Research And Development Of Online Intelligent Identification Technology For Spraying Code In Cans

In view of the current situation of low manual detection efficiency,low visual detection accuracy and poor versatility in spurt code recognition.The topic of this paper is the research and development of online intelligent identification technology for spraying code in cans.It focus on spurt code image preprocessing and image correction technology,spurt code character positioning technology based on YOLOv2,and spurt code recognition technology based on end-to-end deep learning,which can improve the accuracy and generality of spurting code recognition.It has important academic value and practical significance to promote the development of intelligent detection and manufacturing technology.This paper studies the online intelligent recognition system of cans,and summarizes the research status at home and abroad from the classical character recognition method and the character recognition method based on deep learning.The main research content of this paper are as follows:(1)We analyze the reason of the rotating images and the format features of can spray code,and put forward a method to rectifie image by shape matching,while rectifie image by searching the template of the can spray code invariable characteristics.The method was used to test the spray codes of 10 cans,and the results show that the algorithm has the characteristics of strong universality,high stability,high accuracy and fast speed.(2)We design the character location model based on the improved YOLOv2.In view of the problem of the sensitivity of initial cluster center selection for the parameter of candidate box determined by kmeans,we design the method of clustering candidate box by kmeans++,which get the optimal size and number of candidate boxes and improve the location accuracy of spurting code characters.In view of the problem of the darknet-19 network of YOLOv2 has too many parameters and too much computation,we analyze the number of parameters and computations in the depth separable convolution and standard convolution operations.We decide to use Mobile Net V2 convolution structure to replace Darknet-19,which can improve the speed of character positioning of model spurt code.(3)We analyze the recognition mechanism of end-to-end spurting code based on CRNN.In view of the problem of the network model is easy to be disturbed by background stains and stains of spurting code characters,we add an attentional mechanism to the LSTM layerof the original network model,which can improve the ability to extract key feature information of the network model and ignore the interference information such as character background stains and stains,and improve the accuracy of code spurting recognition of the model.(4)In view of the problem of Alex Net network can not recognize the whole string and feature extraction ability is weak,we change the final full connection layer of the network to eight parallel full connection layers for end-to-end recognition and apply multi-scale convolution feature fusion to the first convolutional layer to improve the capability of network model feature extraction,which can improve the recognition accuracy of model spurt code.The online intelligent identification system of cans spraying code studied in this paper has been applied in relevant enterprises.The application effect shows that the system meets the needs of enterprises and proves that the research content of this paper is scientific and effective.