VGG-16 Framework with Dropout for Oracle Bone Inscription Recognition Using Transfer Learning

Abstract

Oracle Bone Inscription (OBI) is a treasure of Chinese culture, as the research on it progresses, higher requirements are placed on its recognition. Due to the long history of oracle bone inscription and the high cost of labelling, its recognition belongs to the task of Few-Shot learning. This paper proposes a recognition network model based on VGG-16 framework with dropout. Transfer learning strategy using similar feature between oracle bone inscription and modern Chinese character is used to solve the issue of lack of training data. Experimental results on OBI376 dataset shows that the transfer strategy proposed is better than a direct transfer from ImageNet weights and a new training. After the feasibility analysis, different fine-tuning strategies are tried to pursue highest accuracy on test dataset on OBI376. Then two dropout operations are added between the three fully connected layers in the model trained through the fine-tuning strategies with highest accuracy. Comprehensive parameter adjustment for the dropout rate is conducted to obtain the optimal model under this framework. Finally, a recognition model with 98.22% accuracy on test dataset is obtained in this paper, which shows that it has good generalization ability and the transfer strategy proposed is indeed useful for oracle bone inscription recognition.