摘要
由检修人员填写的机车检修记录单在电子化归档过程中,采用人工筛查分类,存在效率低、耗时长的问题。为此,文章提出一种改进的基于PaddlePaddle-OCR的文本识别方法,其能准确识别检修单中关键文本内容,并依据所识别的信息实现检修单的快速分类归档。该方法首先基于傅里叶变换进行文本倾斜角度检测校正;接着,采用以可微二值化模块为头部的分割网络来实现文本区域的检测;最后,采用卷积循环神经网络完成文本区域的文字识别。实验结果表明,采用该方法后,检修记录单中关键表头信息的识别准确率在0.94以上,单张图像识别耗时小于17 s,且可批量化操作,能满足机车检修单电子化分类归档的精度与效率要求;同时,相比原版PaddlePaddle-OCR,文中所提方法在文本倾斜场景下的识别效果更好。文章最后还探索了主干网络卷积核尺寸比例对文本检测器与文本识别器的影响。
In the process of electronic filling,locomotive maintenance record sheet filled out by maintenance personnel adopts manual screening and classfication,which has the problem of low efficiency and time-consuming.A text recognition method based on PaddlePaddle-OCR is proposed in this paper to detect and recognize key text information on record sheet which makes sheet classifying more easily and faster.Firstly,the inclination angle of text calculated by Fourier transform is used for image rotation.Secondly,segmentation network with the head of differentiable binarization module detects text regions of image.Lastly,text recognition is finished by CRNN network.Experimental results show that the key text information recognition accuracy is more than 0.94 and the recognition time is less than 17 seconds needed for CPU per image by this method.Requirements of accuracy and efficiency are satisfied with batch computation.Compared with origin PaddlePaddle-OCR,the method proposed in this paper performs better in text inclination scene.The influences on text detector and recognizer caused by ratio of convolution kernel in backbones are also researched in this paper.
作者
颜家云
张慧源
李晨
彭联贴
YAN Jiayun;ZHANG Huiyuan;LI Chen;PENG Liantie(CRRC Zhuzhou Institute Co.,Ltd.,Zhuzhou,Hunan 412001,China)
出处
《控制与信息技术》
2021年第6期77-83,共7页
CONTROL AND INFORMATION TECHNOLOGY
关键词
光学字符识别法
卷积循环神经网络
可微二值化
机车检修单
OCR(optical character recognition)
CRNN(convolutional recurrent neural network)
differentiable binarization
locomotive maintenance sheet
