demosaic: add Menon (2007) DDFAPD and ARI algorithms

[dregsist]

Summary

darktable's demosaicing lineup covers the speed-vs-quality axis well at
the extremes (VNG/PPG for speed, AMaZE for quality), but the mid-range
lacks per-pixel adaptive selection. This PR adds two algorithms that
fill that gap:

• Menon (2007) — Directional De-mosaicing for Frequency-Adaptive
  Pattern (DDFAPD): a two-pass algorithm that separates horizontal and
  vertical interpolation and merges them via local gradient energy
  weighting. Produces sharp edges with reduced zipper artifacts at a
  cost comparable to RCD.

• ARI — Adaptive Resolution Interpolation (Monno et al. 2017):
  generates multiple green-channel candidates (MLRI via Laplacian
  energy direction, and HA via Hamilton-Adams), then selects per-pixel
  using noise-adaptive criteria. Three quality levels: Fast (MLRI only),
  Balanced (2 candidates), Quality (3 candidates with guided-filter
  smoothing on color differences). ARI's per-pixel candidate selection
  also addresses the same high-ISO use case as the existing
  AMaZE+bilinear blend, without requiring manual threshold tuning.

Implementation notes

• Both are pure CPU implementations using OpenMP, consistent with the
  existing RCD and LMMSE modules.
• Negative Bayer values (possible after rawprepare) are clamped to zero
  at algorithm entry, consistent with RCD's _safe_in() approach.
• ARI exposes a quality parameter (1–3) in the UI; stored in the
  params struct, version bumped 6→7 with legacy_params migration.
• Menon reuses the existing procmax scaler used by RCD/LMMSE.
• Both algorithms are Bayer-only; X-Trans paths are unaffected.

References

• D. Menon, S. Andriani, G. Calvagno, "Demosaicing With Directional
  Filtering and a posteriori Decision," IEEE Trans. Image Process.,
  vol. 16, no. 1, pp. 132–141, Jan. 2007.
• Y. Monno, D. Kiku, M. Tanaka, M. Okutomi, "Adaptive Residual
  Interpolation for Color and Multispectral Image Demosaicking,"
  Sensors, vol. 17, no. 12, 2807, 2017.

Testing

• Tested on Bayer RAW files from Canon and Nikon bodies.
• No purple fringing or zipper artifacts observed on high-frequency
  test targets.
• ARI Quality mode shows improved fine-detail retention compared to
  bilinear and VNG on high-ISO crops.
• No regressions in existing demosaic methods (RCD, LMMSE, VNG,
  AMaZE) confirmed.

[TurboGit]

Some style issues. TIA.

[dregsist]

Thank you for the review!

• Copyright year corrected to 2026 in menon.c; standard darktable GPL header added to ari.c as well.
• All multi-parameter lines split to one parameter per line across both files.

[jenshannoschwalm]

Would you be able to hint at raw images plus the region to inspect where you think there is an improvement over current demosaicers?

Menon is ~half performance of rcd here, can you confirm? Btw there definitely is room for perf gains by internal tiling.

[da-phil]

Can we use the integration test repo images to have a quantitative comparison between demosaicers?

[jenshannoschwalm]

Can we use the integration test repo images to have a quantitative comparison between demosaicers?

nope.

[jenshannoschwalm]

I just did a first round of detail-checks on various images i have in my collection and must say, i am not convinced by quality of both algos . Maybe i am testing on "wrong" images, please share examples where you would say the new algos do better. A summary of my results

On high ISO images both algos do quite similar to LMMSE

On low-ISO images both algos

1. introduce maze patterns (as RCD) in areas where VNG4 is stable
2. introduce artefacts at areas with strong gradients

Using RCD as reference

Menon

ARI

Didn't lookup original source. Could it be we need "normalizing" of input/output (as we do for RCD, LMMSE) ?

[dregsist]

Thank you @jenshannoschwalm for running the comparison —
the color fringing on mire1 is clearly not right.

I'll investigate first — likely a bug in the R/B interpolation
direction or a Bayer pattern offset issue in the Menon implementation.
Will also verify ARI output correctness. Fixing both before anything else.

After that, planning to:

• add internal tiling for both Menon and ARI as you suggested
• run CPSNR/SSIM benchmarks (Kodak etc.) across all demosaic methods
  for a proper quantitative comparison

@da-phil yes, will include results against the integration test images as well.

