| 基于改进的CNN喷码式不规则字符识别与提取方法 |
| |
| 引用本文： | 王太学,苗相彬,李柏林,郭彩玲.基于改进的CNN喷码式不规则字符识别与提取方法[J].唐山学院学报,2022,35(3):1-9. |
| |
| 作者姓名： | 王太学 苗相彬 李柏林 郭彩玲 |
| |
| 作者单位： | 西南交通大学 唐山研究生院, 河北 唐山 063000;西南交通大学 机械工程学院, 成都 610031;唐山学院 机电工程学院, 河北 唐山 063000;河北省智能装备数字化设计及过程仿真重点实验室, 河北 唐山 063000 |
| |
| 基金项目： | 国家自然科学基金青年基金(51705436);河北省重点研发科技项目(20327407D) |
| |
| 摘 要： | 乳制品纸包装上的生产批号在喷码过程中由于各种原因部分字符出现粘连或缺失,影响字符的自动化识别。针对这一问题,提出了一种基于改进的CNN喷码式不规则字符识别与提取方法。首先,利用yolov3算法对生产日期区域进行提取;其次,对图像进行预处理;再次,通过一种基于字宽的分割算法结合投影法,利用相邻字符间的像素差异实现对粘连字符的分割;最后,对分割后的单个字符利用改进的CNN进行多标签分类训练得到模型。实验结果表明,改进后的模型对粘连字符和半或残缺字符的识别准确率分别为97.89%和96.71%,相较于模板匹配法、传统的LeNet-5模型、fast R-CNN+NMS模型和yolov3+K-means算法都有所提高。基于该方法设计的字符识别系统,提高了生产日期的在线识别准确率。
|
| 关 键 词： | 喷码字符 目标检测 图像处理 字符识别系统 |
Recognition and Extraction of Irregular Inkjet Characters with Improved CNN |
| |
| WANG Tai-xue,MIAO Xiang-bin,LI Bai-lin,GUO Cai-ling.Recognition and Extraction of Irregular Inkjet Characters with Improved CNN[J].Journal of Tangshan College,2022,35(3):1-9. |
| |
| Authors: | WANG Tai-xue MIAO Xiang-bin LI Bai-lin GUO Cai-ling |
| |
| Institution: | Graduate School of Tangshan, Southwest Jiaotong University, Tangshan 063000, China;School of Mechanical Engineering, Southwest Jiaotong University, Chengdu 610031, China; School of Electrical and Mechanical Engineering, Tangshan University, Tangshan 063000, China;Hebei Key Lab of Intelligent Equipment Digital Design and Process Simulation, Tangshan University, Tangshan 063000, China |
| |
| Abstract: | Due to improper operation, some of the production batch numbers on the paper packaging of the dairy are adhered or missing during the spraying process, which affects the automatic character recognition. To solve this problem, a character recognition and extraction method based on improved CNN is proposed. First, the yolov3 model is used to extract the production date area;Then, the image is preprocessed; After that, a segmentation algorithm based on word width is proposed together with projection method,which uses the pixel difference between adjacent characters to segment the adhered characters; In the last step, the single character after segmentation is trained with more labels classification by the improved CNN to get the model. The experimental results show that the recognition accuracy of the improved model for the adhered characters and the incomplete characters are 97.89% and 96.71% respectively. Compared with the template matching method, the traditional LeNet-5 model, fast R-CNN+NMS model and yolov3+K-means model, the above figures have improved. This character recognition system has improved the online recognition accuracy of the production date. |
| |
| Keywords: | inkjet character target detection image processing character recognition system |
|
| 点击此处可从《唐山学院学报》浏览原始摘要信息 |
| 点击此处可从《唐山学院学报》下载免费的PDF全文 |
|
