BHCDR: Real-Time Bangla Handwritten Characters and Digits Recognition using Adopted Convolutional Neural Network

Abstract

Machine learning algorithm suffers to recognize the Bangla handwriting from images because of the complex design, diversities among different writers and similarity between characters and digits. In recent times, deep learning is becoming very popular among the researchers for Bangla Handwriting Recgnition (BHR) because of its high efficiency i n t erms of memory, time complexity and robust feature extraction. This research aims at improving the performance of baseline Convolutional Neural Network (CNN) by increasing the recognition accuracy with minimizing the computational overhead; this paper presents a real-time Bangla Handwritten Characters and Digits Recognition (BHCDR) system using adopted CNN. Our proposed preprocessing technique, data augmentation and incorporating dropout filters i n t he b aseline C NN a rchitecture h ave achieved the goal. The proposed eight layered architecture has used two convolutional layers followed by two Maxpooling layers with 25% dropout filters from one layer to another and two fully connected layers with 50% dropout followed by a softmax classifier. The proposed model is trained and tested using 118,698 images of Bangla lekha-isolated dataset and 21000 images of CMATERdb dataset for Bangla hand-written characters and digits maintaining the ratio of 4:1 respectively. The proposed model has achieved the mean accuracy of 97.43% for classification with the average computational costs of 44.95 ms/f.