)" in Gimp, which may work better in some cases, but brings its own artifacts.

[GrannyGrump]

@Jellby -- I get options for Weighted Average, Bicubic, Bilinear, Pixel Resize. I use weighted average mostly. My PSP is now almost 10 years old, maybe it IS time for an upgrade.

[derangedhermit]

I think when I was looking at buying PSP, I read they actually replaced the software completely with another package from a company they bought but kept the PSP name, a few major releases ago.

You might have a look at the free Color Quantizer program I mentioned. It has about a dozen resizing algorithms. It only does a few things, but apparently tries to do them well. PNGs only, I think.

[Freeshadow]

For image optimisation you might want to have a look at RIOT