Convolutional Neural Network with Attention Mechanism for Historical Chinese Character Recognition

Haoyu Qin, Liangrui Peng

Abstract: Historical Chinese character recognition is of great significance in digital library applications. This paper proposes to incorporate feature-based attention mechanism into convolutional neural network (CNN) towards a better solution for historical Chinese recognition. A modified GoogLeNet is adopted as the network architecture integrated with batch normalization layers. A squeeze-and-excitation module is utilized after each inception module to learn the weights of attention mechanism. A feature normalization layer is also added before the softmax layer, and compared with different discriminant loss functions including contrastive loss, triplet loss and center loss. Experimental results on both CASIA HWDB offline handwritten Chinese character datasets and historical Chinese samples show that the proposed method achieves improved recognition performance.