KODAK lMAGELINK™ OCR Alphanumeric Handprint Module

NeurIPS 1995

TL;DR

This paper introduces the Kodak Imagelink™ OCR Alphanumeric Handprint Module, a system that uses two neural networks and a novel positioning approach to accurately recognize handwritten alphanumeric characters on forms. It achieves high success rates, even with touching characters, by combining robust character detection with application-specific postprocessing.

What Problem Does This Paper Address?

Accurately recognizing handwritten alphanumeric characters on forms is a persistent challenge in document processing. Traditional Optical Character Recognition (OCR) systems often struggle with variations in handwriting, touching characters, and the need for high throughput. This paper tackles the problem of reliably extracting and classifying alphanumeric fields, specifically for applications like vehicle registration forms, where precise identification is crucial. The authors highlight the difficulty in distinguishing between similar characters (e.g., 'O' and '0', 'I' and '1') and the need for a system that can handle these ambiguities and the occasional overlapping of characters.

Methods

The core of the Kodak Imagelink™ system comprises two neural networks: one for positioning characters and another for classification.

  • Positioning Network: Instead of a single node to detect character centers, this network uses an array of output units. This allows for averaging activations from slightly shifted windows, improving robustness and speed. The network is trained to produce "bumps" of activation indicating character "hearts" (centers), with target activations set for three consecutive nodes to account for the imprecision of human marking.
  • Classification Network: This network also utilizes Gabor projections as its initial feature extraction layer, similar to the positioning network. It then employs a series of hidden layers to classify characters. The system is designed to recognize uppercase letters, numerals, and five special characters.
  • Postprocessing Module: This rule-based module is crucial for disambiguation and error correction. It first checks for inconsistencies in the number of detected characters versus the field width. It then organizes competition between predicted character hearts, especially for closely positioned ones. Crucially, it leverages application-specific knowledge (like valid license plate patterns and control characters for DVLA forms) to validate candidate character strings and assign confidence levels.

Datasets

The paper mentions that both training and testing datasets were drawn from field data collected at the Driver and Vehicle Licensing Agency (DVLA) in the United Kingdom.

  • Training Set: Approximately 60,000 characters from 8,000 fields.
  • Testing Set: About 5,000 characters from 650 fields.

The authors note that the test data included fields with touching characters, making it a realistic evaluation scenario.

Key Insights and Ideas

  • Positioning with an Array of Nodes: The innovation of using an array of output units for character heart detection, rather than a single node, significantly improves the robustness and speed of the positioning stage by allowing for averaging across window shifts.
  • Gabor Projections for Robustness: Employing Gabor projections as the initial feature extraction layer provides inherent noise resistance, contributing to higher recognition rates.
  • Application-Specific Postprocessing is Key: The success of the system hinges on a sophisticated postprocessing module that leverages domain knowledge (like license plate patterns and control characters) to resolve ambiguities and validate entire fields, rather than just individual characters. This is particularly important when individual character rejection is meaningless.
  • Handling Ambiguity: The system explicitly addresses the challenge of distinguishing between visually similar characters by grouping them into a single output class during training and relying on postprocessing for disambiguation.

Key Results

  • Character Level Success Rate (without rejects): Exceeded 96%.
  • Field Rate (without rejects): Above 85%.
  • Character Level Success Rate (with ~20% field rejection): 99.8%.
  • Field Rate (with ~20% field rejection): 99.5%.
  • Positioning Accuracy: More than 92% of character hearts were found within 1-pixel precision, and only 0.4% were missed by more than 4 pixels on the test set.
  • Speed: The system can run at 80 characters/second/board when implemented on specialized parallel hardware.

Limitations and Future Work

  • Limitations:
    • The system's performance can be affected by abnormally wide characters, which may sometimes lead to false alarms.
  • While the system can handle touching characters, the paper notes that in testing mode, the preprocessing module attempts to separate them if possible, and if not, the system performs at approximately 94% accuracy.
  • Future Work: Not explicitly detailed in the paper, but the success of the application-specific postprocessing suggests that tailoring such modules to different form types would be a natural extension. The robustness of the system also implies potential for broader applications beyond vehicle registration.

Summary updated

BibTeX
@inproceedings{1995kodak,
  title = {{KODAK lMAGELINK™ OCR Alphanumeric Handprint Module}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1995},
}
Figure