Intelligent Recognition of Bangla Handwritten Digits

Optical Character Recognition (OCR) devices are essential for digitizing printed and handwritten materials because they transform text into machine-readable formats. Although OCR technologies have seen remarkable advancements in lan- guages such as English, Tamil, and the Bangla script continues to pose significant challenges due to its structural intricacy and the limited availability of high-quality datasets and advanced recog- nition models. Addressing these challenges, this study evaluates the performance of deep Convolutional Neural Network (CNN) architectures namely InceptionV3, ResNet50, and DenseNet121 for Bangla handwritten numeral recognition using the NumtaDB dataset, which is thorough and varied benchmark dataset es- pecially selected for Bangla handwritten digits. We also used an ensemble learning approach to further improve recognition performance by combining the outputs of separate models and utilizing their complementing advantages to increase generaliza- tion and accuracy. Among the models evaluated, InceptionV3 achieved the highest individual accuracy of 99.86% followed closely by ResNet50 and DenseNet121, which obtained 99.73% and 99.69% respectively. The ensemble approach outperformed all standalone models, achieving a superior test accuracy of 99.97%. These results underscore the effectiveness of CNN-based architecture for Bangla numeral recognition and demonstrate the considerable potential of ensemble learning in enhancing the accuracy, robustness, and reliability of OCR systems designed for complex scripts like Bangla.