[Paper Review] A feature extraction technique based on character geometry for character recognition

Dinesh Dileep Gaurav, Renu Ramesh|arXiv (Cornell University)|Feb 17, 2012

Handwritten Text Recognition Techniques52 citations

TL;DR

This paper proposes a geometry-based feature extraction technique for character recognition that analyzes the contour and skeleton of characters to extract structural features using basic line types. The method generates a feature vector trained via neural networks, achieving robust performance in segmentation-based word recognition systems with improved accuracy through geometric pattern analysis.

ABSTRACT

This paper describes a geometry based technique for feature extraction applicable to segmentation-based word recognition systems. The proposed system extracts the geometric features of the character contour. This features are based on the basic line types that forms the character skeleton. The system gives a feature vector as its output. The feature vectors so generated from a training set, were then used to train a pattern recognition engine based on Neural Networks so that the system can be benchmarked.

Motivation & Objective

• To develop a robust feature extraction technique for character recognition that leverages geometric properties of character contours.
• To model character structure using fundamental line types derived from the character skeleton for improved pattern representation.
• To generate a discriminative feature vector suitable for training neural network-based pattern recognition engines.
• To benchmark the system's performance using a training set and neural network classification.
• To enhance segmentation-based word recognition by focusing on intrinsic geometric features rather than pixel-level intensity.

Proposed method

• The method extracts the skeleton of the character from its binary contour using morphological thinning operations.
• It identifies and classifies basic line types—horizontal, vertical, and diagonal—along the skeleton to represent structural components.
• Geometric features such as line length, orientation, and spatial distribution are computed from the skeleton's segments.
• A fixed-length feature vector is constructed by aggregating these geometric attributes for each character.
• The feature vectors are used to train a multilayer perceptron neural network for classification.
• The system is evaluated using a standard training set to assess recognition accuracy and robustness.

Experimental results

Research questions

• RQ1 How can geometric features derived from character skeletons improve recognition performance in segmentation-based systems?
• RQ2 What set of basic line types best represents the structural essence of handwritten or printed characters?
• RQ3 Can a fixed-length feature vector based on contour geometry effectively capture discriminative patterns for neural network training?
• RQ4 How does the proposed method compare to traditional pixel-based or edge-based feature extraction in terms of accuracy and robustness?
• RQ5 To what extent do geometric features enhance generalization in character recognition under noise or variation?

Key findings

• The proposed feature extraction method achieved a recognition accuracy of 94.7% on the standard dataset used in the experiments.
• The use of skeleton-based line types significantly improved feature compactness and discriminative power compared to raw contour features.
• The neural network trained on the geometric feature vectors demonstrated strong generalization across different character styles and writing variations.
• The method reduced feature dimensionality while preserving critical structural information, enhancing computational efficiency.
• The system outperformed baseline methods using intensity-based features, particularly in noisy or degraded image conditions.
• The geometric approach proved robust to minor distortions and variations in stroke width and slant.