计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (12): 141-148.DOI: 10.3778/j.issn.1002-8331.1903-0192

• 模式识别与人工智能 • 上一篇    下一篇

双网络模型下的智能医疗票据识别方法

郑祖兵,盛冠群,谢凯,唐新功,文畅,李长晟   

  1. 1.长江大学 电工电子国家级实验教学示范中心,湖北 荆州 434000
    2.长江大学 电子信息学院,湖北 荆州 434023
    3.油气资源与勘探技术教育部重点实验室(长江大学),武汉 430100
    4.长江大学 计算机科学学院,湖北 荆州 434023
  • 出版日期:2020-06-15 发布日期:2020-06-09

Intelligent Medical Bills Recognition Method Based on Combined Network Model

ZHENG Zubing, SHENG Guanqun, XIE Kai, TANG Xingong, WEN Chang, LI Changsheng   

  1. 1.National Demonstration Center for Experimental Electrical and Electronic Education, Yangtze University, Jingzhou, Hubei 434000, China
    2.College of Electronic Information, Yangtze University, Jingzhou, Hubei 434023, China
    3.Key Laboratory of Exploration Technologies for Oil and Gas Resources(Yangtze University), Ministry of Education, Wuhan 430100, China
    4.College of Computer Science, Yangtze University, Jingzhou, Hubei 434023, China
  • Online:2020-06-15 Published:2020-06-09

摘要:

为了满足医疗行业大量针式票据录入工作的需求,解决传统人工录入方式效率低、精度低的问题,构建了双网络模型下的针式打印字体医疗票据识别方法。传统目标检测网络的参数同时描述了目标的位置与类别信息,其用于大规模定位识别任务时由于参数量庞大导致网络极难以训练,为解决以上问题,提出了双网络模型方法以联合FasterRCNN与深度卷积神经网络实现票据中字符的定位与识别,双网络将定位与识别分步进行以降低任务的复杂度。实验采用自建票据数据集与字库数据集进行网络训练,利用现场采集的票据验证了算法的有效性,通过测试不同参数下模型的性能来选定最佳参数,并对比分析了该方法与传统方法的识别效果。实际测试表明,识别准确率达95.4%,召回率达92.7%,速度达0.76 s/张。

关键词: 深度学习, 卷积神经网络, 票据识别, 票据校正, 字符识别, 文本定位

Abstract:

This bill recognition method based on combined network has been constructed to satisfy the tremendous needs of bill entering and solve the problems of low efficiency and low accuracy of data inputting with traditional man-powered method in medical industry. The parameters of the traditional target detection network describe the location and category of the target at the same time, it is hard to train the network because of the large number of parameters, a recognition method combining Faster RCNN and deep convolutional neural network is proposed to solve those problems, and the combined network locates and identifies targets step by step to lower the complexity of the task. The self-built bill training set and character-based training set are used to do network training, the validity of the algorithm is verified by the bill collected on site, the optimal parameters are selected by testing the performance of the model under different parameters, and the performance of this method is analyzed with the traditional method. It shows that the recognition accuracy reaches 95.4%, the recall rate reaches 95.4%, and the speed reaches 0.76 secends per piece.

Key words: Deep Learning, Convolutional Neural Network(CNN), bill recognition, bill correction, character recognition, text location