Subject:
Update to the FITS Tiled Image Compression Convention
From:
"Pence, William D. (GSFC-6601)" <william.d.pence@nasa.gov>
Date:
Fri, 18 Sep 2009 11:08:49 -0500
To:
FITSBITS <fitsbits@nrao.edu>

This is to announce that a revised 2.1 version of the document 
describing the FITS Tiled Image Compression Convention has been placed 
on the FITS Support office web site at

http://fits.gsfc.nasa.gov/registry/tilecompression.html

The main changes in this version of the document are in Section 4 and in 
Appendix A, which describe a "subtractive dithering" technique for 
quantizing floating point FITS images into scaled integer values, prior 
to compressing the integers with a choice of algorithm (usually Rice). 


The problem with the previous simple quantization technique, in which 
all the image pixels are simply quantized into a fixed grid of discrete 
values, is that this can bias estimates of the background flux level in 
an image when using relatively coarse quantization.  For example, if the 
actual background level in an image has a mean value of 0.4, but all the 
pixels are quantized into levels 0, 0.5, 1.0, ..., then the background 
estimate (e.g., using the median of the pixel values, or the mean after 
several N sigma rejection cycles) will be biased towards the nearest 
quantized level (0.5 in this example).

The subtractive dithering technique (which has been widely used in 
signal processing applications) reduces this bias problem by randomly 
dithering the zero point of the quantization grid for each pixel.  So, 
continuing with previous example, all the pixels are still quantized 
into levels that have a spacing of 0.5, but the whole grid is shifted up 
or down by a random amount on a pixel by pixel basis.  Thus, the median 
of the quantized pixel values will more closely match that of the 
original image.  The big advantage of this subtractive dithering 
technique is that the randomization is accomplished without introducing 
any extra noise into the image.

In collaboration with Rick White (STScI) and Rob Seaman (NOAO),
this new dithering technique has been implemented in the latest 3.20 
version of the CFITSIO library and in the new versions of the fpack and 
funpack FITS image compression utilities.  It should be noted that this 
implementation is backwards compatible with any existing FITS images 
that were quantized using the previous quantization technique.

Bill Pence