Detector for ex-JPEG images

[Alexander Rhatushnyak]

Command-line executable for Windows is here:
http://www.imagecompression.info/gralic/Detector.zip
It takes an image in PPM format, prints the "JPEGish coefficient", and then
if it's less than 1.05, "NO, this doesn't look like an ex-JPEG image",
if it's less than 1.10, "Maybe: either an ex-JPEG with smart post-processing, or an image with filtering/processing",
and if it's bigger than 1.10, "YES, this looks like an ex-JPEG image".

The highest value I observed so far is 18.17.
The higher the value, the higher are odds that you'll be able to visually see the artifacts.

Please report false positives and false negatives.

[m^2]

Nice. I have a question. Why don't your tools support bmp images?

[Alexander Rhatushnyak]

Nice. I have a question. Why don't your tools support bmp images?

BMP format is not as simple as PPM because of line alignment and because BMP files may contain RLE-compressed data. Besides, alignment bytes may contain garbage, so sometimes comparing two BMP files for bit-exactness is not as simple as fc/b file1 file2.

An improved detector is here: http://www.imagecompression.info/gralic/Detector.zip
Now the YES margin is at 1.15, but fewer images fall into Maybe.
First version is here, just in case: http://www.imagecompression.info/gralic/Detector10.zip

[Matt Mahoney]

Here is a quick test using lena.bmp and ImageMagick convert (bmp -> jpg -> ppm) with JPEG quality settings from 10 to 100.

  7,544 lena10.jpg
 12,314 lena20.jpg
 16,388 lena30.jpg
 19,984 lena40.jpg
 23,465 lena50.jpg
 27,224 lena60.jpg
 33,203 lena70.jpg
 37,050 lena75.jpg
 43,247 lena80.jpg
 51,883 lena85.jpg
 92,889 lena90.jpg
 99,750 lena91.jpg
106,738 lena92.jpg
118,300 lena93.jpg
134,324 lena94.jpg
154,279 lena95.jpg
187,263 lena96.jpg
227,727 lena97.jpg
279,893 lena98.jpg
354,348 lena99.jpg
407,462 lena100.jpg
786,447 lena.ppm (bmp -> ppm)

C:\tmp>detector lena10.jpg.ppm
10.836 ! YES, this looks like an ex-JPEG image.

C:\tmp>detector lena20.jpg.ppm
 5.459 ! YES, this looks like an ex-JPEG image.

C:\tmp>detector lena30.jpg.ppm
 4.099 ! YES, this looks like an ex-JPEG image.

C:\tmp>detector lena40.jpg.ppm
 3.512 ! YES, this looks like an ex-JPEG image.

C:\tmp>detector lena50.jpg.ppm
 3.284 ! YES, this looks like an ex-JPEG image.

C:\tmp>detector lena60.jpg.ppm
 2.992 ! YES, this looks like an ex-JPEG image.

C:\tmp>detector lena70.jpg.ppm
 2.633 ! YES, this looks like an ex-JPEG image.

C:\tmp>detector lena75.jpg.ppm
 2.440 ! YES, this looks like an ex-JPEG image.

C:\tmp>detector lena80.jpg.ppm
 2.179 ! YES, this looks like an ex-JPEG image.

C:\tmp>detector lena85.jpg.ppm
 1.941 ! YES, this looks like an ex-JPEG image.

C:\tmp>detector lena90.jpg.ppm
 1.266 ! YES, this looks like an ex-JPEG image.

C:\tmp>detector lena91.jpg.ppm
 1.228 ! YES, this looks like an ex-JPEG image.

C:\tmp>detector lena92.jpg.ppm
 1.194 ! YES, this looks like an ex-JPEG image.

C:\tmp>detector lena93.jpg.ppm
 1.164 ! YES, this looks like an ex-JPEG image.

C:\tmp>detector lena94.jpg.ppm
 1.126   Maybe: either an ex-JPEG with smart post-processing, or an image with filtering/processing

C:\tmp>detector lena95.jpg.ppm
 1.048   NO,  this doesn't look like an ex-JPEG image.

C:\tmp>detector lena96.jpg.ppm
 1.025   NO,  this doesn't look like an ex-JPEG image.

