基于Faster-RCNN的水书古籍手写文字的检测与识别

Detection and recognition of handwritten characters in Shuishu ancient books based on Faster-RCNN

中国非物质文化遗产水书文化面临失传威胁,近年大量深度学习的方法用于手写古籍文字的识别.但水书古籍文字识别面临数据集建立和标注困难、样本不平衡等问题,研究进展不大,且鲜少进行水书古籍页面级的文字检测与识别.首先建立了一个较大规模的水书手写文字数据集,通过几种数据扩增方式,获得包含80个文字类别,共110610个带标签的字符样本.将Faster-RCNN(faster-region based convolutional neural network)算法应用到水书古籍文字识别研究上,以不同组合的数据集作为输入进行实验,在全部80个目标类别上获得了91.95%的平均识别率,实现了页面级的端到端的水书古籍文字的准确定位与识别.实验结果表明,Faster-RCNN模型在目前的数据集上能很好地实现水书手写文字的检测与识别,文中采用的数据扩增方式能明显提升水书手写文字的识别率,为水书文化的保护和传承提供了新思路,对于解决实际应用场景中的水书文字识别问题具有重要意义.

The Shuishu culture of China's intangible cultural heritage is facing the threat of loss.In recent years,a large number of deep learning methods have been used to recognize characters in handwritten ancient books.However,the character recognition of Shuishu ancient books faces problems such as dataset establishment and labeling,sample imbalance,etc.,and there is not much research progress.In this paper,a large scale Shuishu handwritten character dataset is firstly established,discussed several data augmentation methods,and obtained 110610 labeled character samples containing 80 categories.Faster-RCNN(fast region-based convolutional neural network)was applied to the recognition of Shuishu handwritten characters.Experiments were carried out with different combinations of datasets as input,and an average recognition rate of 91.95%was obtained on all 80 target classes,achieving end-to-end accurate localisation and recognition of Shuishu handwritten characters at the page level.The experimental results show that the Faster-RCNN can achieve good detection and recognition of Shuishu handwritten characters on the current dataset,and the data augmentation method used in this paper can significantly improve the recognition rate of Shuishu handwritten characters.This paper provides new ideas for the preservation and transmission of Shuishu culture,and is important for solving the problem of Shuishu character recognition in practical application scenarios.