libbpg/README

BPG Image Encoder and Decoder
-----------------------------

1) Quick introduction
---------------------

- Edit the Makefile to change the compile options (the default compile
  options should be OK for Linux). Type 'make' to compile and 'make
  install' to install the compiled binaries.

- x265 usage: for much increased compression speed (but lower
  quality), you can compile and install x265 and then enable its use
  in the Makefile. x265 supports by default only 8 bits per component
  and does not support monochrome encoding yet (hence no alpha nor
  grayscale images can be encoded with it).

- Emscripten usage: in order to generate the Javascript decoder, you
  must install Emscripten and enable its use in the Makefile.

- An HTML demonstration (with a precompiled Javascript decoder) is
  available in html/index.html (if you use Chrome and want to use
  file:// to access it, launch Chrome with the option
  --allow-file-access-from-files). 

- The BPG file format is specified in doc/bpg_spec.txt.

2) BPG encoder
--------------

The BPG command line encoder is 'bpgenc'. It takes JPEG or PNG images
as input.

- Speed: by default bpgenc uses the JCTVC encoder which has a high
  quality but is slow. If you compiled with x265, you can have a much
  faster encoding with the '-e x265' option. With x265 you can also
  select the encoding speed with the '-m' option (1 = fast, but larger
  image, 9 = slower but smaller image). Warning: x265 does not support
  monochrome (and alpha) yet, so you must use the JCTVC encoder for
  these cases.

- Bit depth: the default bit depth is 8. You can increase it to 10
  ('-b 10' option) to slightly increase the compression ratio. For web
  publishing it is generally not a good idea because the Javascript
  decoder uses more memory.

- Lossless compression is supported as a bonus thru the HEVC lossless
  capabilities. Use a PNG input in this case unless you know what you
  do ! In case of a JPEG input, the compression is lossless related to
  the JPEG YCbCr data, not the RGB data. In any case, the bit depth
  should match the one of your picture otherwise the file size
  increases a lot. By default the lossless mode sets the bit depth to
  8 bits. The prefered color space is set to "rgb". Notes: 
  
    - lossless mode is less tested that the lossy mode but it usually
      gives better results that PNG on photographic images.

    - the JCTVC encoder gives smaller images than the x265 encoder
      with lossless compression.

- There is a difference of interpretation of the quantizer parameter
  (-q option) between the x265 and JCTVC encoder. The default value is
  optimized for the JCTVC encoder, not for x265. We will try to align
  the x265 value to JCTVC in the future.

- By default, the JCTVC encoder is limited to a precision of 12
  bits. You can enable high bit depths (up to 14) by enabling the
  Makefile define: USE_JCTVC_HIGH_BIT_DEPTH. The encoder is sligthly
  slower in this case.

- Color space and chroma format:

    * For JPEG input, the color space of the input image is not
      modified (it is YCbCr, RGB, YCbCrK or CMYK). The chroma is
      subsampled according to the preferred chroma format ('-f'
      option). Images with vertically subsampled chroma are currently
      not supported.

    * For PNG input, the input image is converted to the preferred
      color space ('-c' option). Its chroma is then subsampled
      according to the preferred chroma format.

    * grayscale images are kept unmodified.

- Premultiplied alpha: by default bpgenc uses non-premultiplied alpha
  to preserve the color components. However, premultiplied alpha
  ('-premul' option) usually gives a better compression at the expense
  of a loss in the color components. This loss is not an issue if the
  image is not edited.

- By default, bpgenc does not copy the metadata. You can copy them
  with the '-keepmetadata' option. For JPEG input, EXIF, ICCP and XMP
  are copied. For PNG input, ICCP is copied.

- Objective comparisons: the JCTVC encoder is tuned for PSNR only, not
  for SSIM, so you should use PSNR when making objective comparison
  with other formats. x265 is tuned by default for SSIM.

3) BPG decoder
--------------

The BPG command line decoder is bpgdec. It outputs a PNG or PPM
image. Use a PPM output to get the fastest speed.

- With the '-i' option, you have information about the BPG image (and
no decoded image is output).

- The '-b' option selects the bit depth (8 or 16) of the PNG
  output. It is independent of the internal BPG bit depth.

4) BPG decoding library
-----------------------

BPG images can be decoded in any program with the libbpg
library.

The API is not considered stable yet so that the library is only
provided as a static one.

Currently there is no similar library for encoding so you should
invoke the bpgenc utility.

5) Javascript decoder
---------------------

bpgdec.js is a Javascript decoder supporting the BPG file
format. bpgdec8b.js is a specialized version limited to BPG images
using 8 bits per component. It is a little faster and consumes less
memory (16 MB instead of 32 MB by default, you can change the memory
configuration in the Makefile if you want to handle larger images).

The Javascript decoder substitutes all the <img> tags with a source
having a .bpg extension with a <canvas> tag and decodes the BPG image
into it. Stylesheets are supported (the 'id' and 'class' attributes
are preserved). The 'width' and 'height' attributes are supported only
with pixel units.

The image data is downloaded with the XMLHttpRequest object. So the
BPG images and the BPG Javascript decoder must be in the same domain
unless Cross-Origin Resource Sharing is used.

asm.js gives an interesting speed boost, so we hope that more browsers
will support this Javascript subset.

6) FFmpeg modifications
-----------------------

- Completed support of chroma_format_idc = 0 (monochrome mode).

- Fixed RDPCM support (intra predictions).

- Reduced memory usage for the SAO loop filter.

- Generated the IDCT coefficients dynamically to reduce the code size.

- Added a 'dynamic bit depth' mode where all the bit depths from 8 to
  14 are supported without code duplication but slower decoding.

- Added a modified SPS header to reduce the size of the BPG decoder
  (the solution instead is to generate standard VPS and SPS headers
  from the BPG header).

- Added defines to keep only the HEVC intra code and suppress the
  parsing of all the irrelevant NAL units.

- Stripped FFmpeg from all codecs except HEVC and the necessary
  support code.

7) Licensing
------------

- libbpg and bpgenc are released under the LGPL license (the FFmpeg
  part is under the LGPL, the BPG specific part is released under the
  BSD license).

- bpgenc is released under the BSD license (it includes the JCTVC code
  which is released under the BSD license. The BPG specific part is
  released under the BSD license).

- BPG relies on the HEVC compression technology which may be protected
  by patents in some countries. Most devices already include or will
  include hardware HEVC support, so we suggest to use it if patents
  are an issue.