D³ Net: An Adam-Based Ensemble Voting Method for Bangla Handwritten Digit Recognition Using Multi-CNN Architectures

Bangla handwritten digit recognition is crucial for various applications, particularly in the financial sector, such as bank cheque processing, automated form filling, postal code recognition, and document verification, where accurate identification of handwritten digits ensures proper transaction handling. English handwritten digit recognition has made significant progress thanks to well-established datasets and methods. However, Bangla handwritten digit recognition remains challenging due to variations in handwriting styles, limited annotated datasets, and the complexity of the Bangla script. To overcome these challenges, we introduce D<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>Net, an ensemble voting method built on three distinct lightweight Convolutional Neural Network (CNN) architectures optimized using the Adam optimizer. Trained on the NumtaDB dataset, D<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>Net achieves a remarkable accuracy of 99.52% on the testing set and a Cohen’s Kappa score of 0.99, demonstrating its effectiveness in recognizing Bangla handwritten digits. The system utilizes a multi-CNN approach where each CNN extracts unique features, and the ensemble method combines their predictions to improve overall performance. This robust approach significantly improves recognition accuracy and reliability by effectively addressing the key challenges of handwriting variability, limited datasets, and the complex nature of Bangla digits, ensuring more consistent and precise results in recognition tasks.