Research On Power Worksheet Key Information Recognition Algorithm Based On Deep Learning

The system of power worksheets is the significant reference for the staff to specify and implement the maintenance strategy,which can ensure the safety of the operation of equipments and prevent accidents.Therefore,the worksheets system is of great significance to maintain the stability of power system.Aiming at the recognition of key information like date,name and seal in power worksheets,this thesis performs studies from three aspects,namely the text detection,the text recognition and the seal recognition.The main work is focused on the issue of detecting the location of the seals and the dates which are mixed with printing and handwritten fonts and the issue of recognizing the dates serially without segmentation,which are itemized as follows:1.To tackle the issue of detecting the location of the seals and the dates which are mixed with printing and handwritten fonts,this thesis compares and verifies three detection algorithms based on deep learning which are itemized as follows: the algorithm of detection the mixed dates based on CTPN,the algorithm of detecting the mixed dates and the seals based on Faster R-CNN and the algorithm of detecting the mixed dates and the seals based on YOLOv4.Then we optimize YOLOv4 from the three aspects,namely the backbone,the width and the size of preset anchors and detection head.The corresponding experimental results imply that the m AP of the improved YOLOv4 is 96.9% over the task of detecting the location of the seals and the dates which are mixed with printing and handwritten fonts,at the average running speed of 152 FPS,ensuring the recognition accuracy and the operating efficency.2.In a gesture to recognizing the dates which are mixed with printing and handwritten fonts in worksheets,this thesis studies the algorithm of single character recognition and the algorithm of sequence recognition respectively.Firstly,this thesis studies the recognition method of the traditional image segmentation combined with classification.Specifically,this thesis designed a combined segmentation algorithm by combining connected domain algorithm with dripping algorithm and introduced the classification algorithm based on CNN lately.The experimental results show that the recognition accuracy of the combined segmentation method is 41.5%,which can segment adhesive characters to a certain extent.Then,this thesis proposes an end-to-end full convolution algorithm based on CRNN for image-based sequence recognition tasks,which can predict the recognition results of the whole text through a single inferencing.In order to improve the feature extraction ability of the model,this thesis introduces a series of feature enhancement modules.Finally,two text recognition models,namely Mobiledate-Net and Mobilename-Net,are built for recognizing person names and dates respectively.The experimental results show that the accuracy of Mobiledate-Net on 200 testing images of mixed date is 96.3% at the speed of 156 FPS.The accuracy of Mobiledate-Net on 50000 testing images of printed name is 92.5% at the speed of156 FPS.The accuracy of Mobilename-Net on 3500 testing images of printed date test sets is99.7% at the speed of 143 FPS.3.Models used for seal recognition are all based on lightweight CNNs,namely Res Net18,Shuffle Net v2 and Mobile Net v3,which are subject to analyzing and comparing.Final results show that the Mobile Net V3 achieves 90.1% in recognition accuracy at operating speed of 167 FPS,demonstrating the best performance.