jpegultrascan, an exhaustive JPEG scan optimizer

[MegaByte]

I've just published jpegultrascan, a JPEG lossless recompressor that exhaustively searches all scans to find the smallest combination. Due to its exhaustive nature, it's slower than jpegrescan, which only searches a predetermined set of scan combinations, but not as slow as many PNG recompressors. With my files, I've seen an additional 5-10% file size savings over jpegrescan. It can generate files with a large number of scans, but I haven't run into compatibility problems thus far. It also supports certain obscure JPEG types such as CMYK.

https://github.com/MegaByte/jpegultrascan

As mentioned in the README, I recommend using it with both the IJG version and mozjpeg version of jpegtran (and -t enabled) to get the best compression the fastest.

This version is a Perl script. I have part of a C implementation, but the libjpeg API is a bit of a mess, so I haven't finished it. In addition, in crunching JPEGs in this way, I found several bugs in mozjpeg that I'd rather have resolved before further development.

[Mangix]

Wow you weren't kidding when you said this was slow. Seems to work though.

Isn't mozjpeg an abandoned project?

Does -i create files with the first scan being black and white? fun software bugs with that one...

[Matt Mahoney]

(deleted)

[MegaByte]

Does -i create files with the first scan being black and white? fun software bugs with that one...

The first scan could be monochrome (luminance) or it could be full color depending on whichever combination is smaller. What -i does is allow for DC scans that include more than one but less than the full number of channels. Conversely, -i -1 disallows single channel DC scans, so monochrome wouldn't be possible in that case.

[lorents17]

del

[Malloc Voidstar]

Wow you weren't kidding when you said this was slow. Seems to work though.

Isn't mozjpeg an abandoned project?

Does -i create files with the first scan being black and white? fun software bugs with that one...

https://github.com/pornel/mozjpeg/tree/libjpeg-turbo
Here is version of mozjpeg brought up to date with libjpeg-turbo. Not quite dead yet

edit: moved to https://github.com/mozilla/mozjpeg/tree/libjpeg-turbo

[Jyrki Alakuijala]

I've just published jpegultrascan, a JPEG lossless recompressor that exhaustively searches all scans to find the smallest combination.

Would it be easy to try a lossy, but psychovisually lossless approach based on butteraugli within your recompressor? Butteraugli gives a spatial map indicating where the error exceeds the border of noticeably, so it would be perhaps possible to iterate with it.

[MegaByte]

Good question. The spatial map doesn't lend itself directly to the lossy approach since you can't define different plane splits for spatial sub-sections of the image. However, what you could do is estimate the value of each entire scan to see if you could omit it entirely and still preserve psychovisual integrity. To be fast, you would ideally want a way of performing psychovisual calculation on the scans rather than the entire image, though I'm not sure if that's possible. When I was first testing it out, I wasn't sure I'd be able to iterate through all combinations in a reasonable amount of time, so I was looking into ways to estimate the best combinations by looking at the information content of adjacent bits. I think you'd need to keep track of more than just the best combination though if you were interested in maximizing how many bits you could drop. So all said, yes, I think you could integrate lossiness using a similar approach, but if you wanted the optimal combination for a given quality target, it probably wouldn't come from using the exact algorithm in jpegultrascan. Of course, I imagine you'd get a much better result if you were starting with a lossless image to begin with rather than an already-compressed JPEG.

[lorents17]

At start of a script:

perl jpegultrascan.pl 1.jpg 2.jpg

Gives an error message:

Couldn't read file at jpegultrascan.pl line 350.

I start a script by means of ActivePerl 5.22.1 for Windows (x86). Prompt, please, what I do incorrectly?

[MegaByte]

Do you have jpegtran installed and in your path?

[lorents17]

It has turned out to solve a problem. But there was other problem, the project doesn't optimize the image (the attached file)?

[MegaByte]

That error message is coming from jpegtran. It says your JPEG is corrupted.

[lorents17]

Prompt whether it is possible to expect the version jpegultrascan integrated into jpegtran?

[lorents17]

MegaByte
Tell me, how I can save all versions of a file scan?

[MegaByte]

What do you mean all versions? Each scan or each scan combination? What are you trying to do?

[lorents17]

What is JPEG progressive? Actually it is a number of images where quality of each following image is higher previous the image. JPEG Scan Killer project.

We will take this image



We receive 16 images.



How it is possible to create with the help-scans technology of optimization of JPEG lossy?

