EPIC vs JPEG

[Rob Harper]

     %%%%%%%%%%%%  CLIP & RECOMMENDATION %%%%%%%%%%%%%%%%

I have been comparing the performance of a discrete-cosine transform
image-compression system (cjpeg/djpeg from the Independent JPEG Group)
against a sub-band pyramid coder (epic/unepic from Eero P. Simoncelli
and Edward H. Adelson) of MIT).  Using the same image (a 256x256
grayscale rendering of the famous "Lenna"), which I have compressed
and decompressed at many different levels, I determine subjectively
that their image-processing performance is comparable, with one
important difference: the pyramid coder produces no mosaic effects,
even at very high levels of compression.  For example:

    epic lenna.raw -o lenna.E -l 4 -b 43

compresses `lenna.raw' (65536 bytes) down to 4960 bytes, while

    rawtopgm 256 256 lenna.raw | cjpeg -Q 24 > lenna.jpg

knocks `lenna.raw' down to 4973 bytes.  When expanded (by djpeg and
unepic respectively) and viewed (in my case, with the `xv' viewer),
they are both remarkably good reconstructions, with similar and
surprisingly mild artifacts.  However, the JPEG system produced a
mosaic-like effect across many portions of the restored image.  The
pyramid coder produces no such artifacts that I can detect.

It is possible that a future version of the JPEG system (when
cross-block smoothing is implemented in the decoder) will mitigate the
mosaic effect.

If you wish to experiment with these systems, you may obtain C source
via anonymous FTP from these locations:

   JPEG: ftp.uu.net (137.39.1.9), in graphics/jpeg/jpegsrc.v?.tar.Z
         (new versions are posted from time to time)

   EPIC: whitechapel.media.mit.edu (18.85.0.125), in pub/epic.tar.Z

The "Lenna" image is available as part of the EPIC package, where it
is named "test_image."

Jim McCauley
jem at cup.hp.com
----------