Research Of Text Detection And Recognition Methods In The Underground Garage Environment

Automated driving technology has been developing very rapidly in recent years,and computer vision technology plays a key role in this field.The detection and recognition of license plates and parking space numbers is one of the very necessary tasks in autonomous driving.With the rapid rise of deep learning techniques,the field of text detection and recognition has also been rapidly developed.However,compared with ordinary license plate detection,the underground garage environment is often more complex and mostly under low-light environments,which seriously affects the accurate recognition of location information such as parking space number and the need to accurately identify the location information such as parking space number,area number,and floor number.Therefore,accurately parking a car in a parking space and recording its location information is a challenging vision task,and there are still some problems to be solved.In this thesis,we focus on the detection and recognition of text in underground garage environments and aim to design algorithms that meet practical production needs.The main work of this study is as follows:Aiming the problem that the text of location information such as parking space number in the garage environment requires high detection accuracy and the efficiency of the algorithm for practical applications,this thesis proposes a robust parking space number detection algorithm PSND.The algorithm is improved based on the DB algorithm,and by introducing the cascading feature pyramid enhancement module and context attention block,the proposed method has better feature extraction capability as well as better detection of long and sparse text.In addition,a dataset of 9000 text images of underground garage parking space numbers and other locations is collected and labeled for experimental validation.Finally,experiments are conducted on several publicly available scene text detection datasets,and the results show that the proposed algorithm can improve the accuracy of detection and prove the effectiveness of this method.To address the problem that most existing algorithms divide text detection and text recognition into two separate subtasks,which leads to low model efficiency and error accumulation,this thesis proposes an end-to-end parking space number spotting algorithm PSNS.The algorithm is based on PSND using a more efficient self-regularized attention map to replace the context attention block.Also,a lightweight text recognition branch was also added,resulting in an end-to-end framework that can detect and recognize in real-time.The proposed algorithm confirms by conducting experiments that it can maintain real-time speed on the UPSN dataset and also shows better recognition results.In addition,the proposed algorithm also achieves excellent performance on public datasets,which proves the effectiveness of the proposed algorithm.