JPEG consists of blocks of 8х8 pixels. Perhaps, there is a sense to segment the image and to optimize each part separately. Later to connect parts of the image in the uniform optimized image. Principle of work of the betterjpeg program.

I wanted to learn, your opinion on the matter.

[MegaByte]

That description isn't quite accurate. The *combination* of previous scans ("images") along with the next scan is what makes each image iteration quality higher. Each scan by itself has some arbitrary amount of information depending on its parameters.
As far as I understand the format, you can't choose scans for each block separately; it is a global encoding. However, if you're interested in lossy compression, you could potentially drop information in individual blocks to varying degrees. This is a different problem from optimizing scan choice, though both methods could be combined for even more size gains. It would be most straightforward to perform block-wise optimization and then see which (lossless) scan combination is best suited for that file. In principle, I suppose it's possible that a different combination of block optimization and scan optimization could yield an even smaller file, but searching that space may be very complex. A statistical approach might be possible.

[lorents17]

I thank for your answer!
I wanted to ask you to make the program for optimization with losses with above the described functionality. I think much it will be interesting to test. Unfortunately, my knowledge in given area is very small, and for me this task isn't feasible.

[necros]

C:\Perl\bin>perl jpegultrascan.pl -o . 1.jpg 2.jpg
Use of uninitialized value $_[0] in join or string at jpegultrascan.pl line 92.
Couldn't read file at jpegultrascan.pl line 350.

jpegtran in the same dir
activeperl 5.22.1.2201
is it really possinle to run it under win10?
tried perl jpegultrascan.pl -p . 1.jpg 2.jpg
tried perl jpegultrascan.pl -o . 1.jpg 2.jpg
tried perl jpegultrascan.pl -p jpegtran.exe 1.jpg 2.jpg
tried perl jpegultrascan.pl -o jpegtran.exe 1.jpg 2.jpg

[fhanau]

I just wrote a patch for mozjpeg that further optimizes the huffman encoding. The size difference is really tiny though (about 0.0013%). Just replace the original "jchuff.c" file with this one.

[dado023]

is it possible to create portable version of this tool, with everything needed inside, to make life for us noobs easier

[pico]

deleted

[MegaByte]

Okay, I updated the program to work under Windows. It turned out that there were a few problems (which also would have made jpegrescan not work).
- The binary distribution of jpegtran for Windows doesn't have the same command line.
- Perl pipe open doesn't work under Windows.
- File::Temp has poor random logic when using fork.
Anyway, these were all addressed and it is now working for me (Windows 10, Strawberry Perl 5.24)

[pico]

It seems to work fine now, but for me it fails when trying to use -t saying:

Use of uninitialized value in addition (+) at jpegultrascan.pl line 293.

[MegaByte]

Going to need more info than that...

[pico]

Windows 7 64bit
Portable Strawberry Perl (64-bit) 5.24.0.1-64bit

This command line works:
>perl jpegultrascan.pl 1.jpg a1.jpg

This one does not:
>perl jpegultrascan.pl -t 4 1.jpg a1.jpg

The coutput is:

Calculating sizes of AC scans, plane 0
'-' is not recognized as an internal or external command,
operable program or batch file.
'-' is not recognized as an internal or external command,
operable program or batch file.
'-' is not recognized as an internal or external command,
operable program or batch file.
'-' is not recognized as an internal or external command,
operable program or batch file.


Searching for best AC scan, plane 0
Use of uninitialized value in addition (+) at jpegultrascan.pl line 293.
Use of uninitialized value in addition (+) at jpegultrascan.pl line 293.
Use of uninitialized value in addition (+) at jpegultrascan.pl line 293.
...
...
...

[MegaByte]

It turns out a couple lines were somehow missed in the last commit. Sorry about that; please try again.

[pico]

Works fine now, Thanks!

[MegaByte]

I've created a new script, jpegdropscan, which iteratively shaves off bits and coefficients that contribute least to image fidelity. This initial version uses butteraugli as its quality evaluator. Its logic is not very optimized, so there is potential for future speed up.

As mentioned before, since this is operating on scans, it is whole image-based, so if you wanted to be more block-based, something like guetzli is likely better. The advantage of jpegdropscan is that it does not do any recoding, so you don't lose fidelity to recoded JPEG artifacts.

I've also updated jpegultrascan as I realized certain combinations were missed -- some files now gain an extra 0.05-0.5% improvement. Using jpegultrascan before and after jpegdropscan yields the most size savings. Both can be found in the same repo.

[dado023]

i would like to play with it, are there any Win7 binaries?