A Classical Approach to Handcrafted Feature Extraction Techniques for Bangla Handwritten Digit Recognition

TL;DR

This paper evaluates handcrafted feature extraction techniques combined with classical classifiers for recognizing Bangla handwritten digits, achieving high accuracy across multiple datasets and comparing favorably with recent methods.

Contribution

It benchmarks four classifiers with three handcrafted features on four datasets, identifying the best combination (HOG+SVM) for Bangla digit recognition.

Findings

HOG+SVM achieved up to 98.08% accuracy on CMARTdb.

Fine-tuning classifiers improved recognition performance.

The proposed approach outperforms some recent state-of-the-art methods.

Abstract

Bangla Handwritten Digit recognition is a significant step forward in the development of Bangla OCR. However, intricate shape, structural likeness and distinctive composition style of Bangla digits makes it relatively challenging to distinguish. Thus, in this paper, we benchmarked four rigorous classifiers to recognize Bangla Handwritten Digit: K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Random Forest (RF), and Gradient-Boosted Decision Trees (GBDT) based on three handcrafted feature extraction techniques: Histogram of Oriented Gradients (HOG), Local Binary Pattern (LBP), and Gabor filter on four publicly available Bangla handwriting digits datasets: NumtaDB, CMARTdb, Ekush and BDRW. Here, handcrafted feature extraction methods are used to extract features from the dataset image, which are then utilized to train machine learning classifiers to identify Bangla handwritten digits. We further fine-tuned the hyperparameters of the classification algorithms in order to acquire the finest Bangla handwritten digits recognition performance from these algorithms, and among all the models we employed, the HOG features combined with SVM model (HOG+SVM) attained the best performance metrics across all datasets. The recognition accuracy of the HOG+SVM method on the NumtaDB, CMARTdb, Ekush and BDRW datasets reached 93.32%, 98.08%, 95.68% and 89.68%, respectively as well as we compared the model performance with recent state-of-art methods.