Ancient character recognition with deep learning techniques

Abstract

Many deep learning models have achieved remarkable results in many areas, such as image classification and image generation. At the same time, with the increasing attention given to the digitization of ancient manuscripts, ancient character recognition has become one of the most fascinating research areas. In this article, we try some CNNs such as ResNet, VGG, AlexNet or simply CNN on the dataset named Oracle-MNIST, an open ancient character dataset. In addition, to improve the accuracy of the models, ensemble learning is also adopted. Compared with the accuracy, the number of model parameters and running time, it was found that one simple CNN model trained as a snapshot performed best, and the recognition accuracy rate reached 97.009%.