ir-blob C++ → Python parity checklist

Audit of legacy ir-blob (/legacycode/code/ir-tools/ir-blob.cxx) mapped to Python in nornir_imageregistration/nornir_imageregistration/blob_filter.py.

1. CLI and pipeline defaults

#	C++ (ir-blob.cxx)	Python / buildmanager	Parity
1.1	-sh 1 shrink on load (std_tile, default 1)	Input already at target downsample in pipeline	OK
1.2	-r radius (default 2 in tool, 9 in Pipelines.xml)	radius / -Radius	OK
1.3	-median ITK MedianImageFilter radius (default 0, 7 in IDoc pipeline)	median_radius	Port (nearest boundary)
1.4	-max threshold (default 3)	max_value	OK
1.5	-mask optional; else std_mask (all valid)	mask_path or all-true	OK

2. Median prefilter (median(), common.hxx L6127)

#	C++	Python target	Parity
2.1	ITK MedianImageFilter radius [median, median] (Neumann edges)	scipy.ndimage.median_filter(..., size=2r+1, mode="nearest")	Port
2.2	Skip when median_radius == 0	Same	OK

3. Local variance (calc_variance, ir-blob.cxx L63–149)

#	C++	Python target	Parity
3.1	Window width/height (2r+1) shifted to stay in bounds	_window_bounds_1d + per-pixel slices	Port
3.2	Mean/variance over mask-true samples only	Same	Port
3.3	mass==0 → variance pixel left at float_max sentinel	_VARIANCE_SENTINEL	Port
3.4	Population variance sum((p-mean)²)/mass	Same	Port

4. Blob enhancement (enhance_blobs, L251–420)

#	C++	Python target	Parity
4.1	Collect valid variance samples (!= float_max), qsort, median at [count/2]	_global_variance_median	Port
4.2	metric = min(threshold, (median+1)/(v+1)) for valid pixels	_enhance_blobs	Port
4.3	Invalid pixels → mean metric over valid pixels	Not zero	Port
4.4	Do not divide by threshold before normalize	Removed early / max	Port

5. Output normalization (normalize(image,1,1,0,255,mask), common.hxx L6198–6321)

#	C++	Python target	Parity
5.1	Masked mean/sigma (StatisticsImageFilterWithMask, unbiased variance)	_normalize_with_mask	Port
5.2	(x - mean) / sigma on full image	Same	Port
5.3	Clip to [-3, 3]	_clip	Port
5.4	Linear remap global min/max → [0, 255]	_remap_min_max_inplace	Port
5.5	Save 8-bit PNG (save<native_image_t>)	SaveImage(..., bpp=8)	OK

6. Golden fixtures

Legacy PNG testdata: /legacycode/code/BuildScript/Test/Data/PlatformRaw/PNG/6872/

• Input: 0001_LeveledShadingCorrectedgfp_mosaic_1.png

• Mask: 0001_LeveledShadingCorrectedgfp_mask_1.png

• Golden: 0001_LeveledShadingCorrectedgfp_blob_1.png

• Params: r=3, median=5, max=3 (legacy 6872 fixture; BuildScript Channel.py default median is 3)

Tests crop a central 512×512 region for CI speed; full-image optional via env.

7. Cache invalidation (operators)

After parity fix, delete existing Blob filter images or invalidate input checksums before re-running CreateBlobFilter so STOS brute uses new blob PNGs.

Re-run TestIDocBuild from RunCreateBlobFilter onward (or use IDocBuildTestBootstrapDebugging with earlier steps commented out), then compare StosBrute16 overlays before/after.

IDoc pipeline params (from Pipelines.xml): r=9, median=7, max=3.

8. Validation status

• Golden regression: legacy 6872 fixture passes at r=3, median=5, max=3 (≤1 DN vs precomputed PNG).

• IDoc repro snapshot under TESTOUTPUTPATH/Repros/IDocBuildTest predates blob generation; full slice-to-slice validation requires re-running bootstrap from RunCreateBlobFilter.

• Optional exe parity: configure NORNIR_LEGACY_IR_BLOB + fixture paths; thresholds default to MAE ≤ 0.01 and P99 ≤ 1 DN.