C:\tmp>detector lena97.jpg.ppm
 1.008   NO,  this doesn't look like an ex-JPEG image.

C:\tmp>detector lena98.jpg.ppm
 1.006   NO,  this doesn't look like an ex-JPEG image.

C:\tmp>detector lena99.jpg.ppm
 1.006   NO,  this doesn't look like an ex-JPEG image.

C:\tmp>detector lena100.jpg.ppm
 1.009   NO,  this doesn't look like an ex-JPEG image.

C:\tmp>detector lena.ppm
 1.011   NO,  this doesn't look like an ex-JPEG image.

The JPEG quantization coefficients (0 1 2 3...31...63) for quality 93 through 95 are:

93 (YES)
Q0 = 2 2 2 2...7...14
Q1 = 2 3 3 3...14...14

94 (MAYBE)
Q0 = 2 1 1 2...6...12
Q1 = 2 2 2 3...12...12

95 (NO)
Q0 = 2 1 1 1...5...10
Q1 = 2 2 2 2...10...10

For quality 100, all quantization values are 1. (Q0 = grayscale, Q1 = color).

Also, of note, from maximumcompression.com, I converted a10.jpg and rafale.bmp to ppm:

C:\tmp>detector a10.ppm
 1.005   NO,  this doesn't look like an ex-JPEG image.

C:\tmp>detector rafale.ppm
 1.156 ! YES, this looks like an ex-JPEG image.

a10.jpg has coefficients that look like a quality setting about 98.5

Q0 = 1 1 1 1...1...3
Q1 = 1 1 1 1...3...3

Also, I tested the Kodak set, and all files are detected as NO.

[Alexander Rhatushnyak]

Matt, thank you for testing!
I guess it's obvious that a smart post-processing can eliminate all JPEG-specific artifacts. Looks like ImageMagick applies a smart post-processing step (unlike some other image viewers and convertors).
Nevertheless, here's an improved version: http://www.imagecompression.info/gralic/Detector.zip

Margins are at 1.05 and 1.15 as before, but a10.ppm is detected as Maybe, Rafale is YES with much higher confidence, and among 24 Kodak images only 4 are detected as NO.

There's a bigger Kodak set here: ftp://ftp.kodak.com/www/images/ -- 108 images. If conversion (PCD=>PPM) is done by ImageMagic, 31 of 108 are detected as YES, 49 as NO, 28 as Maybe, if conversion is done by Display ( ftp://ftp.elf.stuba.sk/pub/pc/graph/dispt5.exe ) 33 as YES, 47 as NO, 28 as Maybe. But if it's done by IrfanView 4.28 ( http://www.irfanview.com ) , all 108 are YES.

Among 32 images used in the FLIC/others comparison ( http://encode.su/threads/1222-FLIC-a...age-compressor ), only one is YES, and four are Maybe.

[Matt Mahoney]

How are you doing the detection? I would guess convert the image to quality 100 JPEG (all quantization coefficients 1) in each of the 3 common sampling rates (1x1, 2x1, 2x2), then look at the 128 frequency distributions for repeating patterns.

[Alexander Rhatushnyak]

How are you doing the detection? I would guess convert the image to quality 100 JPEG (all quantization coefficients 1) in each of the 3 common sampling rates (1x1, 2x1, 2x2), then look at the 128 frequency distributions for repeating patterns.

No, that would be much slower, and code size would be bigger, BTW. Hopefully someone will create a thorough ex-JPEG detector caring little about speed. Ex-JPEGs can be compressed much better once you know which of your images are ex-JPEGs and apply a special algorithm.

Previous version of my detector was too much influenced by Kodak images, the corrected version is at http://www.imagecompression.info/gralic/Detector.zip

[m^2]

False negative.

[Alexander Rhatushnyak]

False negative.

Thanks for reporting!
The latest version 1.4 simply says "Maybe. Try better methods" instead of speculating why.
I believe some of the better methods will try a number of rotation variants before analyzing the image.

