This workshop provides a technical introduction to deep-learning-based handwritten text recognition (HTR). We will begin with a general introduction to the topic and outline a typical HTR pipeline. Afterwards, each pipeline’s steps will be discussed in detail, including hands-on exercises using PyTorch.
The presented examples will focus on a handwritten document scenario, however the general pipeline and most of the steps are directly applicable to printed material as well. Differences and possible adaptations will be pointed out along the way.
While transformer-based HTR models are gaining popularity, this workshop will focus on the more classical but still state-of-the-art approach of Connectionist Temporal Classification.
When & Where?
2024-09-24, 12:30 - 15:30 at DSC-Oasen, Georg Sverdrups hus, University of Oslo Library
Registration and more info at: https://www.ub.uio.no/english/libraries/dsc/berut/events/workshops/2024-09-24_ws_pytorch.html
Lesson Plan
- Welcome and introductions
- General Introduction to Handwritten Text Recognition - Slides
- Data for HTR - Slides
- Preprocessing & Augmentation - Slides
- HTR Models and the CTC Loss - Slides
- Inference and Performance Metrics - Slides
- General Thoughts and Experiences from Implementing HTR Models - Slides
- Final questions and wrapping up
Code and Installation Instructions
The code for the workshop, including installation instructions, is available in this github repository.
Material
Please download the material from here