BACKLOG.md

Operation Backlog

The re-arrangeable, priority-ordered list of filters and operations to implement (Step 5 of ROADMAP.md).

How to use this file:

• Reordering rows within a priority band — or moving a row between bands — is reprioritizing. Edit freely; PR to develop.
• When work starts on a row, open a GitHub issue, link it in the Status column, and move status to in progress.
• After Step 3 (CPU backend) lands, every new operation ships both backends (Cuda*.cuh + Cpu*.h) plus a parity test against ground truth in test_data/ — that is part of the definition of done.
• New operations follow the three-file pattern in CLAUDE.md: CUDA/CPU driver, ScrCmd*.h, one SCR_CMD line.

Column notes — Effort: S = days, M = 1-3 weeks, L = 1-2 months (per backend where two exist). Chunk-safe: whether the op fits the streaming-chunk architecture (local neighborhood or mergeable reduction).

Priority 1 — cheap wins (compositions of existing ops)

Operation	Kind	Effort	Chunk-safe	Depends on	Status	Notes
Top-hat	Composition	S	yes	Opener, ElementWiseDifference	proposed	image - open(image); bright feature extraction
Bottom-hat	Composition	S	yes	Closure, ElementWiseDifference	proposed	close(image) - image; dark feature extraction
Morphological gradient	Composition	S	yes	MaxFilter, MinFilter	proposed	max(image) - min(image); edge strength
Difference of Gaussians	Composition	S	yes	Gaussian	proposed	Two sigmas, subtract; blob enhancement
Unsharp mask	Composition	S	yes	HighPassFilter	proposed	HighPass already exists; add amount parameter

Priority 2 — new local kernels (chunking-compatible)

Operation	Kind	Effort	Chunk-safe	Depends on	Status	Notes
Sobel / gradient magnitude	New kernel	S	yes	—	proposed	Separable derivative kernels
Niblack / Sauvola threshold	New kernel	S-M	yes	MeanFilter, StdFilter	proposed	Local adaptive thresholding; Mean/Std already exist
Otsu threshold	Reduction	S-M	yes	histogram reduction	proposed	Histogram is a mergeable reduction (like GetMinMax), so chunk-safe despite being global-valued
Bilateral filter	New kernel	M	yes	—	proposed	Edge-preserving smoothing; range+spatial weights
Anisotropic diffusion	Iterative kernel	M-L	yes	—	proposed	Perona-Malik; halo cost per iteration; natural Step 4 pipeline citizen

Priority 3 — FFT track

Shared prerequisites for this band: cuFFT (GPU) + vendored pocketfft (CPU, BSD-3); overlap-add/save chunking; normalized-convolution boundary mode (see ROADMAP Step 5 notes). Standing check: cuFFT is a large dynamic redistributable — coordinate with conda/NuGet packaging before landing.

Operation	Kind	Effort	Chunk-safe	Depends on	Status	Notes
FFT convolution (large kernels)	FFT	L	yes (overlap-add)	FFT infrastructure	proposed	Wins over spatial above ~15^3 kernels; parity-test vs MultiplySum
FFT bandpass / DoG	FFT	S after FFT conv	yes	FFT convolution	proposed	Frequency-domain band selection
Richardson-Lucy deconvolution	FFT, iterative	L	yes	FFT convolution; Step 4 pipeline (strongly)	proposed	Marquee microscopy feature; iterative conv/divide/multiply loops
Wiener deconvolution	FFT	M	yes	FFT convolution	proposed	Non-iterative alternative to R-L; distinct from existing spatial WienerFilter

Priority 4 — global operations (architecture caveat)

These require global propagation/label passes that conflict with the streaming-chunk architecture. Each row needs a design decision before implementation: either a fits-on-one-GPU gate (error on images that would chunk) or a research-grade border-merge design. Until then, the documented workflow is Hydra for filtering, scikit-image/ITK for global segmentation.

Operation	Kind	Effort	Chunk-safe	Depends on	Status	Notes
Connected components	Global label	L	no — needs gate or border-merge	design decision	proposed	Union-find on GPU is well-studied for single-buffer images
Distance transform	Global sweep	L	no — needs gate or border-merge	design decision	proposed	Prerequisite for watershed seeding
Watershed	Global propagation	L-XL	no — needs gate or border-merge	connected components, distance transform	proposed	Most-requested segmentation op; hardest to chunk correctly
Morphological reconstruction	Global iteration	L	no — needs gate or border-merge	design decision	proposed	Enables h-maxima/h-minima, hole filling

Priority 5 — ideas / unscheduled

Operation	Kind	Effort	Chunk-safe	Depends on	Status	Notes
Hole filling	Composition of reconstruction	S after reconstruction	no	morphological reconstruction	proposed
H-maxima / H-minima	Composition of reconstruction	S after reconstruction	no	morphological reconstruction	proposed	Seed detection for watershed
Local entropy-based threshold	New kernel	M	yes	EntropyFilter	proposed
Structure tensor / orientation	New kernel	M	yes	Sobel	proposed	Fiber/orientation analysis