基于循环神经网络的图像特定文本抽取方法  

Extraction Method of Interest Text in Image Based on Recurrent Neural Network

作　　者：杨恒杰 闫铮[1] 邬宗玲 方定邦 段放 Yang Hengjie;Yan Zheng;Wu Zongling;Fang Dingbang;Duan Fang(College of Information Science and Engineering,Huaqiao University,Xiamen,Fujian 361021,China)

机构地区：[1]华侨大学信息科学与工程学院

出　　处：《激光与光电子学进展》2019年第24期180-187,共8页Laser & Optoelectronics Progress

基　　金：福建省自然科学基金（2017J01116）;华侨大学中青年培育计划（Z16J0070）;华侨大学科研基金（605-50Y18023）;华侨大学研究生科研创新能力培育计划（17014082026）

摘　　要：光学字符识别(OCR)难以针对图像中某些特定文本进行识别,尤其在实际场景中,识别结果通常会包含大量噪声文本。针对这一问题,提出一种基于循环神经网络的双向长短时记忆-条件随机场(BLSTM-CRF)模型。首先利用BLSTM网络捕获OCR识别结果中序列的上下文信息,得到特征序列;然后结合CRF建立模型特征与标签的关系,进行标签预测,通过标签即可得到特定文本。实验结果表明,该方法在场景图像数据集YNIDREAL上可以达到88.52%的准确率,相较于CRF模型,准确率提高了16.39个百分点,证明了本方法的可行性和稳健性。It is difficult to recognize a certain text of interest in the image using the optical character recognition(OCR)method;particularly in natural scenes,the recognition results usually contain a large number of noisy texts.To address this problem,a model termed bidirectional long short term memory-condition random field(BLSTMCRF)based on a recurrent neural network for extracting texts of interest is proposed in this study.First,a BLSTM network is implemented to capture the context information of the sequence obtained by the OCR method,thereby obtaining feature sequences.Second,the relationships between the model features and tags are established by introducing the CRF.Then the text of interest can be obtained through the tags.Experimental results indicate that the proposed method can achieve an accuracy of 88.52%on YNIDREAL dataset.Compared with the CRF model,the accuracy of the proposed method is improved by 16.39percentage points,which proves the feasibility and robustness of the proposed method.

关 键 词：机器视觉 特定文本抽取 光学字符识别 双向长短时记忆网络 条件随机场 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]