摘要
自然场景图像中文本图像的分辨率通常较低,给文本检测和识别等工作带来困难.为了提高文字区域的清晰度、提升文字识别的精度,提出一种新的场景文本图像超分辨网络,在文本超分辨网络(TSRN)的基础上加入置换注意力机制.通过调节相机焦距拍摄同一场景构建真实中文场景文本数据集(CSTD),它包含成对的真实低分辨率和高分辨率图像.在CSTD数据集上进行训练和测试,用PaddleOCR进行文字识别,低分辨率图像的文字识别率提升了3.8%左右;将该网络与TSRN网络做消融实验,测试集的文字识别率提升2.6%左右,证明了算法的有效性.
Text images in natural scenes often face the problem of low resolution,which brings great difficulties to text detection and text recognition.In order to improve the clarity of text region and the accuracy of character recognition,a new scene text image super resolution network was proposed,which is based on the text super-resolution network(TSRN)and adds the mechanism of shuffle attention.A real Chinese scene text dataset(CSTD)was constructed by adjusting the cameras focal length to shoot the same scene,which contains pairs of real low-resolution and high-resolution images.After training and testing on CSTD dataset,the PaddleOCR tool was used for character recognition,and the character recognition rate of LR image is improved by about 3.8%.In addition,ablation experiments on this network and TSRN network was conducted,and the character recognition rate of the test set increases by about 2.6%.The experimental results prove the effectiveness of the proposed algorithm.
作者
袁龙婷
王标
邹佳运
YUAN Longting;WANG Biao;ZOU Jiayun(College of Automation and Information Engineering,Sichuan University of Science and Engineer,Yibin,Sichuan 643002,China;Chengdu Shiguan Tianxia Technology Co.,Ltd,Chengdu,Sichuan 610095,China)
出处
《宜宾学院学报》
2021年第12期43-47,共5页
Journal of Yibin University
基金
四川省科技厅项目“基于移动单兵技术大规模人像布控识别系统关键技术的研究与实现”(2019YJ047)。
关键词
图像超分
注意力机制
文字识别率
image super-resolution
attention mechanism
character recognition rate
