A deep learning approach for Bangla optical character recognition system

Abstract

Bangla Optical Character Recognition has emerged as a crucial aspect of document digitization. A wide range of sources are covered, such as computer-generated text, letterpress, typewriters, outdoor banners, posters, and handwritten documents sourced from diverse channels. Bangla character structure is very complex, handwritten documents pose a significant challenge in recognition. To overcome those challenges, we introduce the custom Convolutional Neural Network model using image classification to detect and recognize the characters. This model can learn from a large-scale Bangla handwritten dataset. BanglaLekha-Isolated[15] and Eksush[19] datasets are tested in our model. To compare the performance of the custom Convolutional Neural Network model, we used the two most popular deep learning model, ResNet-50 and VGG16. By implementing our model, we achieve 95.63% accuracy on Ekush dataset and 96.20% on BanglaLekha-Isolated dataset. Bangla OCR research field is not saturated and also challenging due to its character style is very much complex compared to other language. To conclude, ongoing research efforts are dedicated to advancing Bengali OCR technology, making it more reliable and accessible for users. As with any OCR system, the quality of input images and variations in writing styles remain critical factors, but progress is being made to overcome these challenges