fix(qfactor): correct 1/√2 offset scaling and LL-band visual weight

[osamu620]

Summary

Two related correctness fixes to the Qfactor quantization path, validated against the WG1 guideline N100430 (Controlling JPEG 2000 image quality using a single parameter). Applied symmetrically to QCD_marker, QCC_marker, and the estimate_qfactor inverse predictor.

1. 1/√2 offset scaling

The guideline applies the constant offset as part of Δ_Q, before the √G_Y scaling:

Δ_Q   = 2^P0 · α_Q · M_Q + 1/√2
Δ_ref = Δ_Q · √G_Y

The code added eps0 (the normalized 1/√2) outside the √G_Y multiply, so the offset term was too small by a factor of √G_Y ≈ 1.73. The deviation is largest at high Qfactor, where the offset dominates as M_Q → 0 (the best-quality regime).

-      double delta_Q    = alpha_Q * M_Q;
-      double delta_ref  = delta_Q * G_c_sqrt[0] + eps0;
+      double delta_Q    = alpha_Q * M_Q + eps0;
+      double delta_ref  = delta_Q * G_c_sqrt[0];

2. LL-band visual weight at dwt_levels < 5

The guideline states the LL sub-band is assigned visual weight 1.0 in all cases. The old (i >= W_b.size()) guard only catches the LL band when dwt_levels == 5 (LL index == 15 == table size). For fewer levels the LL band picked up a stray detail-band weight — wrong for chroma (QCC) at <5 levels. Now the last band (always LL) is forced to 1.0, while the i >= size clause still covers extra low-freq bands when dwt_levels > 5.

3. estimate_qfactor kept in lock-step

The tool inverts the same formula, so it carried both bugs. Fixed identically so it stays exact against streams from the corrected encoder.

Verification

• Emitted QCC step sizes match the guideline formula bit-for-bit (independent Python reference), across all 16 subbands of both chroma components.
• Confirmed empirically that anchoring Δ_ref to √G_Y (not the per-component √G_c) is correct: swapping to G_c_sqrt[Cqcc] collapses all components' LL bands to the identical luminance value, dropping the per-component color weighting.
• estimate_qfactor recovers the true Q with 0.0000 residual across Q ∈ {10,30,50,65,80,90,97,100} (all piecewise branches) and dwt_levels ∈ {2,3,4} (exercises the LL-weight fix). Before the fix it estimated Q=79 for a true-Q=80 stream.
• Builds clean; ctest -R "^enc_" and the Part 18 round-trip chain pass.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR corrects two guideline-defined details in the Qfactor-based quantization step-size computation (encoder-side QCD_marker/QCC_marker) and keeps the estimate_qfactor inverse predictor mathematically consistent with the corrected encoder behavior.

Changes:

• Applies the 1/√2 offset inside Δ_Q (before √G_Y scaling) so the offset scales correctly with G_c_sqrt[0].
• Forces the LL band’s visual weight to 1.0 for all dwt_levels by explicitly treating the last band as LL (while still defaulting to 1.0 for bands beyond the 5-level table when dwt_levels > 5).
• Removes an effectively-dead qfactor == 0xFF ternary inside the qfactor-only branch (since 0xFF is handled by the earlier branch).

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
source/core/codestream/j2kmarkers.cpp	Fixes Qfactor quantization math for offset placement and LL-band weighting in emitted QCD/QCC marker step sizes.
source/apps/estimate_qfactor/main.cpp	Mirrors the same formula fixes so estimated Qfactor remains exact against streams produced by the corrected encoder.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.