Hybrid CNN-GRU model for high efficient handwritten digit recognition

Abstract

Recognition of handwritten digits is a challenging research topic in Optical Character Recognition (OCR) in recent years. In this paper, a hybrid model combining convolutional neural network (CNN) and gate recurrent units (GRU) is proposed, in which GRU is used to replace the CNN fully connected layer part to achieve high recognition accuracy with lower running time. In this model, the features of the original image are firstly extracted by the CNN, and then they are dynamically classified by the GRU. Experiment performed on MNIST handwritten digit dataset suggests that the recognition accuracy of 99.21% while the training time and testing time is only 57.91s and 3.54s, respectively.