jpegmini.com

[Black_Fox]

JPEGmini is a patent-pending photo recompression technology, which significantly reduces the size of photographs without affecting their perceptual quality. The technology works in the domain of baseline JPEG, resulting in files that are fully compatible with any browser, photo software or device that support the standard JPEG format.

JPEGmini is capable of reducing the file size of standard JPEG photos by up to 80% (5X), while the resulting photos are visually identical to the original photos. The JPEGmini algorithm imitates the perceptual qualities of the human visual system, ensuring that each photo is compressed to the maximum extent possible by removing redundancies, without creating any visual artifacts in the process. This enables fully automatic, maximal compression of photos with no human intervention required.

JPEGmini can be implemented in end-user devices such as cellular phones, cameras, and personal computers, to reduce storage requirements for photos, and reduce the time and bandwidth required for uploading the photos to online services. It can also be implemented in image-centric data centers, to reduce Total Cost of Ownership (TCO), including storage, bandwidth, electricity, and floor space.

http://www.jpegmini.com/main/home
http://www.jpegmini.com/main/technology

[Biozynotiker]

I tried jpegmini.com with some pics and this works well.
But I get a way smaller filesize with RIOT.

[Intrinsic]

The examples i checked are 100%/PS save as qual 12 jpegs with no huffman optimisation or progressive encoding. All they do is reduce the quality level to around 83%, enable huffman optimisation and use some custom quantization tables(which is where they gain their size, just like PS uses custom tables too). They don't even bother to try progressive encoding which 99% of the time on photo images will reduce the size further. They don't even tailor the quant tables to the individual table they all use the same ones, on the examples i checked anyways.

[Black_Fox]

Would you mind posting the quant tables here?

[Black_Fox]

The quantization tables are not the same neither for their own examples downloadable from the web, nor for my test imagery.

[Karhunen]

I mess around with AWK to make such tables, I know that the following works with cjpeg "-qtables" parameter:
BEGIN{print "# ((2*(y==x))+(y%x)+((y<7)*(y*x)*(y<7))+(50*(y>6))+( x*2)+((x%7)*2)+(x%(10-y)))\n# Generates Table 0\n\n# Table 0 ( Y channel )"}
END { for (y=1; y<9; y++) {printf "\n";for(x=1;x<9; x++) printf ((2*(y==x))+(y%x)+((y<7)*(y*x)*(y<7))+(50*(y>6))+( x*2)+((x%7)*2)+(x%(10-y))) OFS};
printf "\n\n# Table 1 (Cr channel)\n\n7 9 18 28 43 63 91 128\n9 8 17 23 33 48 68 94\n18 17 31 34 43 58 78 105\n28 23 34 55 63 77 98 126\n43 33 43 63 98 108 128 157";
printf "\n63 48 58 77 108 154 174 204\n91 68 78 98 128 174 239 255\n128 94 105 126 157 204 255 255\n\n#Table 2 (Cb channel)\n\n14 18 46 71 109 161 232 255";
printf "\n18 17 43 59 85 122 173 240\n46 43 80 87 111 148 200 255\n71 59 87 142 160 196 249 255\n109 85 111 160 251 255 255 255\n161 122 148 196 255 255 255 255";
printf "\n232 173 200 249 255 255 255 255\n255 240 255 255 255 255 255 255\n"
}
I invoke this with GNU gawk in my Win32 box in a command shell by gawk -f file.awk NUL where file.awk contains that text above. Obviously I am just messing around with Y channel, the other coeficients were taken from http://x264dev.multimedia.cx/archives/541#more-541

[Intrinsic]

Hmm tables were the same on the few i saw(just their examples you could download, didn't try any of my own files)

I was thinking they could use tricks with custom scans files, i guess you are referring to jpgcrush Karhunen? i've used it's scans files before but just for testing, it can gain you a few more % but haven't examined it extensively, i think it was in python and i don't know much about python

[Karhunen]

http://x264dev.multimedia.cx/archives/541#more-541[/URL]

Basically, I use GNU AWK ( gawk ) as a command line C language, as Aho,Weiner,Kernigan developed as a filter for shell scripts. Works great on DOS 2.0 -> Win7
Not to mention *NIX. That horrible script I used from the command line generates a Quant table and I can randomly mess with the numbers. The only entropy I understand is thermodynamical