Research On Coding Detection And Recognition System Based On Deep Learning

With the improvement of people’s living quality and the development of society,the problem of food safety has been widely concerned by people.As an important reference standard for food safety,the date of production has become a key object for people to check when buying food.Jet printing is a widely used production date marking method in food and drug packaging enterprises.However,due to the instability of the industrial inkjet machine,the wrong printing of the production date will bring economic losses to the enterprise,so the quality inspection has become an essential link in the industrial production.Nowadays,quality inspection is carried out in an artificial way.With the acceleration of production speed,an automatic quality inspection method is urgently needed in industrial production to match the current production speed.Therefore,the research on deep-learning-based inkjet detection and recognition algorithm is extremely critical for the actual industrial production.This topic is oriented by actual enterprise needs and aims at practical application.The research work and contributions of this paper are as follows:(1)Aiming at the character printing environment in industrial field,a set of inkjet image acquisition and detection system is built based on PC platform for the collection of image data sets,and the improved algorithm is tested in the field printing environment to verify the practicability of the system,finally equipped with pneumatic sorting device to achieve the elimination of unqualified products.(2)Research on positioning and detection of food packaging box inkjet character region based on YOLOv5 s.Firstly,SE attention module is introduced to improve the feature extraction capability of the network.Secondly,it is added to the rectangular box regression loss function of the network to solve the problem of unbalance of difficult and easy samples and improve the efficiency and accuracy of positioning.Finally,Angle information is introduced to solve the problem of inaccurate positioning of slanting coding,so that YOLOv5 s network can return a closer prediction box.The improved network positioning accuracy reaches 99.8%.(3)Research on the recognition of inkjet characters.Firstly,the input image is scaled to a fixed size for feature extraction through convolutional neural network,and the extracted character sequence features are sent into the bidirectional LSTM to learn the context information of the characters.In addition,in order to solve the problem of difficult recognition of tilted coded characters in positioning,The Spatial Transformer Networks model is introduced to correct slanted images,and the accuracy of character recognition model can reach 99.5% for clean background and 98.2% for complex background.(4)It is difficult to identify complex food packaging box background by drowning injet characters.The data enhancement method of ink-jet fusion is adopted to generate inkjet characters with complex background in batches by random combination of clean characters,and to make targeted data sets for ink-jet quality inspection in different application scenarios.Meanwhile,the multi-task-based positioning algorithm is adopted to realize the task of positioning the ink-jet character region and to separate the ink-jet character from the food packaging background.Remove background interference for character recognition tasks.Finally,combined with the character recognition model,the accuracy of complex background characters reached 99.7%.