[Alexander Rhatushnyak]

More false negatives:
http://photojournal.jpl.nasa.gov/tiff/PIA13760.tif
http://photojournal.jpl.nasa.gov/tiff/PIA13762.tif
http://photojournal.jpl.nasa.gov/tiff/PIA13790.tif
http://photojournal.jpl.nasa.gov/tiff/PIA13793.tif
http://photojournal.jpl.nasa.gov/tiff/PIA13802.tif
http://photojournal.jpl.nasa.gov/tiff/PIA13934.tif
I wonder if you'll be able to visually see the artifacts.

[Ethatron]

There's a bigger Kodak set here: ftp://ftp.kodak.com/www/images/ -- 108 images.

Gallery: http://cdb.paradice-insight.us/?class=7

[Alexander Rhatushnyak]

Also, of note, from maximumcompression.com, I converted a10.jpg and rafale.bmp to ppm:

C:\tmp>detector a10.ppm
 1.005   NO,  this doesn't look like an ex-JPEG image.

C:\tmp>detector rafale.ppm
 1.156 ! YES, this looks like an ex-JPEG image.

...
Also, I tested the Kodak set, and all files are detected as NO.

The latest improved version prints "Yes" for both images from maximumcompression.com, 7 of 24 Kodak images, and two of six images from my previous message.
http://www.imagecompression.info/gralic/Detector.zip

[Alexander Rhatushnyak]

Version 2.2 is able to detect the majority of ex-JPEGs created with quality 100.
Some of the rotated ex-JPEGs may be detected.
http://www.imagecompression.info/gralic/Detector.zip

An example of an image detected as YES but rather close to Maybe:
http://cdb.paradice-insight.us/corpo...C_6543.ppm.bz2
Magnified: http://www.imagecompression.info/gra..._magnified.png
Artifacts (squares) can be seen in the lower right corner, but not in the upper left.

[Matt Mahoney]

It would be interesting if a tool could guess the original JPEG quantization values and be used to convert to JPEG and back losslessly.

[Piotr Tarsa]

Final quantization isn't the only source of rounding errors in JPEG. DCT and even color space transform are lossy. It's impossible to get back the original JPEG. But probably you can get very close. Maybe then encoding guessed JPEG + difference would bring us some space savings, who knows?

You would need first to collect some widely used algorithms for color space transforms and DCT transforms, as there are many, some are less accurate, some are more accurate. Then use some brute force, like precomp does for deflate streams, to find the algorithm used.

[Matt Mahoney]

Yes, or encode as a JPEG plus pixel differences.

[m^2]

Alex, I have a couple of images for you. Not sure if you're interested in small stuff, but I believe they are all false positives.

[Alexander Rhatushnyak]

Alex, I have a couple of images for you. Not sure if you're interested in small stuff, but I believe they are all false positives.

Why do you believe so?
The biggest image has horizontal and vertical edges in low-contrast areas (the left arrow in the attached magnified fragment). And there's an "antenna" on top of 2 (the right arrow).

[m^2]

You're the expert. I looked at the image closeups and didn't see anything suspicious, that's all.

[Alexander Rhatushnyak]

All of WebP's sample images are ex-JPEGs, here's a magnified section of 9_webp.png: http://imagecompression.info/gralic/webp9magnified.png

[Matt Mahoney]

Yes. The original JPEGs are included in the download. Image 9 is the lowest quality of the set, looking at the quantization tables.

C:\tmp\webp-samples>jpegtest *.jpg
10_original.jpg
       0 SOI [0]
       0 SOI
       4 APP0 [16] JFIF ver 1.2 units 1, density 180 by 180
      22 APP1 [6274] Exif  II*               å           î
    6298 APP13 [6728] Photoshop 3.0 8BIM %                      8BIM φ
   13028 APP1 [5067] http://ns.adobe.com/xap/1.0/ <?xpacket begin=''
   18097 APP2 [3160] ICC_PROFILE      HLino    mntrRGB XYZ  ╬       1
   21259 APP14 [14] Adobe d@
   21275 DQT [132]
    Q0 = 1 1 1 1...1...3
    Q1 = 1 1 1 1...3...3
   21409 Baseline DCT [17]  988x1208, 8 bit:(C1=1x1,Q0)(C2=1x1,Q1)(C3=1x1,Q1)
   21428 DRI [4] 124
   21434 DHT [418]
    DC0: (3: 07 08 06 05 04 09)(4: 03 0A)(5: 02 01 00),1,0,0,0,0,0,0,0,0,0,0
    DC1: (3: 06 05 04 03 07 02)(4: 08 01 09)(5: 00),1,1,0,0,0,0,0,0,0,0,0
    AC0: (2: 01 02)(3: 03)(4: 04 11 05)(5: 12 06 21 07),1,3,3,2,3,3,3,2,6,9,117
    AC1: (2: 01 02)(3: 03)(4: 11 04 21)(5: 12 05),4,4,3,5,4,4,4,6,6,5,109
   21854 SOS [12] Coef 0-63, bits 0-0 of (C1 DC0 AC0)(C2 DC1 AC1)(C3 DC1 AC1)
 1093966 EOI [0]

