Deep Learning for Document Analysis, Text Recognition, and Language Modeling

Thomas Breuel, NVIDIA Research, USA

Deep Learning has emerged as the predominant approach to many recognition tasks related to OCR and document analysis. The tutorial will cover applications of deep learning to problems in document analysis:

• models for OCR and text recognition, including recent developments
• DL approaches to layout analysis and preprocessing
• recent advances in DL models for language modeling and OCR
• obtaining training data; semi-supervised and unsupervised methods
• tools for large scale processing

The course will present numerous examples and workbooks based on PyTorch. Basic familiarity with deep learning and Python is recommended.

Outline

• basic OCR models in PyTorch
• convolutional and LSTM models; CTC
• beam search
• attentional models; transformers
• degradation and cleanup of images
• layout analysis and text detection
• statistical foundations of language modeling
• corrective language models and bidirectional LSTM
• natural language models
• theory of semi-supervised and unsupervised learning
• DL approaches to semi-supervised and unsupervised training
• scaling up to large datasets and distributed training

Reading List and Materials

Tutorial slides:

https://github.com/tmbdev/icdar2019-tutorial

Background readings and materials can be found at:

www.tmbdev.net/icdar2019-readings

Related materials from the DAS 2018 Tutorial can be found at:

github.com/tmbdev/das2018-tutorial

(The ICDAR2019 tutorial will contain substantial new material.)

Potential Target Audience

Graduate students and researchers interested in applying deep learning to OCR, scene text recognition, document analysis, and related areas.

There will be some overlap with last year’s DAS 2018 tutorial, and the tutorial will provide a self-contained introduction, but the focus will be on different topics, including the latest version of PyTorch, large scale processing, distillation, semi-supervised training, and distributed training. The tutorial is intended to be useful both to audience members who have seen last year’s tutorial and for audience members who are new to DL for document analysis.

Worksheets for last year’s tutorial are at http://github.com/tmbdev/das2018-tutorial

The Presenter

Thomas Breuel works on deep learning and its applications at NVIDIA Research. Before that, he was a researcher at Google Brain, IBM, and Xerox PARC. He was a professor of computer science and head of the Image Understanding and Pattern Recognition (IUPR) at the University of Kaiserslautern. He has published numerous papers in document analysis, computer vision, and machine learning and is a contributor to several open source projects in OCR, document analysis, and machine learning.