Detecting Signed and Unsigned Documents with Deep Learning - Beyond Transfer...
summary
A pipeline uses transfer learning on a pretrained YOLO object detector plus Google Cloud OCR to localize signed and unsigned signature boxes, trained on 40,000 labeled pages (about 10,000 unsigned and 4,000 signed boxes). Models reach >90% page-level accuracy for binary tests, gain ~20% recall after tripling labeled data, run in Keras/TensorFlow on AWS GPUs, and deploy as a Flask microservice on AWS Fargate with a human‑in‑the‑loop review and high‑recall bias.
AI-written summary · about
chapters (19)
- 0:00Signature detection intro
- 4:03Why signature detection matters in legal documents
- 5:30Existing methods' limits
- 7:15CNN binary classification
- 10:47Transfer learning myths
- 12:12YOLO for localization
- 16:45Anchor boxes & unsupervised
- 18:12Training & hyperparameters
- 19:38OCR for data aug
- 20:40AWS training & deploy
- 21:45Model performance insights
- 22:49Data labeling impact
- 23:33Dataset understanding
- 24:20Q&A: Motivation
- 25:26Q&A: Localization approach
- 26:30Q&A: Segmentation methods
- 27:12Q&A: Two-class setup
- 28:04Q&A: Recall vs review
- 29:09Q&A: Reduce review time
related talks
42mUsing Deep Learning to Understand Documents
51mDeciphering doctor's handwriting with deep learning
19mDocument Reader using Deep Learning with a Small Dataset
46mML Use Cases at Think Big Analytics
40mDeep Learning and Optical Character Recognition
29mdeepdoctection - An open source package for document intelligence
32mApplying Deep Learning for Object Detection from a Telecom Perspective
36mHow to Build a Cloud Native Image Recognition Solution