【作者】 张佳；

【导师】 肖昌炎； 李红梁；

【作者基本信息】 湖南大学 ， 控制工程（专业学位）， 2023， 硕士

【摘要】 金属表面冲压标号,是一种凹陷字符,由于其不易磨损、被修改、成本低等优点被广泛用于记录金属零件的重要信息。早期的标号主要是依靠人工来读取和记录,这大大降低了生产线的工作效率和生产线自动化水平。在生产过程中,由于金属表面冲压受到金属表面反光、金属生锈、标号区域不确定、字符串长度不确定、字符与背景颜色相近导致对比度小等问题影响,使得金属表面冲压标号的识别难度加大。再加上凹陷字符容易存在曝光不均匀、投影失真等问题影响,传统的二维平面图像字符识别算法又存在一定的局限性。本文从满足在线实时检测、降低硬件成本和简化系统等需求出发,通过深度学习和机器视觉技术,来解决工业场景下的金属零配件表面冲压凹陷字符的识别问题,本文主要的工作内容如下:根据实验需求,首先设计了一套用于成像系统,通过光源、相机、镜头等设备设计了一套金属零件图像采集系统,并在现场完成搭建。采取智能控制的方式,触发相机启动拍摄,图像采集完成后传送到工业电脑中进行储存,整个硬件系统保持稳定运行。针对金属表面字符背景相似、光照受限等问题,提出使用基于渐进式可扩展网络对金属零件表面的冲压字符进行定位,PSENet网络通过主干网络Res Net提取特征,通过FPN实现特征融合,再结合尺寸扩展算法,来获得最佳文本包围框。通过与U-Net网络模型进行对比实验,PSENet的网络模型的训练速度更快,定位的结果表现更好。针对冲压标号的识别问题,我们采用卷积循环神经网络模型,它把卷积层、网络层和转录层统一到CRNN模型中,实现了对字符的准确识别。为了解决模糊、低亮度、反光等图像质量不高图像的字符识别率低问题,我们引入了残差网络,使用Res Net-50来替换模型中的卷积网络。通过比较实验结果,改进后的网络模型的识别率有明显提高。文本通过在不同的金属零件图像上进行实验,均能实现对金属表面冲压标号的智能识别,识别率均保持在99%以上。系统能根据字符识别结果,实现零件配件的自动分拣,并实时调用相应的自动焊接程序。本文对金属表面冲压凹陷标号算法进行了研究,通过实验,实现了冲压标号的自动智能化识别。目前,该标号识别系统已在企业生产现场部署,经过一段时间的运行,系统稳定,取得实际效果,满足企业要求。更多还原

【Abstract】 Metal surface stamping labels are a kind of concave characters,they are widely used to record important information of metal parts due to its advantages such as low wear,modification,and low cost.In the early days,labels were mainly read and recorded by manual,which greatly reduced the efficiency and automation of production lines.During the production process,they are more difficult to extract text characters from the metal surface images,because metal surface stamping labels are affected by some problem such as metal surface reflection,metal rust,label-area uncertainty,string length uncertainty,small contrast caused by similar characters and background.In addition,concave characters are prone to uneven exposure,character projection distortion and other problems,the traditional two-dimensional planar image character recognition algorithms are not applicable.In this paper,to satisfy the requirements of online real-time detection,hardware cost reduction and system simplification,we adopt deep learning and machine vision technology to solve the identification the dented characters on the surface of metal parts in industrial scenes.The main study contents of this paper are followed:Firstly,we design a set of imaging system according to the experimental requirements.A metal part image acquisition system was designed using equipment such as a light source,camera,and lens,and was completed on-site.By adopting intelligent control,the camera is triggered to start shooting,and the image is collected and transmitted to an industrial computer for storage.The entire hardware system maintains stable operation.Secondly,to address the issues of similar background and limited lighting of metal surface characters,we propose using progressive scale expansion network to locate stamping characters on the surface of metal parts.The PSENet network extracts features through the backbone network Res Net and achieves feature fusion through FPN.And obtain the optimal text bounding box with the scale expansion algorithms.By conducting comparative experiments with U-Net network models,the training speed of PSENet network model is faster,the positioning performance is better.Thirdly,we adopt a convolutional recurrent neural network to recognize the stamping labels.The CRNN network unify the convolutional layer,network layer,and transcription layer in the the model to recognize the characters correctly.In order to solve the problem of low character recognition rate in images with low image quality such as blurring,low brightness,and reflection,we introduced a residual network and used Res Net-50 to replace the convolutional network in the model.In order to solve the problem of character recognition in low image quality images such as blur,low brightness,reflection,we propose adding a residual network and replacing the convolutional network with Res Net-50.The recognition rate of the improved network model has significantly improved according to the result of comparative experiment.By conducting experiments on different metal part images,the metal surface stamping labels can be read intelligently with the recognition rates above 99%.The system can automatically sort metal parts based on character recognition results and call corresponding automatic welding programs in real-time.In this paper,We have analyzed and studied the algorithm for metal surface stamping labels.Through the experiments,we realize the automatic intelligent recognition of stamping labels.At present,the label recognition system has been deployed on the production site of the enterprise.After a period of operation,the system play very well,achieved practical results,which is meeting the requirements of the enterprise.更多还原