Research On Container Terminal Trailer Number Recogntion System Based On Semi Supervised Learning

With the continuous development of artificial intelligence technology,methods to solve OCR problems have evolved from traditional image algorithms and machine learning methods to deep learning-based computing methods,which improves the efficiency and accuracy of the whole OCR system development.At the same time,with the deepening of deep learning theory and algorithm research,text detection and recognition methods based on deep learning in OCR become more and more popular and widely used in the construction of fully automated container terminal system.The wide application of port scenarios poses challenges to the efficiency and cost of model development.To ensure the accuracy of the model,labeling large amounts of unlabeled text data collected for each particular scenario is usually required,which requires a lot of time and labor costs.In order to reduce the cost of development,text direction recognition,text recognition and semi-supervised learning are deeply studied in this thesis.Taking container terminal Trailer number recognition task as an example,a semisupervised learning vehicle number detection network is designed and implemented.Through automatic labeling software,a real-time Trailer number recognition system is quickly built.The main work of this thesisis as follows:(1)Four text detection algorithms based on deep learning are studied and compared:EAST,PSENet,DBNet,PAN.Through the accuracy,recall rate,F value and recognition speed four evaluation indexes,the above four algorithms are compared,and finally,a basic text detection model using DBNet as the semi-supervised vehicle number recognition system is proposed.(2)Based on the idea of generating two kinds of ideas to improve the performance of semi-supervised learning system based on consistency regularity and pseudo-labels,this thesis presents a text detection algorithm for semi-supervised learning which combines consistency regularity and pseudo-labels.The detection accuracy is improved by using the Main Teacher DBNet consistency model with Segmix input perturbations.Then,a pseudo tag generation method is proposed,which combines CopyPaste with positive and negative sample generation methods to improve the performance of training models for small batches of data samples.(3)Considering the different directions of container terminal set trucks and the irregularity of text direction detected,the text direction recognition and text recognition networks are studied,and the two networks are overlapped to build the recognition network model of this system.Three typical lightweight image classification networks,MobileNet,ShuffleNet and GhostNet,are studied experimentally.MobileNet V3 is selected as the text direction recognition network of the vehicle number recognition system.The text recognition networks CRNN,SAR,SRN are further studied through experimental methods.SRN is selected as the text recognition network of vehicle number recognition system.Based on these two basic network models,a suitable data enhancement method for text recognition task is proposed,which effectively improves the recognition performance of the model.(4)The labelme-based auto-labeling tool OCRlabelme was developed,and pseudolabeling was provided using semi-supervised vehicle number detection model to improve the efficiency of building a vehicle number recognition system.Three methods of deploying a vehicle number recognition system were experimented.The performance of the overall vehicle number recognition system deployed by Tensor RT reached the practical application requirements.