Mai, Christopher

Oracle bone inscriptions (OBIs) are one of the oldest characters in the world and are the predecessors of today's Chinese characters. These oracle characters recorded various human activities of the time and provide insights into Chinese history. To date, almost 4,500 different oracle characters have been discovered, with deciphering still being carried out by people with specialist knowledge. This process is labor-intensive and time-consuming, with around 2,300 characters still to be deciphered. Furthermore, the inscriptions have become increasingly illegible as a result of the aging process, frequently exhibiting characteristics such as noise or incompleteness. To address these issues, in this paper, we present a new convolutional neural network architecture for recognizing OBIs. It is based on the idea of Inception modules and the use of residual connections. To increase the diversity in the dataset, data augmentation techniques were applied. Together with these techniques, the presented architecture achieves an accuracy of 95.93%. For the purpose of comparability, known pre-trained architectures such as InceptionV3, ResNet50, and Inception-ResNet-V2 were used for comparison. The results demonstrate that the proposed architecture exhibits superior performance compared to these models across multiple evaluation metrics while simultaneously establishing a new benchmark on the Oracle-MNIST dataset.