甲骨文公司
性格(数学)
计算机科学
Softmax函数
水准点(测量)
Oracle统一方法
班级(哲学)
人工智能
模式识别(心理学)
机器学习
情报检索
人工神经网络
程序设计语言
搜索引擎
数学
几何学
Web搜索查询
按示例查询
地理
大地测量学
作者
Jing Li,Qiufeng Wang,Rui Zhang,Kaizhu Huang
标识
DOI:10.1007/978-3-030-86549-8_16
摘要
Oracle bone characters are probably the oldest hieroglyphs in China. It is of significant impact to recognize such characters since they can provide important clues for Chinese archaeology and philology. Automatic oracle bone character recognition however remains to be a challenging problem. In particular, due to the inherited nature, oracle characters are typically very limited and also seriously imbalanced in most available oracle datasets, which greatly hinders the research in automatic oracle bone character recognition. To alleviate this problem, we propose to design the mix-up strategy that leverages information from both majority and minority classes to augment samples of minority classes such that their boundaries can be pushed away towards majority classes. As a result, the training bias resulted from majority classes can be largely reduced. In addition, we consolidate our new framework with both the softmax loss and triplet loss on the augmented samples which proves able to improve the classification accuracy further. We conduct extensive evaluations w.r.t. both total class accuracy and average class accuracy on three benchmark datasets (i.e., Oracle-20K, Oracle-AYNU and OBC306). Experimental results show that the proposed method can result in superior performance to the comparison approaches, attaining a new state of the art in oracle bone character recognition.
科研通智能强力驱动
Strongly Powered by AbleSci AI