Hybrid Deep Learning Model for Handwritten Bangla and English Digit Recognition
M. Ali Akbar, Md. Saiful Islam · 2023
The majority of the works on character recognition are based on manual feature extraction and then transferring them into respective classifiers or using CNN solely. To overcome the limitations of manual feature extraction, CNN can be used as an automated feature extractor. On the other hand, many noteworthy machine learning algorithms are capable of beating CNN in the classification methods. So, to resolve the problem in both cases three hybrid model systems titled CNN-KNN, CNN-SVM, and CNN-Decision Tree respectively are proposed in this work to recognize handwritten Bangla and English Digit recognition. A CNN model is also proposed which solely works as a feature extractor and a classifier to see the difference in results with proposed hybrid models. BanglaLekha-Isolated dataset for handwritten Bangla digit and MNIST dataset for handwritten English digit is used to train the proposed models. The accuracy rate of the proposed CNN model is 96.4 % for Bangla and 98.5% for English digit recognition. The accuracy rate of the proposed hybrid CNN-KNN, CNN-SVM, and CNNDecision Tree models are 97.2%, 96.7% & 92.7% for Bangla, and 98.7%, 98.7%, and 96.3% for English digit recognition respectively.