jeudi 18 décembre 2014

JPEG-in-TIFF with mixed qualities

Some time ago, I wrote about advanced JPEG-in-TIFF uses in GDAL, i.e. TIFF files compressed with the JPEG codec. Recently, I've got again to that topic when I realized that recent libtiff versions produced files a bit bigger than necessary by repeating the quantization tables in each tile or strip, instead of omitting them in favor of the global quantization tables that are stored in a global location, the JpegTables TIFF tag, shared by all tiles or strips. This issue is now solved.

A quantization table is a set of 64 integer coefficients that are used to divide values coming from the Discrete Cosine Transform, before passing them to the Huffman compression stage and finally reaching the JPEG binary stream. The higher the coefficients, the lower the quality. The decoder must have the quantization table used by the encoder to properly recompute the values needed to do the Inverse Discrete Cosine Transform. The TIFF specification supplement 2 allows the quantization table to be stored in the JpegTables TIFF tag for all tiles or strips. Up to now, I thought this was a requirement, but indeed, reading more closely the specification, I found this was only an option. So there are several legit possibilities :
  1. centralized quantization tables stored in JpegTables TIFF tag, and none in each tile/strip. This is what GDAL generates with its internal version of libtiff (or if using older libtiff, such as 3.8.2, or the to-be-released libtiff 4.0.4) will now produce.
  2. centralized quantization tables stored in JpegTables TIFF tag, and redefined in some tile/strip if needed.
  3. no JpegTables TIFF tag, and quantization tables defined in each tile/strip.
Options 2 and 3 makes it possible to have TIFF files with strips/tiles of different qualities. Actually even option 1 could be used to have up to 2 different qualities. Why 2 ? The reason is that JPEG specification allows a maximum of 4 different quantization tables (also technically it could reference up to 16 tables, but the limit of 4 is actually enforced by libjpeg, so better not play with that!). For RGB imagery compressed in the YCbCr color space, libjpeg uses two quantization tables: one for the Y component and another one for the Cb and Cr components. Hence the capability of storing a set of table for 2 quality settings.

So, as mentionned in my previous article about GeoPackage, it would actually be possible to generate JPEG-in-TIFF with different qualities depending on area(s) of interest, for example medium quality (and high compression) on most areas, and high quality on specific spots. I would expect such variants to have good compatibility with existing readers of JPEG-in-TIFF.

If you are interested in exploring that, you can contact me.

1 commentaire:

  1. JPEG2000 allows variable quality across the image (e.g. defining an area of interest that is high quality, when the rest of the image is lower quality, so that could be an interesting capability when converting from J2K.

    RépondreSupprimer