restoration op• 数据种类:voxel × voxel → voxel
• 调用: import fullseye as fs; fs.ledger.vol_richardson_lucy(vol, psf, iterations=10, clip_tiny=1e-12)(要直接调用实现,import volrestore; volrestore.vol_richardson_lucy(vol, psf, iterations=10, clip_tiny=1e-12);从台账取用则 ops3d.get("vol_richardson_lucy"))
用已知 PSF 对非负体做 Richardson–Lucy 反卷积。
> 以下的详细说明为原文 —— 摘要与标题已翻译。
*psf* is a 3-D non-negative kernel (any odd/even size smaller than the
volume; it is normalised to sum 1 internally so overall intensity is
preserved). *iterations* trades sharpness against noise amplification —
5-30 is the practical range (see the module notes on semi-convergence).
Negative voxels are refused (RL is a Poisson model) — except *rounding
dust*: values no lower than `-NEGATIVE_DUST_TOL * max|vol|` (1e-9
relative; an FFT-blurred observation typically carries -1e-16) are clipped
to 0 instead of rejected, so the module's own forward model feeds back in.
Returns the deblurred `(D, H, W)` float64 volume (non-negative).
Measured on the test scene (binary sphere pair blurred by a sigma-2
Gaussian): the RMSE to ground truth falls to 0.81x the blurred
observation's at 10 iterations and 0.68x at 50 — genuine but *gradual*,
because the residual is dominated by the spheres' hard edges, which RL
recovers slowly. What converges fast is the *forward consistency*:
re-blurring the estimate reproduces the observation almost exactly (that
is the quantity the RL update actually optimises).
• 示例数据目录(下载 URL / 许可证) —— 2-D 用 skimage.data(BSD/公有领域)加合成图,3-D 给出真实数据源(Stanford/PDS 等)的下载 URL。
• 算子来历与参考文献 —— 该算子族所依据的研究/方法出处。
• deconv_fft_restore — py -3.11 examples_3d/deconv_fft_restore.py
voxel 作为输入)voxel_to_mips · voxel_to_mesh · signed_distance_field · to_points · sobel3d · hessian3d · curvature_maps · edt_jfa
restoration)*Provenance: volrestore.py — 3D 算子登记表。本条目由 tools/opdocs.py md 自动生成(请勿手工编辑)。*
© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.