Steganoscription : exploring techniques for privacy-preserving crowdsourced transcription of handwritten documents

Abstract

The focus my research is the historical document format represented by the Central State Hospital (CSH) dataset, handwritten medical records. The specific problem innate to the CSH dataset in question is how to transcribe sensitive, cursive-handwritten documents via a manual vehicle- such as crowdsourcing. Manual methods are necessarily no matter the sophistication of the optical character recognition system used because of the inconsistencies within cursive script. To address this problem I've developed an application that enables users to transcribe sensitive, handwritten, document images while preserving the privacy of the context around the transcribed text via random word selection and visual manipulation of the displayed text. This is made possible through several algorithms that process documents from a top-down approach. These system operations detect and segment lines of text in images, reverse the slant common to cursive script, detect and segment words, and finally, manipulate word-images before they are displayed to users; combinations of color, noise, and geometric manipulations are currently supported and used randomly. This system, called Steganoscription, combines the concepts of steganography and transcription.