摘要
为获取污水处理设备铭牌图像中设备编号、名称等基本参数信息,针对现有的文本检测算法存在的漏检率高、准确度低等问题,基于传统的CTPN算法,增加图像校正模块,采用Res Net50网络进行特征提取,提高特征提取效果;识别阶段采用Dense Net+BLSTM+CTC算法模型实现对铭牌图像中不定长字符序列的识别,改进传统的NMS算法,将检测区域进行融合,减少文本框边界丢失导致文本识别错误的问题。在铭牌数据集上进行测试,检测精确率达到95.8%,识别准确率为92.6%,与传统的算法相比,性能有所提升,验证了该模型的有效性。
To obtain the basic parameter information such as equipment number and name in the nameplate image of sewage treatment equipment,aiming at the problems of high missed detection rate and low accuracy for the existing text detection algorithms,an image correction module was added based on the traditional CTPN algorithm,and ResNet50 network was used for feature extraction to improve to the effect of feature extraction.In the recognition stage,DenseNct+BLSTM+CTC algorithm model was used to recognize the indefinite length character sequence in the nameplate image,the traditional NMS algorithm was improved,the detection area was fused,and the problem of text recognition error caused by the loss of text box boundary was reduced.Tested on the nameplate dataset,the detection accuracy is 95.8%and the recognition accuracy is 92.6%.Compared with the traditional algorithm,the performance is improved and the effectiveness of the model is verified.
作者
郭毛琴
谢红薇
张效良
GUO Mao-qin;XIE Hong-wei;ZHANG Xiao-liang(Department of Software,Taiyuan University of Technology,Taiyuan 030024,China)
出处
《计算机工程与设计》
北大核心
2022年第10期2904-2910,共7页
Computer Engineering and Design
基金
国家自然科学基金项目(61872262)
山西省自然科学基金项目(201801D121143)。
关键词
铭牌识别
文本检测
文本识别
深度学习
自然场景
nameplate recognition
text detection
text recognition
deep learning
natural scene