CNN-based Bronze Inscriptions Character Recognition

Bronze inscriptions characters are always recognized manually by paleographic professionals, which is a very hard and time-consuming task. Automatic computer recognition can be very helpful to this problem. Bronze inscriptions character recognition by computer is a classic image classification task which is the speciality of convolutional neural network (CNN). In this paper, a targeted model was designed and tested based on CNN-related technologies. Specifically, we explored many advanced CNN models and modified them to accommodate our needs as our base models to recognize bronze inscriptions characters. Then we brought in attention mechanism for this model by adopting spatial transformer network (STN) which was a learnable component to perform spatial transformation to attention the important regions in images. Furthermore, a robust loss function was introduced to implement implicit semantic data augmentation (ISDA) to help regularize the whole model and improve the final performance. A new large bronze inscriptions character dataset was used in our experiments helping our model get an excellent performance. A lot of experiments with detailed analysis were conducted and the final 91.21% accuracy showed the feasibility and effectiveness of our work.