Research on Segmentation and Recognition of Oracle Topographies Based on Deep Learning and Image Processing

This paper proposes a new technical approach to the problem of automatic segmentation and recognition of single characters in original oracle bone topography. The problem of mis-segmentation in oracle bone image segmentation is effectively solved by the super-pixel method and feature-based image processing technique. Meanwhile, the computational efficiency and recognition accuracy of the model are improved by utilizing the Faster-RCNN algorithm and VGG-16 network architecture. For the small sample learning problem, this paper adopts the siamese similarity network (SSN) based on deep metric learning to improve the recognition performance of oracle bone characters. Finally, the balance between model learning efficiency and recognition accuracy is achieved by fine-tuning the model parameters.