[Paper Review] Handwritten Bangla Digit Recognition Using Deep Learning

Md Zahangir Alom, Paheding Sidike+2 more|arXiv (Cornell University)|May 7, 2017

Handwritten Text Recognition Techniques1 references80 citations

TL;DR

This paper evaluates multiple deep learning models for Handwritten Bangla Digit Recognition on CMATERdb 3.1.1, finding CNN with Gabor features and dropout achieves the best accuracy of 98.78%.

ABSTRACT

In spite of the advances in pattern recognition technology, Handwritten Bangla Character Recognition (HBCR) (such as alpha-numeric and special characters) remains largely unsolved due to the presence of many perplexing characters and excessive cursive in Bangla handwriting. Even the best existing recognizers do not lead to satisfactory performance for practical applications. To improve the performance of Handwritten Bangla Digit Recognition (HBDR), we herein present a new approach based on deep neural networks which have recently shown excellent performance in many pattern recognition and machine learning applications, but has not been throughly attempted for HBDR. We introduce Bangla digit recognition techniques based on Deep Belief Network (DBN), Convolutional Neural Networks (CNN), CNN with dropout, CNN with dropout and Gaussian filters, and CNN with dropout and Gabor filters. These networks have the advantage of extracting and using feature information, improving the recognition of two dimensional shapes with a high degree of invariance to translation, scaling and other pattern distortions. We systematically evaluated the performance of our method on publicly available Bangla numeral image database named CMATERdb 3.1.1. From experiments, we achieved 98.78% recognition rate using the proposed method: CNN with Gabor features and dropout, which outperforms the state-of-the-art algorithms for HDBR.

Motivation & Objective

• Motivate improved recognition of handwritten Bangla digits due to variability in writing styles.
• Compare deep learning approaches (DBN, CNN variants) for HBDR without extensive feature engineering.
• Assess the impact of dropout and Gabor/filter features on recognition performance.
• Benchmark against state-of-the-art methods on CMATERdb 3.1.1 to establish a strong baseline.

Proposed method

• Evaluate Deep Belief Networks (DBN) and Convolutional Neural Networks (CNN) on CMATERdb 3.1.1.
• Explore CNN variants including dropout, Gaussian filters, and Gabor filters.
• Describe CNN architecture: two convolution layers, two subsampling layers, and one fully connected classification layer.
• Use RBM-based pretraining for DBN with contractive divergence learning.
• Train and test across multiple iterations to compare performance with SVM and other methods.

Experimental results

Research questions

• RQ1 How do different deep learning architectures (DBN vs CNN) perform on HBDR for Bangla digits?
• RQ2 Does applying dropout and using Gabor or Gaussian filters improve CNN-based HBDR performance?
• RQ3 How does the best DL approach compare to state-of-the-art methods on CMATERdb 3.1.1?

Key findings

• CNN with Gabor features and dropout achieves the highest reported accuracy of 98.78%.
• CNN with random Gaussian filters yields 97.70% accuracy, and CNN with Gabor yields 98.30%.
• CNN with dropout and Gaussian filters achieves 98.64% accuracy, outperforming standard CNNs.
• DBN achieves 97.20% accuracy, outperforming SVM at 95.50%.
• Among evaluated methods, CNN with Gabor + Dropout outperforms state-of-the-art approaches on the same dataset.
• CNN-based approaches outperform previous non-DL methods on CMATERdb 3.1.1 for HBDR.