In spite of advances in object recognition technology, Handwritten BanglaCharacters Recognition (HBCR) (such as alpha-numeric and special) remainslargely unsolved due to the presence of many ambiguous handwritten charactersand excessive cursive in Bangla handwritings. Even the best existingrecognizers do not lead to satisfactory performance for practical applications,and have much lower performance than those developed for English alpha-numericcharacters. To improve the performance of HBCR, we herein present Banglahandwritten characters recognition methods by employing the state-of-the-artDeep Convolutional Neural Networks (DCNN) including VGG Network, AllConvolution Network (All-Conv Net), Network in Network (NiN), Residual Network,FractalNet, and DenseNet. The deep learning approaches have the advantage ofextracting and using feature information, improving the recognition of 2Dshapes with a high degree of invariance to translation, scaling and otherdistortions. We systematically evaluated the performance of DCNN models onpublicly available Bangla handwritten character dataset called CMATERdb, andachieved the state-of-the-art recognition accuracy when using DCNN models. Suchimprovement fills a significant gap between practical requirements and theactual performance of Bangla handwritten characters recognizers.
展开▼
