(1) Mask the raw detector data, and compress in the Y-direction. In EWS and MIA this is done for speed in a C-language program called oir1dCompressData, which is called from IDL with the data files and mask files as input.
(2) Subtract the two MIDI interferometric channels. This removes about 90% of the background.
(3) Decide which scans were off-Fringe and use them to estimate the background noise.
(4) For the on-Fringe scans, estimate the mean square power (per wavenumber bin), subtract the off-Fringe power, and if you have courage, take the square root.
(5) Estimate from the non-interferometric, photometric exposures, what the total flux F(k) in each wavelength bin should be, and divide this into the reduced correlated flux F(k)V(k) to estimate V(k), the visibility.
(6) Repeat this process for a calibrator to estimate the instrumental visibility loss. Some disadvantages of this procedures are: You have to have a good estimate of the off-Fringe noise power; if the nominal signal/noise in a single wavenumber bin is small: (S/N)<1, then the S/N of the correlated flux after M samples is only (S/N)^2&radic(M).