Reading Between the Lines: Image-to-Segment Relationship Development and Analysis

1. Introduction
Crowdsourced transcription has been adopted successfully by institutions large and small in order to unlock cultural heritage data in handwritten documents. In the case of mental health documents, however, public transcription is not an option as the act of transcription involves reading them, which has implications for privacy of patients, doctors, and employees alike, potentially resulting in social repercussions to family and descendants. We are exploring mechanisms to crowdsource the transcription of such privacy-sensitive documents while maintaining the anonymity of individuals named in handwritten records by constraining the context of such mention as well as by exploiting other characteristics of documents. We report our promising initial results and describe our approach for generating structural metadata to identify multi-page units within large registers to improve the granularity of access.

1.1. Collection
Now called the Central State Hospital (CSH) [1], the first mental health institution for African-Ameri- cans in the USA was founded in 1870 near Petersburg, VA. The meticulous records maintained by the custodians of this historically significant institution have are now stored as tiff files at 400 dpi resolu- tion in a folder structure that reflects minimal structural information. The documents include hospital administrative and medical records of all stripes. The early records are handwritten and must be tran- scribed before these can be analyzed to observe patterns in administrative practices, as well as patient care. While the data set contains several types of handwritten cursive documents, including patient rec- ords, we are basing the development on board meeting minutes in order to minimize the risks in case of accidental exposure of these records.

Fig. 1: Cursive document with line breaks using adaptive thresholds.

1.2. Prior work
There is a vast body of literature on off-line handwriting recognition, focuing on methods for automatic character identification and transcription into machine-readable text [2]. Two examples of character recognition work are Tomai et al., [3] who provide a framework for mapping words in a transcript to a word image in a document and Guillevic et al., [4] who provide a character recognition approach to unconstrained, small-lexicon cursive handwriting. However, our corpus is heterogeneous, authored by many individuals, and uses a broad vocabulary.

2. Approach
We are using the Gamera libraries to segment documents for identifying individual words, in an effort to minimize the context available to potential transcribers, much like captchas are used to transcribe old, hard to OCR documents. As each document is different, and the text across lines overlaps in various ways, ruling out the use of bounding box-based methods. We using flexible, self-adjusting thresholds to detect lines. Fig. 1 shows the identified line breaks, with some breaks going through ascenders or descenders within characters in a line based on thresholds detected using document histograms. Additionally, we are using X-Y histogram profiles to identify and mark meeting minutes that span multiple pages (but begin on a fresh page) to improve the access granularity for these documents. Fig. 2 shows the current document structure using solid lines (pages within a register) and the intermediate minute structure superimposed by dashed lines.

Fig. 2: Generation of multi-page document structure

3. Discussion and Future Work
References
By employing a top-down, histogram-based approach to line and word recognition, our methods are adaptable to a large body of handwritten documents that space text at varying distances. In addition, the histograms also enable the generation of structural metadata that improves the access granularity.

The next step is to identify words within each line, where the threshold-based approach will help us locate spaces between inclined words. Transcribing at the word-level will enable us to expose varying levels of documents contexts to potential transcribers, depending upon a computation assessment of privacy sensitivity of content (for example, very short words are less likely to contain individual information). Prevention of identity disclosure in these records is critical due to the stigma associated with treatment for mental health issues. The methods we develop for transcribing handwritten documents will also be applicable to other privacy-sensitive historical records, most immediately, those of similar mental health institutions that followed the CSH.

4. References
The Central State Hospital. http://www.csh.dbhds.virginia.gov/

Cattoni, R., Coianiz, T., Messelodi S., and Modena ,C. M. (1998). 'Geometric Layout Analysis Techniques for Document Image Understanding: a Review' January 1998

Tomai, C. I., Zhang, B. and Govindaraju V. (2002). Transcript Mapping for Historical Handwritten Document Im- ages. Proceedings of the Eighth International Workshop on Frontiers in Handwriting Recognition.

Guillevic, D. and Suen, C. (1995). Cursive Script Recognition applied to the Processing of Bank Cheques.