Page-level handwriting forgery detection (PageEncoder, run main_s0)
Page encoder for writer verification: ResNet-18 (1 channel, random init) over K=64 ink-centred 128px patches per page, attention pooling, 256-d L2-normalised page embedding. Trained with SupCon plus feature and relational distillation from a frozen DINOv2-small teacher (the teacher is not needed at inference). English handwriting only (IAM + CVL).
Compare a questioned page with N reference pages of the claimed writer by cosine similarity of embeddings.
Validation EER (best checkpoint): IAM 0.005 (optimistic, shared scanner/pen), CVL 0.059.
Files: best.pt (state_dict), config.yaml (embed_dim, pool), best_val.json.
Runs on CPU; uses a GPU automatically if present. Not a forensic tool: outputs are evidence of similarity, not proof.
- Downloads last month
- -
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support