[HazenBabcock]

Summary

A new Spliner parameter, opt-in, for how the padding margin around each frame is filled. mirror stays the default.

<edge_fill type="string">planar</edge_fill>          <!-- mirror (default) | planar | edge -->
<edge_fill_weight type="float">30.0</edge_fill_weight> <!-- margin pixel = 1/30 of a real pixel -->

 PeakFinderFitter.analyzeImage
   loadImage
-    padArray(frame, margin)                       # always mirror
+    peak_finder.padImage(frame)                   # mirror | planar plane fit | edge copy
   peak_finder.newImage
+  peak_fitter.setScmosTerm(marginVariance(image)) # planar/edge: (k - 1) x fill in the margin
   peak_fitter.newImage
   for each iteration
     backgroundEstimator
-      bg_filter.convolve(image)
+      bg_filter.convolve | PlanarBackground.estimate | realPixelsBackground

• What the fills are. planar is a Gaussian-weighted plane (first-order normalized convolution, done with a MatchedFilterBank). edge copies the outermost pixels straight out.
• What changes with either fill:
  • The background comes from the real pixels only.
  • The margin stays in the fit at 1/k weight. That weight goes in through the C fitter's sCMOS term (MultiFitter.setScmosTerm(), a ctypes write, with no C changes).
• Scope. Only ParametersSplinerSTD has the parameter. The deconvolution finders keep mirroring.

Evidence

The 2016 SMLM challenge's AS-Exp N1 movies, Spliner alone, Jaccard:

• Before (mirror): N1.LD 0.9194, N1.HD 0.7345. On N1.LD it finds only 49-77% of the defocused emitters 450-900 nm from the edge (813 and 519 emitters on the two kinds of side).
• After:
  • planar: 0.9542 and 0.7638.
  • edge: 0.9540 and 0.7703.
  • Recall of those same edge emitters rises to 87-96%.

On 320 x 256 frames with a real cell's background (about twice the frame median along one edge):

• planar keeps the mirror's false positives at the bright edge (53 against 52 on N1.LD); edge adds some (121).
• On 64 x 64 and 320 x 256 frames, planar costs 1.3-1.4x the mirror's time per frame, interleaved on one core. edge costs nothing.

Tests:

• test_edge_fill.py (new): 6 passed. It covers:
  • the edge copy;
  • the plane fit: exact on a plane wherever a fit window reaches, and a plain Gaussian filter in the interior;
  • parameter checks;
  • the margin variance reaching the C fitter each frame;
  • Spliner runs with each fill.
• The full suite: 294 passed.
• On test_spliner.dax, the default and an explicit mirror give byte-identical localizations to master.
• Both fills give byte-identical localizations to the prototype the numbers above come from (with fftw_estimate 1).

Merge Danger

Door: two-way

The parameter is opt-in, the default path gives byte-identical output, and there are no C or file-format changes.

Blast Radius: small

PeakFinderFitter.loadImage() and loadBackgroundEstimate() now pad through peak_finder.padImage(), and analyzeImage() reads peak_finder.edge_fill. Every finder in the repository provides both. fitting.PeakFinder has them (mirror unless the parameter is set), and so does the Spliner deconvolution finder, which gained them in this PR. A custom finder outside the repository that doesn't inherit from fitting.PeakFinder, used with fitting.PeakFinderFitter, would need the same two additions.

🤖 Generated with Claude Code