[dregsist]

The previous ARI implementation had quality issues (color fringing, maze patterns on mire1) because it cut too many corners in the adaptive selection heuristic. This rewrites ARI to match the Monno 2015 paper line-by-line (structurally equivalent to the authors' MATLAB reference, within ±0.002 dB CPSNR of a Python port).

Quality is indeed better:

• Kodak low-ISO CPSNR 39.94 dB, +0.8 dB over amaze
• Chroma zipper smallest across all tested methods and ISOs
• SIDD real raw iso_clean: +0.38 dB over amaze+dual

However, the paper-exact pipeline is slow:

• 30-80x slower than existing methods
• 53s per 15 MP patch at 16 threads (standalone)
• 113s through darktable pipeline on mire1.cr2
• Memory-bandwidth bound; OpenCL is unlikely to reach practical speeds

Optimizations applied (all preserve paper-exact output):

• Rolling sum box filter replacing double integral image
• Row-major vertical pass for cache-line reuse on wide images
• Paired GF sum sharing (~33 % box-call reduction)
• Separable Gaussian 5x5 for criterion smoothing
• Separable bicubic 7x7 for R/B residual upsample
• Removed redundant refined-estimate recomputation in RI-H
• Criterion-side adjacent-loop fusion

Menon (2007) is separately included as a reference implementation. It is fast (~amaze speed), but benchmarking across Kodak-24 + SIDD-medium did not show a clear advantage over amaze — chroma zipper is similar or worse (especially cz_peak), and CPSNR is within 0.1 dB.

Both algorithms looked strong on paper; extensive benchmarking showed neither clearly beats the existing methods across the tested conditions. I'll leave the merge decision to maintainers. The benchmark numbers themselves should be useful regardless; full writeup + reproducible scripts + Python reference ported from the original MATLAB are available at:

https://github.com/dregsist/darktable-rawforge/tree/demosaic-benchmarks

menon.c: keep the unused FIR/k_b locals (they're referenced via OpenMP firstprivate clauses) to fix -Werror=unused-variable with GCC 15.

[jenshannoschwalm]

Thanks for reworking ARI and especially the "benchmarking" :-)

1. I think the good noise data for the dual demosaicers on high ISO images are mostly due to the low-pass effect of the low-frequency content (bilinear part of VNG4 plus smoothing)
2. The "amount of calculated detail" is difficult to benchmark.
3. I guess we agree that there is no big point to include menon in dt.
4. I looked at the re-implementation of ARI, that indeed uses a large amount of memory (memory bound performance) per pixel. As that is comparable to markje3 we could guess-assume a two-fold performance gain if using internal CPU tiling for L2 cachesize. When implementing such code in OpenCL we mostly find a perf-gain of ~4-fold on capable hardware. Expected performance still pretty slow. Not sure if we should include this ...

[dregsist]

Thanks for looking at this and for the honest read.

Fully agree with your assessment. Both methods I picked from recent papers hoping they'd bring something new, but the benchmarks across Kodak-24, SIDD real raw, and chroma-zipper metrics made it clear:

• ARI: quality wins at low ISO (+0.8 dB Kodak, +0.38 dB SIDD clean over amaze+dual), but 30–80× slower is just not viable for an interactive tool. Even with OpenCL the paper algorithm (11 iter × RI+MLRI × per-pixel argmin) is memory-bandwidth bound, so the gap to amaze/rcd wouldn't really close.
• Menon: fast enough, but no margin over amaze at any ISO tested. Hard to argue for inclusion on quality grounds.

The existing rcd/amaze/+dual lineup is well-balanced and I don't see a use case these two would unlock in practice. I'll leave the merge/close decision to you — happy either way.

The benchmark writeup and reproducible scripts are at dregsist/darktable-rawforge/demosaic_test in case the numbers are useful for future demosaic discussions.

[jenshannoschwalm]

I tried the reference ARI again (on my pretty fast machine) it a) is very slow b) seems to give slightly more details without introducing artefacts compared with rcd or amaze c) it has better "chroma noise resistance". Although when using capture sharpening on rcd the details advantage is not visibly any more to me.

I am still working on pixelpipe caching improvements, with current ROI concept ARI is unfortunately not practical.