II image normalization issue

mschuster

Well-known member
Hi Juan,

Please recall the reason for adding the NOISE00 keywords: SA interpolation changes noise statistics. Eg., reference frame noise nearly unchanged (due to near identity homography), but noise of all other frames decreases due to bilinear interpolation. Example:

noiseMRS_c (calibrated)_c_r (aligned)
reference frame2.814e-042.813e-04 (nearly unchanged)
a non-reference frame2.938e-041.833e-04 (decreased)

On a dim project with low scale, I note a similar problem with the scale estimator: bilinear interpolation reduces scale significantly. Results are improper image scaling and loss of SNR.

sqrt(BWMV)_c (calibrated)_c_r (aligned)
reference frame4.762e-044.764e-04 (nearly unchanged)
a non-reference frame4.929e-043.669e-04 (decreased)

Other scale estimators (eg., Sn) exhibit the same issue.

Note: this problem was discovered by my new II SNR tool.
 
Hi Mike,

I'm afraid this is unavoidable. Bilinear interpolation generates very strong aliasing artifacts, especially for small rotations, which act like a low-pass filter. It is not recommended for image registration except in special cases. See this article (I wrote it more than 10 years ago so it is quite outdated and even wrong in some aspects, but still provides good information and interesting comparisons).

To prevent (or improve on) these problems, you can use Lanczos 4/3 or bicubic spline. Or, as an alternative to avoid interpolation, use drizzle integration. Why are you using bilinear interpolation?

As for your new tool, I am very interested. Can we see it soon?
 
Hi Juan,

Update: Lanczos 4/3 end bicubic spline improve but do not eliminate the problem. Drizzle is no better because it measures scale on aligned frames. The non-reference frame has larger scale than the reference before alignment and smaller scale after. Hence a better frame is inappropriately deweighted relative to the reference.

Actually for my projects I don't use bilinear interpolation. This work was just a test to investigate the impact of alignment on II SNR.

The tool likely will remain unreleased. A released version covering all or at least the most used preprocessing variations would require too much development work and require a large amount of time and space during preprocessing. For my projects, a sparse matrix library and multithreading suffice to manage the workload.
 
Last edited:
Back
Top