Using Deep Learning to Understand Documents
summary
An end-to-end computer-vision model extracts key fields directly from invoice images, trained on millions of images and using a convolutional encoder + GRU decoder to produce character-level output for amounts and dates. The system scores each page independently for multi-page docs, uses image augmentations for rotation/blur robustness, and returns confidences that gate automation. Compared with OCR-plus-NLP (AWS/Google or Tesseract) it reports higher accuracy, lower latency and lower cost on Bill.com workloads, though initial training requires large labeled volumes or transfer learning.
AI-written summary · about
chapters (16)
- 0:00Bill.com and document processing
- 2:09Extracting fields from invoices
- 4:39OCR two-step approach limits
- 7:12End-to-end computer vision benefits
- 10:05Rotated and skewed invoice images
- 15:23Deep learning building blocks
- 19:21Model architecture for field extraction
- 22:15Training progress and loss curves
- 24:40Multi-page invoice processing
- 26:05Explainability via rastering and mapping
- 28:55Data labeling and transfer learning
- 33:08Confidence scoring and garbage handling
- 36:14Augmentation vs inference processing
- 38:45Date formats and new templates
- 40:09Transformer models potential
- 41:34Q&A
related talks
40mDeep Learning and Optical Character Recognition
01h 19mLearn to Unlock Document Intelligence with Open-Source AI
29mBetter Understand Your Documents with AI
01h 33mComputer Vision in Azure
32mDetecting Signed and Unsigned Documents with Deep Learning - Beyond Transfer...
51mDeciphering doctor's handwriting with deep learning
30mIntroduction to Computer Vision with OpenCV
28mSupercharge Information Extraction with Vision-Based LLMs