Extraction of Handwritten Text from Carbon Copy Medical Form Images

Robert Milewski, Venu GovindarajuDocument Analysis Systems 2006 · DOI

Abstract

Automated extraction of handwritten text from carbon copy medical forms, such as the New York State Pre-Hospital Care Report (PCR), presents significant challenges due to heavy carbon mesh noise and low contrast. This paper introduces a novel binarization algorithm utilizing a sinusoidal scanning trajectory to effectively separate handwriting from background noise. By dynamically navigating around handwritten strokes to identify true background pixels, the proposed method achieves a word recognition accuracy improvement of up to 21% over standard techniques like Otsu’s method.

Introduction

In emergency medical services, the New York State Pre-Hospital Care Report (PCR) is a critical document for patient care and health surveillance. However, these forms are often carbon copies filled out under suboptimal conditions, such as in moving ambulances. The resulting images are plagued by "carbon mesh noise," where the background texture and handwritten strokes share similar grayscale intensities. Standard binarization methods frequently fail to distinguish between light pen strokes and carbon residue, leading to broken characters and poor OCR performance.

Handwritten text obscured by heavy carbon mesh noise on a carbon copy form.

Methodology

The core of the proposed solution is a "Sine Wave" binarization approach. Unlike traditional adaptive binarization that uses fixed square windows, this method employs a dynamic scanning strategy:

  1. Dual Masking: A central $5 \times 5$ mask evaluates the target pixel, while eight $3 \times 3$ outer masks sample the surrounding area.
  2. Sinusoidal Trajectory: The outer masks move along a sine wave path rather than a straight line. This allows the algorithm to "hop" over thick or curved strokes to find cleaner background paper regions.
  3. Decision Logic: A pixel is classified as foreground (text) if it is significantly darker than the background regions identified along the sinusoidal paths, governed by a carbon intensity similarity rule.

Evaluation

The method was evaluated using a dataset of 30 NYS PCR forms, yielding 1,440 word images. Performance was measured using a Lexicon Driven Word Recognizer (LDWR) across various lexicon sizes.

Example of a New York State Pre-Hospital Care Report (PCR) form section used in the study.

The Sine Wave (SW) method achieved 56.5% accuracy with a 100-word lexicon, significantly outperforming Otsu’s method (35.3%). With post-processing filters like despeckle and blob removal, accuracy reached 59.1%. The SW method consistently maintained a 4-7% performance lead over optimized baseline methods across all tested lexicon sizes (up to 4,000 words).

Conclusion

The sinusoidal scanning trajectory provides a robust mechanism for evading handwritten strokes to accurately estimate background intensity in high-noise documents. While current work relies on manual word segmentation, future research will focus on automated segmentation and analyzing system reliability through accept/reject rates to better serve the needs of medical surveillance.

Summary updated

BibTeX
@inproceedings{milewski2006extraction,
  author = {Robert Milewski and Venu Govindaraju},
  title = {{Extraction of Handwritten Text from Carbon Copy Medical Form Images}},
  booktitle = {Document Analysis Systems},
  year = {2006},
  doi = {10.1007/11669487\_10},
}
Figure