[Paper Review] Multistage Hybrid Arabic/Indian Numeral OCR System

Yasser M. Alginaih, Abdul Ahad Siddiqi|arXiv (Cornell University)|May 6, 2010

Handwritten Text Recognition Techniques12 references3 citations

TL;DR

This paper presents a multistage hybrid OCR system for recognizing printed Arabic and Indian numerals in postal codes, combining binary, zoning, and fuzzy feature extraction with Hamming Net, Euclidean Distance, and Fuzzy Neural Network classifiers. The system achieves a high recognition rate, outperforming existing methods in the literature, and is designed to improve accuracy and efficiency in automated mail sorting systems.

ABSTRACT

The use of OCR in postal services is not yet universal and there are still many countries that process mail sorting manually. Automated Arabic/Indian numeral Optical Character Recognition (OCR) systems for Postal services are being used in some countries, but still there are errors during the mail sorting process, thus causing a reduction in efficiency. The need to investigate fast and efficient recognition algorithms/systems is important so as to correctly read the postal codes from mail addresses and to eliminate any errors during the mail sorting stage. The objective of this study is to recognize printed numerical postal codes from mail addresses. The proposed system is a multistage hybrid system which consists of three different feature extraction methods, i.e., binary, zoning, and fuzzy features, and three different classifiers, i.e., Hamming Nets, Euclidean Distance, and Fuzzy Neural Network Classifiers. The proposed system, systematically compares the performance of each of these methods, and ensures that the numerals are recognized correctly. Comprehensive results provide a very high recognition rate, outperforming the other known developed methods in literature.

Motivation & Objective

• To develop a robust, high-accuracy OCR system for printed Arabic and Indian numerals used in postal codes.
• To reduce errors in automated mail sorting by improving recognition performance over existing systems.
• To investigate the effectiveness of combining multiple feature extraction and classification techniques for numeral recognition.
• To systematically compare different feature-classifier combinations for optimal recognition outcomes.
• To provide a scalable and efficient solution suitable for deployment in postal service automation.

Proposed method

• The system employs three feature extraction methods: binary features based on pixel patterns, zoning features dividing the numeral image into grid regions, and fuzzy features using fuzzy logic to handle variations in shape and distortion.
• Three classifiers are used in sequence: Hamming Net for fast initial classification, Euclidean Distance for similarity-based matching, and Fuzzy Neural Network for final refinement using fuzzy inference.
• The multistage architecture processes input images through feature extraction, followed by sequential classification stages to improve recognition confidence.
• Feature vectors from each method are evaluated independently and combined using a decision fusion strategy to enhance accuracy.
• The system is trained and tested on a dataset of printed Arabic and Indian numerals from real postal code samples.
• Performance is evaluated using standard recognition rate metrics, with results compared against known OCR methods in the literature.

Experimental results

Research questions

• RQ1 How do different feature extraction methods (binary, zoning, fuzzy) impact the recognition accuracy of Arabic and Indian numerals?
• RQ2 What is the comparative performance of Hamming Net, Euclidean Distance, and Fuzzy Neural Network classifiers in recognizing printed numerals?
• RQ3 Can a multistage hybrid system combining multiple features and classifiers outperform single-method OCR systems?
• RQ4 What is the optimal configuration of feature-classifier combinations for minimizing recognition errors in postal code processing?
• RQ5 To what extent does the proposed system reduce errors in automated mail sorting compared to existing OCR solutions?

Key findings

• The proposed multistage hybrid OCR system achieves a significantly higher recognition rate than previously reported methods in the literature.
• The combination of fuzzy features with the Fuzzy Neural Network classifier yielded the highest individual recognition accuracy among all tested configurations.
• The multistage architecture effectively reduced error rates by allowing error correction through sequential classification stages.
• The system demonstrated robustness to variations in font, size, and image quality commonly found in real-world postal code images.
• The use of multiple feature types improved system resilience to noise and distortion, especially in low-quality scans.
• The overall system performance outperformed existing OCR systems for Arabic and Indian numerals, particularly in postal code recognition tasks.