Comparing to an imperfect "ground truth"

[Aneesh Vartakavi <aneeshvartakavi@xxxxxxxxx>]

Dear list,

I'm actually working on image reconstruction in Optical Coherence
Tomography (OCT), but I believe the problem is relevant to this group
as well.

My goal is to compare the performance of a few image reconstruction
algorithms using a few metrics. My problem is that the experimentally
obtained reference image is not free from defects. Intuitively, a
simple metric like MSE (between reference image and reconstructed
image) would not be applicable.

A similar problem could be 'cleaning up' of a noisy speech signal.
Without access to the original, clean speech signal, I'd like to
compare how different algorithms perform with respect to the original,
noisy signal. The differences are apparent to a listener, but how can
I quantify this 'difference' without a clean reference to go by?

Any insight would be helpful!

Regards,

Aneesh Vartakavi

Indian Institute of Science (IISc),
Bangalore, India.