Research On Recognition Of Electric Energy Meter Rotation Display Number In Complex Environment

With the rapid development of the national power industry,the intelligentization of the power grid has brought great convenience to both power companies and individual users,and devices such as smart meters are gradually integrated into our daily lives.Therefore,in order to ensure that the energy meter can work normally and avoid the loss of economic benefits to enterprises or individuals due to errors caused by old equipment,the power grid company will rotate the old electric energy meters that have reached the age in accordance with relevant national regulations.During the rotation process,the staff needs to use mobile devices to take pictures of the meters to be replaced and upload them for retention,and then the backstage staff will collect and input the information such as meter number,asset number and electric energy meter display into the system.Among them,the display information of the old meter needs to be inputed manually by means of eye observation.However,because the electric energy meter is often installed outdoors,its position is remote and the surrounding environment is complex,the electric energy meter images taken will have problems such as tilt,reflection,exposure,shadow and surface stain.In view of the above problems and combined with the rapid development and maturation of the field of deep learning text detection and recognition algorithms in recent years,this thesis treat the task of recognize the number of electric energy meters as one of the natural scene text recognition,by studying and improving the related text detection and recognition algorithms,we propose a electric energy meter display area detection and display number recognition algorithm,and develop an interface for electric energy meter display number recognition system.The work of this thesis is as follows :（1）In this thesis,the DBNet text detection algorithm is improved.By lightweighting the backbone feature extraction network,the complexity of the network is reduced,and the detection speed is improved.The spatial and channel attention mechanism module is introduced to make the network focus on more important features with more information,and suppress the features with small weight.The Atrous Spatial Pyramid Pooling（ASPP）module that integrates multiple expansion convolution kernels is used to extract the feature map more semantic information and enhance the segmentation effect of the display region.In the feature fusion stage,the bidirectional feature fusion network with weight coefficient is used to fuse the feature map information of different scales to improve the detection effect of the display area.Through the model training and experimental comparison,it shows that our improved algorithm can better detect the electric energy meter display area in complex environment.（2）In order to solve the problem of tilted and distorted display numbers,and the dataset is not large enough,this thesis corrects and enhances the dataset image.Then,based on the improvement of CRNN text recognition algorithm,a lightweight Feature-Fusion CRNN algorithm is proposed.The feature extraction network is lightweight and the multi-scale feature maps are fused,which reduces the problem of information leakage caused by single feature map prediction,and then introduced the SPP module and improved its network structure,which can further enrich the information expression ability of feature maps.The experimental results on the electric energy meter dataset show that the proposed Feature-Fusion CRNN algorithm can quickly and accurately identify the numbers in the display area.（3）Based on the above display area detection and display number recognition algorithm,this thesis designs and develops an interface of the electric energy meter display number recognition system based on C# language and SQL Server database,which realizes the automatic identification of the electric energy meter image display number,the visualization of the results and the storage to the database,verifies the effectiveness and feasibility of the research work of this thesis.