Rapid Feature Extraction for Optical Character Recognition

M. Zahid Hossain, M. Ashraful Amin, Hong YanarXiv 2012 · arXiv

Optical Character Recognition (OCR) for handwritten text remains a significant challenge due to the inherent variability in human writing styles. While mature solutions exist for Latin scripts, scripts like Bangla, used by hundreds of millions, still lack highly efficient and accurate OCR systems. This paper presents a "Celled Projection" (CP) method designed to bridge the gap between computational speed and recognition accuracy.

The Challenge of Feature Extraction

The core of any OCR system is feature extraction—the process of converting a character image into a mathematical descriptor. Traditionally, researchers have chosen between two paths:

  1. Structural Features: These describe geometric properties like loops and endpoints. While robust, they often require "thinning" (skeletonization), which is computationally expensive and prone to errors during processing.
  2. Statistical Features: These measure properties like pixel density. They are faster but often lack the local detail needed to distinguish between similar characters under distortion.

The proposed Celled Projection method aims to combine the speed of statistical methods with the local structural awareness typically found in more complex approaches.

Methodology: Celled Projection (CP)

Standard projection histograms sum foreground pixels across entire rows or columns, losing local spatial information. The CP method introduces a "divide and conquer" strategy:

  1. Partitioning: The input image $G$ is divided into $k$ regions or "cells."
  2. Local Projections: Projections are calculated independently for each specific region.
  3. Concatenation: These local values are combined into a single feature vector $V$.

This approach captures local geometric information that global histograms miss, making the system more resilient to slants and minor handwriting variations. The algorithm operates directly on binarized images using simple arithmetic, avoiding the overhead of skeletonization.

Pseudocode for the Horizontal Celled Projection algorithm

Experimental Setup

The method was evaluated using a dataset of 12,000 Bangla handwritten numeral samples from 120 writers. Images were normalized to a $16 \times 16$ resolution. The study compared CP against five established statistical methods: Crossings, Fourier Transforms, Moments, Global Projection Histograms, and Zoning.

To ensure the robustness of the features, three distinct classifiers were employed:

  • k-Nearest Neighbor (KNN)
  • Probabilistic Neural Networks (PNN)
  • Feed-Forward Backpropagation Networks (FBPN)

Results and Analysis

The experimental results demonstrate that the Celled Projection method significantly outperforms traditional statistical features across all tested classifiers.

Recognition rates (%) for various feature extraction methods across different classifiers

As shown in the table, the CP method (specifically using 4 horizontal and 4 vertical cells) achieved a peak recognition rate of 94.12% with the PNN classifier. In contrast, standard methods like Moments and Fourier Transforms struggled, peaking at approximately 68% and 73%, respectively.

Notably, CP maintained high accuracy even with the simplest classifier (KNN at 94.10%), suggesting that the feature extraction itself provides high class separability. Furthermore, the training process for CP with the FBPN classifier was twice as fast as other feature sets on average, highlighting its suitability for real-time commercial applications.

Conclusion

The Celled Projection method provides a rapid and accurate solution for handwritten numeral recognition. By focusing on local projections, it achieves robustness against handwriting distortions without the computational cost of structural analysis. While currently validated on numerals, future work will explore scaling this approach to the full, complex Bangla alphabet and higher-resolution imagery to handle more intricate scripts.

Summary updated

BibTeX
@misc{hossain2012rapid,
  author = {M. Zahid Hossain and M. Ashraful Amin and Hong Yan},
  title = {{Rapid Feature Extraction for Optical Character Recognition}},
  year = {2012},
  eprint = {1206.0238},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1206.0238},
}
Figure