Research On Text Recognition Of Substation Equipment Nameplate Based On Deep Learning

The power equipment ledgers reflect basicallly the status,ownership and distribution of various types of equipments in power system.A large part of the ledger data come from equipment nameplates.The quality of the equipment ledgers has a great influence on the equipment management in power system.However,there are still some problems in the equipment ledgers of China’s power system,such as the missing and low accuracy of ledger data.Traditionally,the nameplate data are logged manually in the equipment ledgers and the consistency between the nameplates and the ledgers are checked also manually,which has the disadvantages of low efficiency and poor accuracy.Therefore,a deep learning-based image recognition technology is used in this work to extract text information in nameplate pictures.This technology can also be used to record and check equipment ledger data,which improves the work efficiency and data accuracy.The research object of this thesis is the nameplates of substation equipments,and the main work inlcudes the positioning of namplates,the detection and the recognition of texts in nameplates.Firstly,a nameplate positioning model was designed based on the object detection algorithm YOLO(You Only Look Once),and the loss function of the model was improved to solve the problems of unreasonable detection box loss and imbalance of positive and negative samples.In addition,the data augmentation was used to solve the problem of insufficient data,and the best hyperparameters of the model were obtained by means of clustering method and parameter pre-search method.The experimental results show that the improved model can effectively locate the areas of nameplates in the images with complex background.Moreover,the pruning technique was used to compress the trained model.This greatly reduced the model parameters and FLOPs(Floating Point Operations)with only little decrease in accuracy,which improves the speed of nameplate positioning.Secondly,a nameplate text detection model based on EAST(Efficient and Accurate Scene Text Detector)algorithm was designed on the results of nameplate positioning,and a fast NMS(non-maximum suppression)algorithm was proposed to solve the problem that the detection costs too much time due to the large amount of nameplate texts.The results show that the proposed text detection model can effectively detect different types and scales of texts in nameplate images.Moreover,the proposed NMS algorithm can greatly reduce the non-maximal suppression processing time without reducing accuracy,and can effectively solve the problem of incomplete detection of long text.In addition,a new method for calculating the inclination angle of nameplate texts based on perspective transformation was proposed to assist the EAST algorithm.The experiment shows that this method has more advantages in detecting the texts with larger angles,but it is less robust.Thirdly,a nameplate text recognition model based on CRNN(Convolutional Recurrent Neural Network)algorithm was designed on the results of the above text detection.According to the characteristics of the nameplate texts of substation equipments,a large number of text image data were synthesized artificially to solve the problem of small amount of text data and difficult labeling.In the experiment,the text recognition model was first pre-trained using open source text data sets,and then the real nameplate text data and synthetic data were used to fine-tune the pre-training model,which improved the performance of the model.In addition,a dictionary-based text correction method was used to correct the text recognition results,improving the recognition accuracy.Finally,the nameplate positioning module,text detection module and text recognition module were integrated to achieve a complete nameplate text recognition,and the corresponding software system was designed.Furthermore,the nameplate text recognition system was applied to the verification of the equipment ledgers,which can be used to evaluate the accuracy of the ledger data and provid some suggestions for modification.