 1093968 end of file

1_original.jpg
       0 SOI [0]
       0 SOI
       4 APP0 [16] JFIF ver 1.1 units 1, density 200 by 200
      22 DQT [67]
    Q0 = 3 2 2 3...10...20
      91 DQT [67]
    Q1 = 3 4 4 5...20...20
     160 Baseline DCT [17]  800x 440, 8 bit:(C1=2x2,Q0)(C2=1x1,Q1)(C3=1x1,Q1)
     179 DHT [31]
    DC0: (2: 00)(3: 01 02 03 04 05)(4: 06)(5: 07),1,1,1,1,0,0,0,0,0,0,0
     212 DHT [181]
    AC0: (2: 01 02)(3: 03)(4: 00 04 11)(5: 05 12 21),2,4,3,5,5,4,4,0,0,1,125
     395 DHT [31]
    DC1: (2: 00 01 02)(3: 03)(4: 04)(5: 05),1,1,1,1,1,1,0,0,0,0,0
     428 DHT [181]
    AC1: (2: 00 01)(3: 02)(4: 03 11)(5: 04 05 21 31),4,3,4,7,5,4,4,0,1,2,119
     611 SOS [12] Coef 0-63, bits 0-0 of (C1 DC0 AC0)(C2 DC1 AC1)(C3 DC1 AC1)
  136778 EOI [0]

  136780 end of file

3_original.jpg
       0 SOI [0]
       0 SOI
       4 APP0 [16] JFIF ver 1.2 units 0, density 100 by 100
      22 APP12 [37] Ducky       N           5 . 1 . 2
      61 APP14 [14] Adobe d└
      77 DQT [132]
    Q0 = 3 2 2 2...6...14
    Q1 = 3 3 3 5...14...14
     211 Baseline DCT [17]  450x 288, 8 bit:(C1=1x1,Q0)(C2=1x1,Q1)(C3=1x1,Q1)
     230 DHT [183]
    DC0: (2: 04 05)(3: 02 03 06)(5: 00 01 07),1,1,0,0,0,0,0,0,0,0,0
    DC1: (2: 02 03)(3: 00 01 04)(4: 05)(5: 06),1,1,0,0,0,0,0,0,0,0,0
    AC0: (2: 01 02)(3: 03)(4: 11 04)(5: 00 21 31 12 05),2,4,3,5,5,5,6,5,3,3,5
    AC1: (2: 00 01)(3: 11 02)(4: 03)(5: 21 31 12 04),1,2,3,6,6,2,1,4,3,1,0
     415 SOS [12] Coef 0-63, bits 0-0 of (C1 DC0 AC0)(C2 DC1 AC1)(C3 DC1 AC1)
   46766 EOI [0]

   46768 end of file

4_original.jpg
       0 SOI [0]
       0 SOI
       4 APP1 [18538] Exif  MM *                 &#198;           ?
   18544 APP2 [88] MPF MM *      ░       0100░           ░          2
   18634 DQT [132]
    Q0 = 1 1 2 2...8...15
    Q1 = 1 2 2 3...15...15
   18768 Baseline DCT [17] 2048x1536, 8 bit:(C1=2x1,Q0)(C2=1x1,Q1)(C3=1x1,Q1)
   18787 DHT [418]
    DC0: (2: 00)(3: 01 02 03 04 05)(4: 06)(5: 07),1,1,1,1,0,0,0,0,0,0,0
    AC0: (2: 01 02)(3: 03)(4: 00 04 11)(5: 05 12 21),2,4,3,5,5,4,4,0,0,1,125
    DC1: (2: 00 01 02)(3: 03)(4: 04)(5: 05),1,1,1,1,1,1,0,0,0,0,0
    AC1: (2: 00 01)(3: 02)(4: 03 11)(5: 04 05 21 31),4,3,4,7,5,4,4,0,1,2,119
   19207 SOS [12] Coef 0-63, bits 0-0 of (C1 DC0 AC0)(C2 DC1 AC1)(C3 DC1 AC1)
  795866 EOI [0]

