SondhiNet: A CNN-Classifiers Framework for Enhanced Bangla Handwritten Character Recognition with Ensemble Integration

Bangla, one of the most widely spoken languages globally with approximately 265 million speakers, holds profound cultural and linguistic significance. Bangla handwritten character recognition (BHCR) is a vital step in the development of Optical Character Recognition (OCR) systems for preserving and digitizing this rich linguistic heritage. BHCR has critical applications in fields such as document digitization, educational tools, and automation in government and private sectors. However, BHCR presents unique challenges due to the intricate nature of the script, including visually similar characters, complex compound shapes, and a lack of balanced datasets for effective training. Traditional Convolution Neural Network (CNN) and transfer learning methods often struggle to address these complexities adequately. To overcome these limitations, we introduce SondhiNet, a novel framework that integrates CNN-based feature extraction with dimensionality reduction through Principal Component Analysis (PCA), followed by classification using Decision Trees (DT) and Support Vector Machines (SVM). SondhiNet improves accuracy by leveraging an ensemble (voting) mechanism to aggregate predictions effectively. Evaluated on three balanced datasets—Ekush with 122 classes, BanglaLekha with 84 classes, and CMATERdb with 231 classes—SondhiNet achieved accuracies of 97.92%, 97.64%, and 98.47%, respectively. These results underscore the effectiveness of our approach in addressing the challenges of Bangla handwritten character recognition and advancing its practical applications.