Text Recognition In Images Of Traffic Scenes

With the development of deep learning in recent years,artificial intelligence technology is increasingly applied to traffic scenes.Recognizing text in images is a key link in the application of artificial intelligence technology in traffic scenes,and plays an important role in related fields such as traffic violation punishment and intelligent driving.However,due to the complexity of natural scenes,current text recognition algorithms do not work well in traffic scene images.Therefore,this paper selects two important text objects in traffic scenes,license plate and traffic sign,as the research objects,conducts in-depth research on scene text recognition technology and constructs a complete algorithm for license plate recognition and traffic sign recognition.The main work is as follows:Aiming at the problem of license plate recognition,by analyzing the defects of traditional license plate recognition algorithm,a two-stage license plate recognition algorithm based on deep learning is proposed in this paper,which can be divided into two stages: license plate detection and license plate recognition.In the license plate detection stage,for the target characteristics of license plates,the YOLO network is improved by lightening the network structure and adding attention mechanism in this paper.And in the text recognition stage,for the text characteristics of license plates,the CRNN network is optimized by adding the spatial transformation module and improving the feature extraction network.The experimental results show that the improvements in this paper effectively improve the model performance.On the test set,our algorithm surpasses the traditional algorithm in both recognition accuracy and recognition speed.Aiming at the problem of traffic sign recognition,firstly,according to the different ways of conveying information,the traffic signs are divided into two types: symbolic traffic signs and text traffic signs.In order to solve the problem that existing research pays less attention to text traffic signs,a general traffic sign recognition algorithm is proposed in this paper,which can be divided into three stages: traffic sign detection,text detection,and text recognition.In the traffic sign detection stage,we use YOLOV3 network to locate and classify the traffic signs in the image.In the text detection stage,the DBNet network is improved by expanding the receptive field for the false detection of large area text.In the text recognition stage,in order to recognize the vertical text,a multi-directional text recognition algorithm is constructed by adding a pre-processing step to the CRNN algorithm.For the problem of insufficient training data,we also propose an image synthesis algorithm for text traffic signs in this paper.Finally,due to the lack of standard dataset of text traffic signs,the Traffic Sign＿ch dataset is built for experiments.The experimental results show that our algorithm has achieved ideal recognition results on both types of traffic signs.Especially for text traffic signs,our algorithm surpasses the general scene text recognition algorithm in both text detection and text recognition.Through the targeted improvement strategy,the algorithm designed in this paper has reached ideal performance indicators in the scene of license plate recognition and traffic sign recognition.