Abstract
A critical preprocessing step for Optical Character Recognition (OCR) systems is the separation of handwritten annotations from machine-printed text. This paper presents a word-level classification system designed to address this challenge in South Indian documents. By framing text separation as a texture analysis problem, the authors extract nine statistical texture features and employ a k-Nearest Neighbors (k-NN) classifier. The proposed method achieves an average accuracy of 99.26% across multiple scripts, including Kannada, Telugu, Malayalam, Hindi, and Roman, demonstrating significant script independence without the need for complex structural rules.
Introduction
In multilingual environments like India, documents such as application forms, bank cheques, and annotated research papers frequently contain a mix of handwritten and machine-printed text. Effectively separating these two types of text is essential for downstream OCR tasks. While significant research has focused on Roman, Arabic, and Chinese scripts, South Indian scripts have remained relatively under-explored.
Existing methods often rely on script-dependent features like stroke direction or chain codes, which lack generalizability. This research proposes a robust, script-independent approach that operates at the word level, striking a balance between the fragility of character-level analysis and the imprecision of block-level separation.
Methods
The core innovation of this work is treating text separation as a texture analysis problem. This approach mirrors human perception, which often distinguishes handwriting from print based on the overall "texture" or structural arrangement of ink rather than individual character recognition.
The pipeline consists of three primary stages:
- Preprocessing & Segmentation: Documents undergo binarization via Otsu’s method. Morphological operators are applied to remove noise and long lines. Word segmentation is then performed using dilation and connected component analysis to define bounding boxes for individual words.
- Feature Computation: For each word image, nine statistical texture features are calculated based on intensity histograms and statistical moments: mean, standard deviation, smoothness, third moment (skewness), uniformity, entropy, local standard deviation, local range, and local entropy.
- Classification: These 9-dimensional feature vectors are processed by a k-Nearest Neighbors (k-NN) algorithm to categorize each word as either handwritten or printed.
Datasets
To evaluate the system, the authors compiled a comprehensive multilingual dataset of 10,000 word images (5,000 printed and 5,000 handwritten).
- Custom South Indian Dataset: Comprising 8,000 word images, this dataset includes samples of Kannada, Telugu, Malayalam, and Hindi (Devanagari) collected from 100 different writers across various professions.
- IAM DB 3.0: To ensure extensibility to Western scripts, 2,000 English/Roman word images were integrated from this public database.
Key Results
The system was evaluated using 10-fold cross-validation. With $k=5$, the k-NN classifier achieved a peak average accuracy of 99.26%. Script-specific performance was remarkably high:
- 100% accuracy for Hindi, Telugu, and Malayalam.
- 98.55% accuracy for Kannada.
- 98.00% accuracy for Roman scripts.
The confusion matrix revealed only 74 misclassifications out of the 10,000-word sample, highlighting the robustness of the statistical texture features.
Discussion and Conclusion
The success of this method underscores the power of simple statistical moments when applied to the right problem representation. By prioritizing texture over geometric shape, the system achieves high accuracy across diverse scripts without requiring language-specific tuning.
However, limitations exist. Misclassifications typically occur when handwriting is exceptionally neat, mimicking the uniformity of print, or when printed text is degraded. Future work will focus on benchmarking against other state-of-the-art methods and expanding the dataset to include more overlapping text scenarios. The authors also intend to release their custom South Indian script dataset to support further research in this under-resourced area.
Summary updated