H.264 for Image coding

[sharp81]

Hallo Everybody,

As I was searching for information on H.264 , just the intra prediction mode I came across this informative website. I found it very interesting to figure out how the whole process is done. I am trying to learn and maybe implement a part of the H.264 standard. I am interested only in the intra prediction ie I-frames and I want to use it for still images rather than videos. The idea is to check performance of various predictor modes and to select the best possible predictor. And after that I want to compare the blockwise and pointwise compressions. Since I am beginner in the field of Digital Image Processing, I want to first try to do a very simple implementation on this. I was checking the block prediction algorithm but I dont quite understand it works. Could you give me some ideas how to proceed with a complete system design and maybe some coding help? I am considering starting with only 4x4 block implementation

Also every paper mentions using a reference pixels on top and left to predict the block. But how do I select the reference pixels?

Thanks in advance.

[sauvage78]

Technically I can't help but your post reminds me of this:

http://www.bilsen.com/index.htm?http...ilsen.com/aic/

If you don't already know about this, maybe you'll find some interesting info there.

[jmartinr]

Have you read this: http://forum.doom9.org/showthread.ph...97#post1299497?

[sharp81]

yeah, I have read that forum post. But I think just using the intra prediction from H.264 shld give better results than jpeg. Which I dont know yet.

So I want to check if the PSNR and entropy is better.

This page on Advanced Image Coding is very nice and interesting. Thanks for the link. But I would be glad if somebody could explain how I could go abt the implementation.

Thanks.

[Dark Shikari]

Originally Posted by sharp81

yeah, I have read that forum post. But I think just using the intra prediction from H.264 shld give better results than jpeg. Which I dont know yet.

Yes you do. It does.

Feel free to use x264 to confirm if you want

[sharp81]

Hi Dark Shikari, Is there a documentation on the source cod of x264? Its very hard to understand a complete code and algorithm without documentation.

[Dark Shikari]

Originally Posted by sharp81

Hi Dark Shikari, Is there a documentation on the source cod of x264? Its very hard to understand a complete code and algorithm without documentation.

I didn't say to use the source. Just x264 with --tune psnr --preset placebo --frames 1 --crf X, for some value of X, is sufficient to generate your x264-compressed, psnr-optimized images.

For psy optimization, replace psnr with "stillimage".

[sharp81]

is it possible to obtain some explanation on the source code? So that I know how it is working behind the scenes.

[Dark Shikari]

Originally Posted by sharp81

is it possible to obtain some explanation on the source code? So that I know how it is working behind the scenes.

Hop on the x264 IRC on Freenode and ask questions.

[sharp81]

The bilsen project looks interesting. Unfortunately I dont have a delhi to compile it. Is there anything similar available written in C/C++?

[cyberbeing]

Originally Posted by Dark Shikari

Yes you do. It does.

Feel free to use x264 to confirm if you want

http://forum.doom9.org/showthread.php?t=142910
Here was a comparison done between JPEG, JPEG2000, HDPhoto, and x264 back in 2008. It was better than JPEG back then, so it should be even better today.

[sharp81]

i guessed so that it must be definitely better than jpeg. Becoz the prediction algorithm is definitely better. Here the dct is also done on the difference picture rather than on the picture itself like in jpeg.

What I found out about the Intra coding is finding the pixel values from reference is done by intrapolation and extrapolation and taking the weighted average. But I still am not able to get the math here. Is it possible for someone to explain the whole math.

Thanks

[sharp81]

does the Jpeg2000 use the same prediction method as the H.264 and only the part where it differs is it uses DWT instead of DCT. Is that true?

[Dark Shikari]

Originally Posted by sharp81

does the Jpeg2000 use the same prediction method as the H.264 and only the part where it differs is it uses DWT instead of DCT. Is that true?

No; the prediction methods in H.264 are mutually incompatible with large wavelet transforms.

[sharp81]

from what i understand in Jpeg there is no prediction and instead DCT is directly done on the picture which is divided into blocks. How exactly does Jpeg2000 work and how is the H.264 intra mode better than Jpeg2000. Is there a reference source code available?

[foxyshadis]

H.264 JM Reference Source Code
JPEG2000 Reference Source Code

JPEG2000 has no block or pixel prediction. It's a direct wavelet transform of the source, same as JPEG is a direct DCT transform. The primary gain at low bitrates comes from the LL band of the wavelet transform being recursively transformed, and selectively throwing away higher resolution information.

H.264 can predict a block based on DC, up, left, up-left, up-right, and even down-left. Here's a very good h.264 intra primer (pdf) you could look at.

[sharp81]

Thanks a lot for the info. Can you tell me which are the C files, that would be interesting just to consider the intra prediction from the JM software. What's the major difference between this JM software and x264?

[Underground78]

Originally Posted by sharp81

What's the major difference between this JM software and x264?

JM is more something like a proof of concept than an encoder you can use for real daily encoding ... In fact, it is slow as hell and x264 probably has some black magic quality optimizations JM does not have ... But JM source code is probably easier to understand.

[Dark Shikari]

Originally Posted by Underground78

But JM source code is probably easier to understand.

I doubt it. It's practically commentless and has no IRC channel where one can easily talk to the people who wrote it.

[foxyshadis]

I found ffmpeg clear and reasonably concise when I was implementing h.264 parsing in python, trying to minimize outside dependencies. I just used x264 directly for encoding, no way I could compete with that.