Results 1 to 7 of 7

Thread: Extreme Learned Image Compression

  1. #1
    Member
    Join Date
    Aug 2008
    Location
    Planet Earth
    Posts
    772
    Thanks
    63
    Thanked 270 Times in 190 Posts

    Extreme Learned Image Compression

    Generative Adversarial Networks for Extreme Learned Image Compression
    Our Algorithm (2379 Bytes) vs BPG (2565 Bytes)

    https://data.vision.ee.ethz.ch/aeiri...mecompression/
    https://github.com/Justin-Tan/generative-compression

  2. The Following 3 Users Say Thank You to Sportman For This Useful Post:

    Christian (15th May 2018),Jarek (14th May 2018),schnaader (14th May 2018)

  3. #2
    Member
    Join Date
    Nov 2013
    Location
    Kraków, Poland
    Posts
    645
    Thanks
    205
    Thanked 196 Times in 119 Posts
    WaveOne ( https://arxiv.org/pdf/1705.05823.pdf ) is also on GANs and promises a lot ... but this one is open-source ...

  4. The Following User Says Thank You to Jarek For This Useful Post:

    Sportman (14th May 2018)

  5. #3
    Member
    Join Date
    Jun 2015
    Location
    Switzerland
    Posts
    667
    Thanks
    204
    Thanked 241 Times in 146 Posts
    I'm afraid that the RD-curves of such methods are only good at very low bpp. The success of the lossy algorithms at very low bpp does not directly indicate good performance at high bpp. Very likely PIK is better at 1.5 bpp than a GAN-method at 2.5 bpp.

    If you are interested in the current state of lossy image compression, probably the easiest place to compare formats against BPG is:

    http://wyohknott.github.io/image-for...ry&pik=m&bpg=m

  6. #4
    Member
    Join Date
    Nov 2013
    Location
    Kraków, Poland
    Posts
    645
    Thanks
    205
    Thanked 196 Times in 119 Posts
    WaveOne claims clear superiority on wide range of parameters: http://www.wave.one/icml2017
    Simple heuristic compressors have tiny chance with NN learned on millions of samples of how the actually used images look like ...
    CNN learns wavelet-like decomposition, including feature extraction and even the downsampler - such that it is the most difficult for adversarial NN to distinguish results of compress-decompress from the dataset.

  7. #5
    Member
    Join Date
    Jun 2015
    Location
    Switzerland
    Posts
    667
    Thanks
    204
    Thanked 241 Times in 146 Posts
    Quote Originally Posted by Jarek View Post
    Simple heuristic compressors have tiny chance with NN learned on millions of samples of how the actually used images look like ...
    Possibly true, but no one just figured out how to do that yet, or it needs a few gigabytes of modeling in the decoder to be useful at high quality, or a few minutes of decoding time per image. So far only low image quality solutions have been demonstrated to work, and even them are computationally very expensive.

    I believe we will be compressing lossy images and video with heuristics for another ~15 years if not longer.

  8. #6
    Member
    Join Date
    Nov 2013
    Location
    Kraków, Poland
    Posts
    645
    Thanks
    205
    Thanked 196 Times in 119 Posts
    WaveOne claims great speed ... on GPU, the new paper uses TensorFlow ...
    And generally we can add size of model to optimized parameters ... but sure, it might need 5-10 years to reach wide use in smartphones ... unless some nasty patents on the way (?)

  9. #7
    Member
    Join Date
    Jun 2015
    Location
    Switzerland
    Posts
    667
    Thanks
    204
    Thanked 241 Times in 146 Posts
    Quote Originally Posted by Jarek View Post
    WaveOne claims great speed ... on GPU, the new paper uses TensorFlow ...
    And generally we can add size of model to optimized parameters ... but sure, it might need 5-10 years to reach wide use in smartphones ... unless some nasty patents on the way (?)
    Are you claiming that waveone beats other (heuristic) methods at 1.5 bpp?

    What kind of verification can I conduct to arrive to the same conclusion?

Similar Threads

  1. UCI Image Compression
    By maadjordan in forum Data Compression
    Replies: 5
    Last Post: 19th August 2017, 23:15
  2. Image compression for microcontroller
    By branislav in forum Data Compression
    Replies: 6
    Last Post: 16th November 2015, 23:24
  3. xrle-eXtreme Run Length Encoding
    By algorithm in forum Data Compression
    Replies: 5
    Last Post: 20th April 2015, 23:02
  4. 3d image compression
    By m^2 in forum Data Compression
    Replies: 5
    Last Post: 9th July 2012, 08:52
  5. "Extreme" compression of DNS domains
    By nickety in forum Data Compression
    Replies: 20
    Last Post: 22nd October 2011, 01:20

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •