Fixes CCD noise in an HDR input file, in EXR format, and saves the result to a
new HDR file, also in EXR format.

The denoising is performed using an algorithm related to:
	Jeff Orchard, Mehran Ebrahimi and Alexander Wong. "Efficient
	Nonlocal-Means Denoising Using the SVD". Proceedings of the IEEE
	Conference on Image Processing. Pages 1732-1735. 2008.
This paper accelerates the original non-local means algorithm by restricting
the search for similar pixels to only a neighborhood of the pixel of interest,
rather than the entire image. The window parameter is the size of this
neighborhood. The weight on each pixel is the Gaussian-like function
exp(-gamma*(d^2)), where d is the Tukey-window weighted L2 distance between
windows of size 2*radius+1 around the pixel of interest, and the candidate
pixel, both projected onto the PCA subspace of the specified dimension.  The
gamma parameter to this weight function is determined by first finding the
descriptor distances between many adjacent pixels (the number of samples is
determined by the pixels parameter), and then choosing gamma such that
differences of the alpha*100th percentile have the weight given by the beta
parameter. The PCA subspace is found using the subspace iteration algorithm,
applied to a covariance matrix calculated from the specified number of pixels,
with termination threshold epsilon.

This program is extremely slow! Be patient.

Usage: hdr_denoise.exe [OPTION...]
  -h, --help                      show help
  -i, --input=INPUT               input file (EXR)
  -o, --output=OUTPUT             output file (EXR)
  -p, --pixels=FLOAT(=10000)      how many pixels should we use?
  -r, --radius=INT(=4)            descriptor radius
  -d, --dimension=INT(=32)        descriptor dimension
  -a, --alpha=FLOAT(=0.2)         alpha parameter
  -b, --beta=FLOAT(=0.5)          beta parameter
  -w, --window=INT(=63)           window side length
  -e, --epsilon=FLOAT(=1e-12)     epsilon parameter