TLDR: A research paper presents a deep convolutional neural network (CNN) model for recognizing handwritten Devanagari characters. Using the Devanagari Handwritten Character Dataset (DHCD) of 61,000 images across 36 classes, the model achieved 96.36% accuracy on testing data and 99.55% on training data, demonstrating a highly effective automated approach for digitizing handwritten Hindi text.
Handwritten character recognition is a field gaining significant attention due to its potential to enhance technological tools like search engines and recommender systems. While many languages have advanced digitization tools, the ancient Indian Devanagari script, used for over 120 languages, has lacked proper solutions for digitizing handwritten text. This gap presents a challenge, especially given the script’s widespread use and the need to preserve and access historical data.
A recent research paper, “Devanagari Handwritten Character Recognition using Convolutional Neural Network,” by Diksha Mehta and Prateek Mehta, addresses this challenge by proposing an automated approach to extract and recognize handwritten Hindi characters from images of Devanagari script. Their work aims to save time and digitize obsolete data, contributing significantly to the preservation and accessibility of information in this widely used script. You can read the full paper here: Devanagari Handwritten Character Recognition using Convolutional Neural Network.
The Challenge of Devanagari
Devanagari, an abugida script, is the world’s fourth most extensively used writing system, comprising 47 main letters (14 vowels and 33 consonants). Digitizing handwritten Devanagari is particularly complex due to the inherent variations in individual handwriting styles. Furthermore, several Devanagari characters look very similar, differing only by subtle features like dots or horizontal lines, which adds to the difficulty of accurate recognition.
Leveraging Convolutional Neural Networks
The researchers employed a deep convolutional neural network (CNN) for this task. CNNs are a powerful type of deep learning model particularly well-suited for image processing tasks like character recognition. Unlike traditional methods that require explicit feature definition, CNNs can learn the most relevant features directly from raw pixel data. This capability makes them highly effective for complex pattern recognition.
The proposed system utilizes two deep convolutional neural network layers. A CNN typically consists of several key layers:
- Convolutional Layer: This is the core of a CNN, where filters (kernels) slide over the input image to detect patterns like edges or textures. This layer significantly reduces the number of trainable parameters compared to fully connected networks, making training more feasible.
- Rectified Linear Unit (ReLU): An activation function applied element-wise, helping the network learn complex, non-linear relationships.
- Pooling Layer: This layer reduces the size of the image representations, making them more manageable and helping to prevent overfitting. It achieves this by down-sampling, for example, by taking the maximum value from a small region (max-pooling).
- Fully Connected Layer: Similar to a traditional neural network, these layers connect all neurons from the previous layer, processing the extracted features to make the final classification.
The Devanagari Handwritten Character Dataset (DHCD)
A crucial component of this research is the Devanagari Handwritten Character Dataset (DHCD). This open dataset contains 61,000 images of 36 Devanagari characters, with each character class having 1,700 images. The dataset was meticulously prepared by scanning handwritten texts from various authors, ensuring a wide range of handwriting variations. The dataset was split into 80% for training (48,960 images) and 20% for testing (12,240 images). Before training, images were preprocessed by converting them to grayscale, inverting pixel intensities (white character on a dark background), and standardizing the background to zero-rate pixels.
Promising Results
The model was trained over 30 epochs. The experimental results demonstrated high accuracy. During training, the model achieved an impressive accuracy of 99.55% with a loss of 0.0145. For the testing phase, which validates the model’s performance on unseen data, it achieved a recognition accuracy of 96.36% with a loss of 0.2102. These results indicate that the deep convolutional neural network approach is highly effective for Devanagari handwritten text recognition.
Also Read:
- Advanced AI Combines Neural Networks for Superior Handwritten Digit Recognition
- Unlocking History: How AI Extracts Information from Handwritten Birth Certificates
Future Scope
The researchers conclude that while the model performed very well, there is always room for improvement in machine learning models. Future work could involve training the model with even larger and more diverse datasets to further enhance its recognition capabilities and robustness.
