Isolated Bangla handwritten character recognition with convolutional neural network

The handwritten character recognition (HCR) problem has been studied extensively during the last few decades with varying level of success. Although one of the earliest optical character recognition work was done using an artificial neural network, due to low computational speed and other computational resource constraints, researchers had to move away to formulate the HCR problem of different languages (including Bangla) using the hand crafted features based classification methods. The recent progress of deep learning technologies and significant developments of parallel computing hardware have established a solid platform for the researchers in producing state-of-the-art reliable performance in many fields. Even though deep learning is proving its applicability in different computer vision problems since AlexNet was proposed in 2012, due to lack of significantly large Bangla handwritten character dataset, Bangla HCR could not progress far. Recently, a few initiatives have been taken to build large scale Bangla handwritten character datasets and made them available for public use. In this paper, we propose a modified ResNet-18 architecture (a convolutional neural network architecture) in recognizing Bangla handwritten characters. The proposed method is applied to two recently created isolated Bangla handwritten datasets. The used datasets are relatively large and practical to apply deep learning. Using the proposed method, we achieve the state-of-the-art recognition performance.