Study On Wagon Character Recognition System Based On Computer Vision

Mine wagon is the main vehicle for coal transportation in China.The current trend requires developing an automated and intelligent way for loading and unloading goods in mine wagon due to the great demand for transportation.But nowadays,most recording work is done by workers based on observing information of China's railway wagon transportation though the camera,which is error-prone and time-consuming.Consequently,it is greatly desired to develop a wagon character recognition system based on computer vision,which can increase the automation and the productivity effectively.In this paper,to identify mining wagon character,an attempt is made to employ deep learning to extract text information.And this research is carried out from the three aspects as follows:(1)Prior to information record,wagon character recognition is a must-have stage.Connectionist text proposal network is employed to recognize wagon text.To achieve a better performance,the model trained on the public dataset is made corresponding adaptation according to the wagon character dataset.The detection algorithm achieves an F1-score of 0.9107.(2)Wagon character recognition is a post-stage.To solve the problem caused by noise interference,ambiguity,and defects in gondola characters,this paper proposes a detection model based on generative adversarial networks(Defect-Restore Generative Adversarial Networks,DR GAN).This model is robust to noise due to the restriction of the coding space.In addition,we compare the performance of four algorithms including CRNN,CRNN with Attention,Attention-based OCR and DR GAN.Experimental results show that the accuracy of the proposed algorithm is significantly higher than that of the other three models in the case of random masks and noise interference.In addition,DR GAN achieves the best accuracy of 97.76% on the undisturbed data set among the four algorithms.(3)In this research,we design a software for wagon character recognition system.The software provides functionalities mainly including account login,automatic identification and results save grab.The thesis contains 29 figures,10 tables and 68 references.