refine_hr#

de_ebsd.index.refine.refine_hr(patterns, master, detector, rotations, pattern_center=None, refine_pattern_center=False, n_tiles=16, samples_per_tile=8, n_steps=15, damping=0.001, keep_best=True, device=None, all_pixels=False, batch_size=None, sample_by='master')[source]#

High-precision joint orientation + pattern-center refinement.

The sub-0.1-degree-capable path. Rather than touching every detector pixel, it importance-samples a few thousand band-edge points spread across a tile grid and runs batched Levenberg-Marquardt on (R, pattern_center) against the continuous master, over an objective normalized per tile (robust to intensity ramps). The sample budget (n_tiles, samples_per_tile) is the accuracy/speed knob.

Parameters:

patterns: A detector image (H, W) or stack (Np, H, W). The raw sampled values are used – the experimental signal is never back-projected or interpolated.
master, detector: The continuous de_ebsd.sim.MasterPattern and the EBSDDetector geometry.
rotations: Seed orix.quaternion.Rotation, typically a coarse index result.
pattern_center: None uses detector.pattern_center. A (3,) value shares one center across the stack; (Np, 3) gives a per-pattern center. Returned in the same shape.
refine_pattern_center: Refine the pattern center jointly (6 DOF) and return it, or (default) leave it untouched and refine orientation only (3 DOF).
n_tiles, samples_per_tile: Tiles per detector side and samples kept per tile; the defaults give 256 * 8 = 2048 samples. Raising them improves accuracy proportionally.
n_steps: Levenberg-Marquardt steps (default 15); second-order, so a coarse seed sharpens well inside this budget.
damping: Initial damping, adapted per pattern each step (x0.5 accepted, x3 rejected).
keep_best: Never return a (R, pattern_center) scoring below the seed, making refinement monotone.
device: Defaults to master.data.device; otherwise everything is moved there.
all_pixels: Use every detector pixel instead of the band-edge subsample (slow reference path for validation).
batch_size: Patterns per mini-batch. None picks a memory-aware size; the result is identical for any batch size.
sample_by: Which gradient picks the sample pixels. "master" (default) renders the simulated pattern at the seed and samples its noise-free gradient, so samples land on true band edges. "experimental" samples the raw pattern, which on real data is noise-dominated and adds substantial refinement scatter; it is retained for noise-free/synthetic use.

Returns:

rotationorix.quaternion.Rotation: The refined orientation(s), in the convention MasterPattern.project consumes.
scoretorch.Tensor: Mean per-tile NCC per pattern, in [-1, 1]; >= the seed when keep_best.
pattern_centertorch.Tensor: Only when refine_pattern_center: the refined pattern center, shaped like the input.