Research On Text Detection And Recognition Methods In Power Nameplate Images

This thesis takes the intelligent power-station nameplate information collection system as the background,and conducts research on algorithms and system implementation.On the one hand,the key issues such as text detection,text recognition and etched character recognition on power nameplate images are studied.On the other hand,a power-station nameplate information collection system including an Android client and a server is built.The main work and constributions are itemized as follows:1.A text detection algorithm based on YOLOv3 is proposed which detects text parts with YOLOv3 first,then connects text parts into text lines,finally detects the deflection angle of the text line and corrects it.The algorithm has a correct rate of 95.59% with a recall rate of99.3% and the algorithm execution speed reached 20 frames per second.The detection accuracy and real-time performance meet the requirements of the system.2 The power-station nameplate text recognition task is completed with CRNN.The model is trained with transfer learning and ACNET.Transfer learning is employed to solve the problem of the lack of sufficient training data.Using ACNET asymmetric convolution blocks instead of square convolution layers during model training can further improve Model accuracy.Experiments show that training models using transfer learning and ACEnet perform better.The accuracy of the nameplate text recognition can reach 91.09%,and the detection time is controlled below 0.02 s.3 To solve the problem that the accuracy of text recognition of etched characters on power nameplates is low,a method of image augmentation for etched characters is proposed.First,the TET-GAN is used to generate the sculpted text image,and then the Meta Net is used to stylize the background of the text image.The generated text image is used to fine-tune the text recognition model.Experiments show that the accuracy is improved by 13.48% compared with the model before fine-tuning.4.A power nameplate information collection system was set up.The power nameplate information collection system is divided into two parts: Android client on a mobile phone and server.Android client simply pre-processes the image after capturing the image,and then transmits image to the server.Text detection and text recognition models are deployed on the server,and Tensor RT is used to optimize the model inference process.After receiving image sent from Android client,server performs text detection and text recognize.Finally,keyword analysis is performed on all the recognized texts to obtain key information.