hrebsd

de_ebsd.index.crosscorr.hrebsd(patterns, reference, detector, pattern_center=None, reference_pattern_center=None, *, roi_size=128, grid=8, margin=0.05, upsample=50, n_passes=2, bandpass=None, whiten=0.0, min_peak=0.0, batch=256, device=None)

Deviatoric strain and lattice rotation relative to a reference pattern.

Cross-correlates a grid of regions between each test pattern and the reference, fits the gnomonic shift equation for the trace-free displacement gradient, then remaps the reference into the measured configuration and repeats. That second pass is what makes the result trustworthy once lattice rotations exceed a degree: it drops the residual rotation to first-pass accuracy, and pattern-center error reaches the strain only through that residual.

Parameters:

patterns

Test patterns (H, W) or (N, H, W), background-corrected; see de_ebsd.index.preprocess.

reference

(H, W), or (N, H, W) for one reference per test point. Strain is measured relative to whatever state this pattern is in.

detector

EBSDDetector, for the gnomonic geometry.

pattern_center, reference_pattern_center

Pattern center at the test and reference points, (3,) or (N, 3); both default to detector.pattern_center. Passing them separately is worth doing over any appreciable scan – a millimetre of travel moves the pattern center by several percent of the screen – and the remap absorbs the difference.

roi_size

Region side in pixels. Larger regions correlate more precisely but average the deformation over a wider gnomonic span.

grid

Regions per side, so grid ** 2 in total, against 8 free parameters.

margin

Edge inset as a fraction of each side.

upsample

Sub-pixel resolution of the correlation peak, 1 / upsample pixels.

n_passes

Correlation passes; 1 disables remapping and is the classic single-pass formulation. A pattern-center difference is still applied on pass 1.

bandpass

(low, high) Gaussian band-pass in cycles per pixel (Nyquist is 0.5), or None for flat. Rarely needed on real patterns; raise low above zero to suppress slow background gradients that survived preprocessing.

whiten

Spectral whitening exponent, 1.0 being full phase correlation. Defaults to 0 (plain correlation), which measured better than whitening on real patterns. See _measure_shifts() for the failure mode it exists to defend against.

min_peak

Regions whose normalized cross-correlation falls below this are dropped from the fit. Regions are weighted by that correlation regardless.

batch

Patterns correlated at once, which bounds memory: it scales as batch * grid**2 * roi_size**2.

device

Defaults to the test patterns’.

Returns:

HREBSDResult

Detector-frame tensors, the raw region shifts, and a per-point residual.