EfficientNet-B0 with Batch-Hard Triplet Loss Fine-Tuning for Multi-Input Signature Forgery Detection
Published:
A writer-independent signature verification system that freezes a pretrained EfficientNet-B0 backbone and fine-tunes only its embedding head with batch-hard triplet loss, mapping signatures to L2-normalized embeddings compared by cosine similarity and evaluated on writers held out entirely from training.
Recommended citation: Sunoto, D. N., Febrianto, Frolentika, Davalynda, J., Sulistiawati, W., & Tarigan, G. A. (2026). "EfficientNet-B0 with Batch-Hard Triplet Loss Fine-Tuning for Multi-Input Signature Forgery Detection."