基于深度学习的工业视觉箱体字符识别与判断  

Character Recognition and Judgment of Industrial Vision Box Based on Deep Learning

作　　者：葛永杰 王丽丹[1,2,3,4] 陈定喜 段书凯 干秀灵[1,3] GE Yongjie;WANG Lidan;CHEN Dingxi;DUAN Shukai;GAN Xiuling(College of Electronic and Information Engineering,Southwest University,Chongqing 400715,China;National and LocalJoint Engineering Laboratory of Intelligent Transmission and Control Technology,Chongqing 400715,China;Chongqing Key Laboratory of Brain-Inspired Computing and Intelligent Control,Chongqing 400715,China;Chongqing Brain Science Collaborative Innovation Center,Chongqing 400715,China;Midea Group,Foshan,Guangdong 528311,China;School of Artificial Intelligence,Southwest University,Chongqing 400715,China)

机构地区：[1]西南大学电子信息工程学院,重庆400715 [2]智能传动和控制技术国家地方联合工程实验室,重庆400715 [3]类脑计算与智能控制重庆市重点实验室,重庆400715 [4]重庆市脑科学协同创新中心,重庆400715 [5]美的集团,广东佛山528311 [6]西南大学人工智能学院,重庆400715

出　　处：《计算机工程》2022年第1期296-304,共9页Computer Engineering

基　　金：国家重点研发计划(2018YFB1306600);国家自然科学基金(62076207,62076208,U20A20227,61672436);重庆市基础科学与前沿技术研究专项重点项目(cstc2017jcyjBX0050)。

摘　　要：工厂生产线上的商品包装外箱文本印刷存在残缺,无法及时检出会影响流通销售。制作工业商品外观信息数据集,提出基于深度学习的工业视觉箱体字符识别与匹配判断方法。合并YOLOv3中的卷积层和批量归一化层,引入GIoU作为边界框损失函数并设计自适应调整定位坐标的方法,优化在原始图像上进行文本检测定位的速度与精度。同时,训练并对比CRNN和Tesseract两种识别引擎在已裁剪文本图片上的识别性能,设计字符匹配方法判断字符识别正确与否并输出结果,从而减少误判。对基于该方法的系统进行生产线实测,实验结果表明,其识别准确率可达99.5%,单件商品的外观拍照、检测识别、输出结果耗时仅3 s左右,表明所提方法能够实现实时监测。If the incomplete text printing on commodity packaging boxes produced by factory production lines cannot be detected in time,the sales and circulation of the commodities will be affected.This paper presents a deep learning-based box character recognition and matching method for industrial vision,and also makes a data set of industrial commodity appearance information for the method.By merging the convolutional layer and the batch normalization layer of YOLOv3,and introducing GIoU as the loss function of the boundary box,a method for adaptive positioning coordinate adjustment is designed,which improves the speed and accuracy of text detection and location on the original image.Then the recognition performance of the trained CRNN and Tesseract engines on cropped text images is compared.The designed character matching method is used to judge whether the character recognition result is correct,and the result is output,which reduces the misjudgment.The system based on this method is tested on a production line,and the experimental results show that the system displays an accuracy of 99.5%.It takes about 3 s to take a photo of the appearance,detect and recognize the characters,and output the result of a single product,which demonstrates that the proposed method enables real-time monitoring.

关 键 词：深度学习 YOLOv3算法 卷积递归神经网络 字符识别 外观信息 实时监测 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]