Signature Verification in Banks using Few Shot Learning
summary
Implements a Siamese (Signet) CNN with contrastive loss trained on the BigSig260 dataset (Hindi/Bengali) using SageMaker + PyTorch, producing 128-d embeddings and a similarity score. Trains on pairs (3k pairs each split), validates on disjoint users, and reports ~97% accuracy at a ~0.8 threshold when training/validating/testing on small disjoint user sets (3 users each). Produces a ~130 MB model that supports single-reference verification without retraining; cross-language transfer drops to ~70%.
AI-written summary · about
chapters (21)
- 0:00Introduction to signatures
- 1:05Signature forgery and verification
- 2:50Online vs offline verification
- 3:55Limitations of standard CNN
- 5:59Few shot learning in banking
- 7:44Siamese networks for verification
- 9:51Contrastive loss function
- 11:13Model training setup
- 13:20Dataset preparation
- 14:44Testing with unseen users
- 16:11Image augmentation benefits
- 17:15Attention mechanism potential
- 18:01Signet network architecture
- 19:43Training process details
- 20:45Threshold selection for accuracy
- 21:46Prediction results and demo
- 22:47Advantages of Siamese networks
- 23:51Disadvantages and challenges
- 24:54Resources and references
- 25:58Language dependency in signatures
- 27:49Closing and contact information
related talks
41mSemantic Multimodal Multilingual Similarity Engine
05mFine-tuned SigLIP Image
27mTaming the Wild West of ML: Practical Model Signing with Sigstore on Kaggle
32mDetecting Signed and Unsigned Documents with Deep Learning - Beyond Transfer...
26mTaming ML’s Wild West: Practical Model Signing with Sigstore on Kaggle
35mHow deep learning can help interpret American Sign Language
40mVisual concept learning from few images
48mSignature Methods for Time Series Data