ReadMe

This source computes the PSNR and the SSIM index, a quality index
defined by Wang, Bovik, Sheihk and Simoncelli that uses a top-down
approach to image quality assessment, based on the following
article of the above authors:

Image Quality Assessment: From Error Measuement to Structural
Similarity, IEEE Transactions on Image Processing, Vol. 13, 2004.

The code was written by Felix Oum and Thomas Richter, and is based
on the IMCO framework, developed by Thomas Richter for the
image compression project of the TU Berlin. For legal restrictions,
and the licence conditions, please check README.Legal.

Thomas Richter,
	January 2007


-----------------------------------------------------------------------

Usage: ssimdiff
        [-CC #] : set the number of CPUs to use in parallel
        [-bl]   : print SSIM separately for each level
        [-exp val ]     : use masking exponent "val", default is 2.0
        [-err file]     : create an image showing the error distribution
        [-lin]          : compute the (m)ssim as linear value, not in dB as by default
        [-nowav]        : compute a single scale ssim instead of multiscale mssim
        infile1:         the original file name.
        infile2:         the distorted file name.
ssimdiff currently understands .ppm and .pgm files.

If called without additional arguments, the source computes the multiscale
SSIM between the input and the output image, and prints the result in dB.
The relation between the "linear" ssim and the logarithmic output in generated
by this tool is

output = -10 log_10 (1 - ssim)

If you think about this for a while, this brings ssim approximately on the
same scale as PSNR.

-CC  	   :	   This option enables multithreading, and speeds up the computation.
     	   	   Due to differences in the final pooling, the output might be slightly
		   different from the single-threaded version.

-bl  	   :	   Print the multi-scale ssim for each level separately

-exp val   :	   Defines an alternative masking exponent for the ssim. By default,
     	   	   this value is 2.0 corresponding to the SSIM as described by Wang,
		   Bovik and Sheihk.

-err file  :	   Generates an error visibility map as predicted by SSIM.
     	   	   This file may require massive contrast or luminance adjustments
		   to generate visible differences on the screen.

-lin	   :	   Output the linear SSIM as described by Wang, Bovik & Sheihk as
	   	   opposed to the dB value.

-nowav	   :	   Compute the single-scale SSIM instead of the multi-scale SSIM.
	   	   By default, a 13/7 filter is used to generate the scales of the
		   multi-scale SSIM. If this option is set, no filtering is performed
		   and the single-scale SSIM is used instead.

If the images are RGB color images, sRGB input is assumed. Note that Wang, Bovik and
Sheihk do not define a color SSIM. In this version, any color input data is first
transformed to YCbCr, and then SSIM is computed independently for each component,
then pooled with weights of 95%, 2% and 3%, respectively. Hence, chroma differences
are mainly ignored. Note that the frequency weights (aka exponents) are independent
of the color channel, which is certainly not correct.

Note that SSIM is not fully specified, thus this implementation may generate results that
are different from the matlab code provided by the original authors. Pooling, filtering
and windowing may be slightly different, giving numerically different results.