Template-based handwriting scoring for preschool letters (pixel overlap / error ratio) — looking for metrics & related work

1

Hi everyone,
I’m working on a research component where I need to score how accurately a preschool child wrote a single letter (not just classify the letter). My supervisor wants a novel scoring algorithm rather than “train a CNN classifier.”

My current direction is template-based:

  • Preprocess: binarize, center, normalize size, optionally skeletonize

  • Have a “correct” template per letter

  • Overlay student sample on template

  • Compute an error score based on mismatch: e.g., parts of the sample outside the template (extra strokes) and parts of the template missing in the sample (missing strokes)

I’m looking for:

  1. Known metrics / approaches for template overlap scoring (IoU / Dice / Chamfer / Hausdorff / DTW / skeleton-based distance, etc.)

  2. Good keywords/papers for handwriting quality scoring or shape similarity scoring, especially for children

  3. Ideas to make it more robust: alignment (Procrustes / ICP), stroke thickness normalization, skeleton graph matching, multi-view (raw + contour + skeleton) scoring

Also—my supervisor mentioned something like using a “ratio” (she referenced golden ratio as an example), so if there are shape ratios/features commonly used for letters (aspect ratios, curvature, symmetry, stroke proportion, loop size ratio), I’d love suggestions.

Thanks!