R-UNet++: A Local Segmentation Network for the Classification of Oracle Bone Materials
Weize Gao, Shanxiong Chen, Bofeng Mo, Yang Ye, Benpeng Su · Journal of Computer-Aided Design & Computer Graphics · 2022
The rejoining of oracle bone (OB) fragments with the same material is a significant topic in the field of oracle bone inscriptions research. Facing to the challenges of OB materials classification, a classification framework based on R-UNet++ is proposed. R-UNet++ draws on the dense convolution block link in UNet++, and on this basis introduces the improved strategies by attention module, bilinear upsampling method and residual unit, which effectively inhibits the noise response generated by multi-scale feature fusion and enhances the fine- grained segmentation capability in final. In this classification framework, first, R-UNet++ is able to accurately divide the difference information between classes. Then, ResNet50 is adopted to further extract features from images segmented by R-UNet++ and completes the OB materials classification task. With the real OB dataset, segmentation and classification experimental results show that R-UNet++ has strong ability to accomplish the exact segmentation task. Compared with other state-of-the-art classification networks, the classification accuracy has a higher improvement, which fully verifies the feasibility and efficiency of proposed classification framework.