Structure Tensor Coherence¶
The structure tensor is built from the local, locally-averaged outer product of the image gradient with itself — capturing not just gradient magnitude but how consistently oriented the gradients are in a neighborhood. Its two eigenvalues feed into a coherence score, close to 1 where the local structure is strongly, consistently oriented (e.g. along a sharp, straight edge or elongated grain boundary) and close to 0 where it’s isotropic or textureless (no dominant direction). Coherence is a good complement to Gabor’s per-orientation response when what matters is simply how directional a region is, not which direction specifically.
See also
Gabor — orientation-selective texture, rather than a single coherence score
Hessian — a related eigenvalue-based analysis of local shape
Feature Extraction — every other feature