基于STN-CRNN的自然场景英文文本识别研究  

Research on English Text Recognition in Natural Scenes Based on STN-CRNN

作　　者：汪洪涛 李魁 潘昊[2] 丁力 WANG Hongtao;LI Kui;PAN Hao;DING Li

机构地区：[1]武汉理工大学网络信息中心,湖北武汉430070 [2]武汉理工大学计算机学院,湖北武汉430070 [3]武汉尚赛光电科技有限公司,湖北武汉430206

出　　处：《武汉理工大学学报（信息与管理工程版）》2021年第6期574-579,共6页Journal of Wuhan University of Technology：Information & Management Engineering

基　　金：国家自然科学基金项目(21376185).

摘　　要：针对复杂自然场景中文本识别精确度不高的问题,提出了采用STN的文本修正与改进CRNN的文本识别方法,并设计了Res_CRNN文本识别模型。通过采取性能更好的残差网络Resnet50提取更为复杂的序列信息,运用新设计的CTC损失函数进行训练。实验结果表明,将STN文本修正网络与Res_CRNN模型相结合,即可形成了一个端到端的文本修正与识别模式,在该模式下能有效地提高文本识别的精度与效率。For the problem of low accuracy of text recognition in complex natural scenes,the text correction method of STN and the improved CRNN text recognition method are adopted,and the Res_CRNN text recognition model is proposed on the basis of it.By adopting the better-performing residual network Resnet50 to extract more complex sequence information,the newly designed CTC loss function is used for training.The experimental results show that by combining the STN text correction network with the Res_CRNN model,an end-to-end text correction and recognition mode is formed,which improves the accuracy and efficiency of text recognition.

关 键 词：自然场景 英文文本识别 连接时序分类 空间转换网络 卷积递归神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]