Research On Character Recognition Algorithm And Integrated Implementation Of Cylindrical Utility Pole Signs Based On Lightweight Network

As an important reference mark for power supply department in power grid management,the utility pole signs record the information of the utility pole number,the line information and the power supply company information and is an important basis for equipment risk investigation and positioning and transmission line optimization.Accurate and efficient extraction of utility pole signs is an important basis for power infrastructure information collection and subsequent data analysis;Automated,lightweight,and intelligent collection methods based on mobile intelligent terminals have become an important means to improve the efficiency of power grid operation and management,and are of great significance for reducing power grid operation and management costs and promoting the intelligent construction of the national grid.At present,the difficulty of character recognition mainly lies in the missing and misrecognition of edge characters of cylindrical utility pole signs,and the accuracy of character recognition is poor.The deep learning model has a large number of parameters and computation,which cannot meet the requirements of lightweight and real-time operation of mobile intelligent terminals.In this thesis,geometric image correction and deep neural network are used to study and implement a lightweight text detection and character recognition model for cylindrical utility pole signs,and a prototype system is built to complete the actual scene experiment test.The main research results are as follows:(1)A special recognition data set is constructed and a cylindrical image correction preprocessing method based on back projection is designed.The data set of real cylindrical utility pole signs is constructed and annotated by three methods: field shooting,network crawling and authoritative publishing.Aiming at the problem of pixel compression on both sides of the edge of the sign plate of cylindrical utility pole signs,projection modeling is carried out in the horizontal and vertical directions,and the projection correction equation is designed to convert the cylindrical image into a flat image by pixel.After the correction,the correct rate of character recognition of utility pole signs is increased by 40.3% compared with direct recognition.(2)A segmentation-based lightweight utility pole signs text detection model Tiny-DBNet is researched and implemented.The model uses deep separable convolutions and attention mechanism to build a lightweight feature extraction network.After the fusion of features at different levels,it is connected to the differentiable binarization module of DBNet to obtain adaptive thresholds in a learnable way,perform text and non-text segmentation.Experimental verification is carried out using real cylindrical pole signs data and open dataset.The detection speed of Tiny-DBNet is improved by 3 times and the parameter scale is reduced by 45.15%under the small loss of 0.60% precision which meet the needs of rapid detection of lightweight handheld terminals(3)The character recognition model based on CRNN is optimized and improved.It studies the optimal combination scheme of feature extraction network,text detection network and character recognition network,and uses the basic network of Tiny-DBNet combined with CRNN algorithm to build a lightweight Tiny-DBnet-CRNN fusion model for utility pole signs recognition which changes the scattered and low coupling structure of modules and reduces training levels and redundant operations.It takes 1s to recognize a sign of a utility pole in single and batch processing modes,and the recognition accuracy rate reaches more than 95%,which provides core algorithm support for system module integration and handheld image acquisition application scenarios.(4)A character recognition prototype system for utility pole signs is builded and based on this,the scene verification analysis is carried out.Four modes of "field collection mode,background processing mode,single processing mode and batch processing mode" and four layers of architecture of "data layer-algorithm layer-integration layer-display layer" are designed for the prototype system of information collection of utility pole signs and deploy in the form of "H5mini-programs".50 cylindrical and flat utility pole signs are verified under the streets,and all 48 signs are output with data and text characters,forming a utility pole signs integration system with strong scene application universality,high algorithm recognition accuracy,and fast character recognition speed.To sum up,in view of the cylindrical pole sign character recognition,this thesis designs the including cylinder correction,text detection,character recognition algorithm,the reliability of the model is verified by experiment,through the actual scene validation and application,effectively increase the pole plate collection operation efficiency and accuracy,reduce the power grid operation and management cost,promote the construction of the national grid.