Advice on Developing an Accurate, Efficient OCR Pipeline

Notify Health protects children in Nigeria from preventable illness and death by delivering automated voice and text immunization reminders to their caregivers. Over the next few months, we will be developing and testing an early-stage prototype for Optical Character Recognition (OCR) digitization of paper immunization records. If successful, this technology will enable us to reach thousands more children at a lower cost.

We are seeking advice on how to develop our OCR pipeline - for example, where is OCR likely to fail or produce inaccurate results? What are the best ways to improve accuracy (e.g. data collection, preprocessing, model options, annotation, and validation practices)? What risks might we otherwise overlook? Sample images will be provided during the session.