摘要
本文以弯曲超长的钢卷标号为研究对象,提出一种调整字符区域方位的预处理方法和钢卷标号识别的方案。首先使用改进U-Net模型分割字符区域并将文本区域调整至图像上方,然后使用改进Mask TextSpotter V3进行标号识别。实验结果表明,相较于原模型,改进的U-Net模型在反光等干扰的情况下提升了字符区域分割效果,交并比提升了14%;改进的Mask TextSpotter V3模型改善了原模型易受噪声影响、模糊图像文本检测准确率低的问题,在整体长字符检测与识别中,模糊图像的文本检测召回率提高了4%,字符错误率降低了12.49%。整体模型框架针对变长弯曲的多方向长字符在真实场景中准确率能够稳定在98%及以上,在对钢卷标号或类似对象的识别上有一定的先进性、实用性和参考性。
In this paper,a preprocessing method to adjust the orientation of character region and a scheme to recognize the steel coil label are proposed,which is based on the research object of the steel coil label with ultra long bending.First,the improved U-Net model is used to segment the character area and adjust the text area above the image,and then the improved Mask TextSpotter V3 is used for label recognition.The experimental results show that,compared with the original model,the improved U-Net model enhances the character region segmentation effect,and the intersection over union ratio is increased by 14%;the improved Mask TextSpotter V3 model eliminates the vulnerability of the original model to noise and the low accuracy of text detection in fuzzy images.In the overall long character detection and recognition,the text detection recall of fuzzy images is increased by 4%,and the character error rate is reduced by 12.49%.The overall model framework can stably maintain the accuracy of 98%and above in the real scene for the multi-directional long characters with variable length and bending.It is progressive,practical and referential in the recognition of steel coil labels or similar objects.
作者
光鼎立
郑天意
王健
肖昌炎
GUANG Dingli;ZHENG Tianyi;WANG Jian;XIAO Changyan(Liuzhou Iron&Steel Company,Liuzhou 545001,China;China Telecom in Hunan Branch,Changsha 410001,China;Electrical and Information Engineering College of Hunan University,Changsha 410082,China)
作者简介
光鼎立(1984-),男,广西贵港人,学士,主要从事检斤计量的智能化和无人化实施,主要研究方向为深度学习与成像技术。E-mail:guang0618@126.com;郑天意(1998-),女,河南商丘人,硕士,主要研究方向为机器视觉和人工智能。E-mail:zhengty@hnu.edu.cn;王健(1999-),男,硕士研究生,主要研究方向为目标检测、场景文本识别等。E-mail:wj2021@hnu.edu.cn;肖昌炎(1972-),男,湖南湘潭人,教授,博士生导师,入选江苏省企业领军人才计划,主要研究方向为多模态计算成像、工业与手术机器人导航、嵌入式视觉系统、AI目标识别与缺陷检测等。E-mail:c.xiao@hnu.edu.cn
