Lossless JPEG

[ImpulseAdventure]

Decode lossless JPEG files

[2010/10/16] First created at https://sourceforge.net/p/jpegsnoop/feature-requests/24/

[MarcusJohnson91]

I'm trying to understand how the DHT table is composed for Lossless JPEG, it'd be great if you could add support for the SOS header in SOF3 mode

Here is a sample image I'm using: https://www.dropbox.com/s/2vxl1h6wmo97lse/Lossless.jpg?dl=0

[lclevy]

https://github.com/lclevy/libcraw2/blob/master/docs/cr2_lossless.pdf

[ImpulseAdventure]

Thank you Laurent -- that is an excellent summary you created.

[MarcusJohnson91]

I'm not understanding how a code is mapped to a pixel value?

from the image I posted earlier, the Huffman table is: 2[4, 5, 6] 4[2, 3, 7] 5[0] 6[1] 7[8] 8[A] 9[9] 10[B] 11[C]

where the number outside of the brackets is the number of bits, and the number(s) inside the brackets are the values.

how do I find out which pixel 11[C] maps to?

11[C] would be what binary string, and what would that binary string map to pixel wise?

[ImpulseAdventure]

Hello Marcus!

First, I must caveat this by saying that I have not spent much time decoding lossless JPEG images in detail, but perhaps this may help you get started with your question.

Unlike BMP-formatted images, a singular JPEG Huffman code doesn't directly map to a pixel value. I would recommend starting with a read through my JPEG Huffman coding tutorial first, as the differential coding scheme is covered in moderate detail with a simple two-block example. You will find that each Huffman code basically describes the "change in value" from a previous block or predicted block.

More specific to your question: Your Huffman code "11[C]" is mapped as follows (according to JPEGsnoop with DHT Expand option enabled):

     Codes of length 11 bits:
      11111111110 = 0C                   (Total Len = 23)

Once you encounter the bit sequence 11111111110 during decode, the next 12 bits will represent a signed "difference" value. In lossy compression modes, this difference value would typically describe the change in the luminance (since you have a single-channel image) for the DC coefficient versus the previous "block" or the change in an AC frequency coefficient. Thus one doesn't really convert a single coefficient (or huffman code) into a "pixel value". The DC / AC coefficients are based on a frequency-domain representation of your image (see FFT), which is quite removed from the concept of a pixel value.

In the case of lossless JPEG images, the changes are based not on the previous DCT block but a predicted value from neighboring blocks. See https://en.wikipedia.org/wiki/Lossless_JPEG

Hope that helps!