[Paper Review] Implicit segmentation of Kannada characters in offline handwriting recognition using hidden Markov models

Manasij Venkatesh, Vikas Majjagi+1 more|arXiv (Cornell University)|Oct 16, 2014

Handwritten Text Recognition Techniques12 references4 citations

TL;DR

This paper proposes an implicit segmentation approach for offline Kannada handwritten character recognition using continuous-density Hidden Markov Models (HMMs), leveraging the script's agglutinative nature to decompose complex characters into base shapes. By modeling character sequences with HMMs, the method avoids explicit segmentation, reduces classification complexity, and improves accuracy by 10% over independent character classification, achieving 61.0% on the Chars74k dataset, with further 4% improvement using additional training data.

ABSTRACT

We describe a method for classification of handwritten Kannada characters using Hidden Markov Models (HMMs). Kannada script is agglutinative, where simple shapes are concatenated horizontally to form a character. This results in a large number of characters making the task of classification difficult. Character segmentation plays a significant role in reducing the number of classes. Explicit segmentation techniques suffer when overlapping shapes are present, which is common in the case of handwritten text. We use HMMs to take advantage of the agglutinative nature of Kannada script, which allows us to perform implicit segmentation of characters along with recognition. All the experiments are performed on the Chars74k dataset that consists of 657 handwritten characters collected across multiple users. Gradient-based features are extracted from individual characters and are used to train character HMMs. The use of implicit segmentation technique at the character level resulted in an improvement of around 10%. This system also outperformed an existing system tested on the same dataset by around 16%. Analysis based on learning curves showed that increasing the training data could result in better accuracy. Accordingly, we collected additional data and obtained an improvement of 4% with 6 additional samples.

Motivation & Objective

• To address the challenge of high classification complexity in Kannada handwritten character recognition due to the large number of morphologically related characters.
• To overcome limitations of explicit segmentation in offline handwriting by enabling implicit segmentation through HMM modeling.
• To improve recognition accuracy by exploiting the agglutinative structure of Kannada script to increase effective training samples.
• To evaluate the impact of training data size on model performance and demonstrate data efficiency through learning curves.
• To establish a baseline for offline HWR in Kannada, a low-resource Indic script, with a robust feature and model framework.

Proposed method

• Uses a sliding window approach to extract gradient-based features from fixed-width vertical strips across character images, generating sequential observation vectors.
• Employs left-to-right continuous-density HMMs with tied state parameters to model character sequences, enabling joint training and segmentation.
• Applies a 10-state HMM architecture to reduce model complexity and improve training efficiency compared to a 15-state model.
• Uses Gaussian Mixture Models (GMMs) for observation density estimation, with component count optimized via validation accuracy.
• Applies implicit segmentation by allowing the HMM to discover underlying base shapes (e.g., consonants and diacritics) without requiring segmentation boundaries in training data.
• Performs data augmentation by collecting 6 additional samples per character for 500 characters, retraining the model to assess data impact.

Experimental results

Research questions

• RQ1 Can implicit segmentation via HMMs improve recognition accuracy in offline Kannada handwriting recognition by reducing the number of distinct character classes?
• RQ2 How does exploiting the agglutinative structure of Kannada characters through HMM-based modeling affect training data efficiency and model generalization?
• RQ3 To what extent does increasing training data size improve recognition performance, as indicated by learning curves?
• RQ4 How does the proposed HMM-based system compare to existing methods, such as DCT-based nearest neighbor classifiers, on the Chars74k dataset?
• RQ5 Can the implicit segmentation approach be extended to handle complex character combinations like Consonant-Consonant-Vowel (CCV) sequences in future work?

Key findings

• The proposed HMM-based implicit segmentation method achieved 61.0% recognition accuracy on the Chars74k dataset, representing a 10% improvement over treating all 569 characters as independent classes.
• The system outperformed a DCT-based nearest neighbor classifier (33.3%) by 15.9 percentage points, demonstrating superior performance on the same dataset.
• Learning curve analysis revealed that increasing training data could significantly improve accuracy, with a 4% improvement observed after adding 6 additional samples per character.
• The model with 10 states achieved better performance and faster training than a 15-state model, indicating that reduced parameter count improved generalization.
• Validation accuracy peaked at 10 GMM components, with a decline beyond that point, indicating overfitting due to excessive model complexity.
• The results suggest that the primary bottleneck in performance is insufficient training data, not model architecture, as further data collection led to measurable accuracy gains.