  795870 SOI [0]
  795870 SOI
  795874 APP1 [116] Exif  MM *                 >           F (
  795992 DQT [132]
    Q0 = 1 1 1 1...2...5
    Q1 = 1 1 1 1...4...5
  796126 Baseline DCT [17]  640x 480, 8 bit:(C1=2x1,Q0)(C2=1x1,Q1)(C3=1x1,Q1)
  796145 DHT [418]
    DC0: (2: 00)(3: 01 02 03 04 05)(4: 06)(5: 07),1,1,1,1,0,0,0,0,0,0,0
    AC0: (2: 01 02)(3: 03)(4: 00 04 11)(5: 05 12 21),2,4,3,5,5,4,4,0,0,1,125
    DC1: (2: 00 01 02)(3: 03)(4: 04)(5: 05),1,1,1,1,1,1,0,0,0,0,0
    AC1: (2: 00 01)(3: 02)(4: 03 11)(5: 04 05 21 31),4,3,4,7,5,4,4,0,1,2,119
  796565 SOS [12] Coef 0-63, bits 0-0 of (C1 DC0 AC0)(C2 DC1 AC1)(C3 DC1 AC1)
  936603 EOI [0]

  936605 end of file

6_original.jpg
       0 SOI [0]
       0 SOI
       4 APP1 [37930] Exif  II*               &#198;           ▓           ╩
   37936 DQT [132]
    Q0 = 2 2 2 2...7...14
    Q1 = 2 3 3 3...14...14
   38070 Baseline DCT [17] 3264x2448, 8 bit:(C1=2x1,Q0)(C2=1x1,Q1)(C3=1x1,Q1)
   38089 DHT [418]
    DC0: (2: 00)(3: 01 02 03 04 05)(4: 06)(5: 07),1,1,1,1,0,0,0,0,0,0,0
    AC0: (2: 01 02)(3: 03)(4: 00 04 11)(5: 05 12 21),2,4,3,5,5,4,4,0,0,1,125
    DC1: (2: 00 01 02)(3: 03)(4: 04)(5: 05),1,1,1,1,1,1,0,0,0,0,0
    AC1: (2: 00 01)(3: 02)(4: 03 11)(5: 04 05 21 31),4,3,4,7,5,4,4,0,1,2,119
   38509 SOS [12] Coef 0-63, bits 0-0 of (C1 DC0 AC0)(C2 DC1 AC1)(C3 DC1 AC1)
 1867350 EOI [0]

 1867802 end of file

7_original.jpg
       0 SOI [0]
       0 SOI
       4 APP0 [16] JFIF ver 1.1 units 1, density 314 by 314
      22 APP13 [80] Photoshop 3.0 8BIM       3  Z   %G         x  OLYM
     104 APP2 [576] ICC_PROFILE      0ADBE    mntrRGB XYZ  ╨       3 ;
     682 APP1 [657] Exif  MM *                             &#229;
    1341 DQT [67]
    Q0 = 2 1 1 2...6...11
    1410 DQT [67]
    Q1 = 2 2 2 3...11...11
    1479 Baseline DCT [17] 3648x2736, 8 bit:(C1=2x2,Q0)(C2=1x1,Q1)(C3=1x1,Q1)
    1498 DHT [31]
    DC0: (2: 00)(3: 01 02 03 04 05)(4: 06)(5: 07),1,1,1,1,0,0,0,0,0,0,0
    1531 DHT [181]
    AC0: (2: 01 02)(3: 03)(4: 00 04 11)(5: 05 12 21),2,4,3,5,5,4,4,0,0,1,125
    1714 DHT [31]
    DC1: (2: 00 01 02)(3: 03)(4: 04)(5: 05),1,1,1,1,1,1,0,0,0,0,0
    1747 DHT [181]
    AC1: (2: 00 01)(3: 02)(4: 03 11)(5: 04 05 21 31),4,3,4,7,5,4,4,0,1,2,119
    1930 SOS [12] Coef 0-63, bits 0-0 of (C1 DC0 AC0)(C2 DC1 AC1)(C3 DC1 AC1)
 3587689 EOI [0]

