Handwritten Bangla character and numeral recognition using convolutional neural network for low-memory GPU

In this work, a convolutional neural network (CNN) based architecture is proposed for low memory GPU to recognize the handwritten isolated Bangla characters and numerals. The merit of the proposed architecture is the lesser number of trainable parameters as compared to the standard deep architectures and enabling it to train the proposed architecture on the low-memory GPU. The features from various layers of CNN are fused to handle the multi-scale nature of a character. The spatial pyramid pooling on the fused features produces a fixed size feature vector. It helps to reduce the number of parameters of the proposed model. Extensive experiments have been conducted on various versions of publicly available Bangla character dataset CMATERdb. The proposed architecture yields competitive results as compared to the fine-tuned standard deep architectures such as AlexNet, VGGNet, and GoogLeNet.