 3587691 end of file

9_original.jpg
       0 SOI [0]
       0 SOI
       4 APP1 [19120] Exif  MM *                 ▓       "   ╚
   19126 APP2 [117] FPXR        &#163; /   E x t e n s i o n   L i s t
   19245 APP2 [427] FPXR         ■*       `aVT┴╬ &#224;S ? &#237;∙[     `aVT┴╬ &#224;
   19674 APP2 [61623] FPXR         ■*      1   └o╨ ╜  `&#249; &#237;&#199;    1   └o╨ ╜
   81299 APP2 [17818] FPXR       ≡&#191;W-Tφ =√?═&#196;≡j≥&#198;√─[ δ)╟≈Dp╚▀&#252;ατ≡ ?╟ k
   99119 APP9 [15571]
  114692 DQT [132]
    Q0 = 5 5 5 5...8...23
    Q1 = 5 5 5 5...8...24
  114826 Baseline DCT [17] 3072x2304, 8 bit:(C1=2x2,Q0)(C2=1x1,Q1)(C3=1x1,Q1)
  114845 DHT [418]
    DC0: (2: 00)(3: 01 02 03 04 05)(4: 06)(5: 07),1,1,1,1,0,0,0,0,0,0,0
    DC1: (2: 00 01 02)(3: 03)(4: 04)(5: 05),1,1,1,1,1,1,0,0,0,0,0
    AC0: (2: 01 02)(3: 03)(4: 00 04 11)(5: 05 12 21),2,4,3,5,5,4,4,0,0,1,125
    AC1: (2: 00 01)(3: 02)(4: 03 11)(5: 04 05 21 31),4,3,4,7,5,4,4,0,1,2,119
  115265 DRI [4] 1728
  115271 SOS [12] Coef 0-63, bits 0-0 of (C1 DC0 AC0)(C2 DC1 AC1)(C3 DC1 AC1)
 1175640 EOI [0]

 1175642 end of file


C:\tmp\webp-samples>dir
 Volume in drive C is OS
 Volume Serial Number is 66E6-426E

 Directory of C:\tmp\webp-samples

10/01/2010  11:27 PM    <DIR>          .
10/01/2010  11:27 PM    <DIR>          ..
10/01/2010  11:27 PM             6,148 .DS_Store
09/30/2010  11:22 AM         1,093,968 10_original.jpg
09/30/2010  11:21 AM         1,575,804 10_webp.png
09/30/2010  11:22 AM           136,780 1_original.jpg
09/30/2010  11:21 AM           576,422 1_webp.png
09/30/2010  11:22 AM            46,768 3_original.jpg
09/30/2010  11:21 AM           170,569 3_webp.png
09/30/2010  11:22 AM           936,605 4_original.jpg
09/30/2010  11:21 AM         2,707,632 4_webp.png
09/30/2010  11:22 AM         1,867,802 6_original.jpg
09/30/2010  11:21 AM         7,734,863 6_webp.png
09/30/2010  11:22 AM         3,587,691 7_original.jpg
09/30/2010  11:21 AM        13,722,548 7_webp.png
09/30/2010  11:22 AM         1,175,642 9_original.jpg
09/30/2010  11:21 AM         4,819,986 9_webp.png
              15 File(s)     40,159,228 bytes

[Karhunen]

Here is a rendered wood texture from POV 3.7, which comes up 1.203 ( positive ):

The attached POV and Portable aNyMap in zip file below

Are there settings for the detector that enumerate the regions that flag as compression artifacts?

[Alexander Rhatushnyak]

You are right, this image is a false positive.
Hopefully someone will create a better detector. It may be much slower, I'd guess speed is not an issue for the majority of practical applications.