| Pointer instruments have the advantages of anti-vibration,anti-high temperature,anti-magnetic interference,etc.,so they are still widely used in substations.Relying on manual meter reading is not only costly and risky,but also laborious and time-sensitive.Therefore,it is of great significance to realize the automatic identification of the readings of the substation pointer meter.The traditional instrument reading identification process is: instrument target detection,instrument panel noise reduction,instrument distortion correction,instrument scale detection,instrument pointer segmentation,and reading calculation.Obviously,multi-process nodes and lack of instrument level correction will reduce the recognition accuracy,and there is no instrument category and range recognition,which weakens the robustness of the algorithm.Aiming at the above problems,the subject conducts research from three aspects: meter target detection,meter automatic level correction and meter reading interpretation,thus designing an end-to-end pointer meter reading recognition framework based on deep learning.Aiming at the problems of weak semantic expression ability and low target recognition rate of traditional target detection algorithm,this paper designs a target detection algorithm for substation instrumentation based on Res Net50 and improved UNet network.The instrument features are extracted with Res Net50 as the skeleton network.The UNet network fusion feature is optimized by Amplification Pooling（APPooling）,and the fusion feature detection instrument is used based on the regression idea of the SSD target detection algorithm.In addition,in order to obtain range information,category information and assist subsequent instrument correction,a shared convolution multiplexing feature map is introduced,that is,the output feature processed by UNet is used as the first shared convolution.The main reason for the long identification process and large error of pointer meter readings is that it is difficult to design a meter reading identification framework that integrates the target detection process and the reading identification process.The main obstacle to designing this framework is that the data collected by the substation has tilt,rotation and distortion and other characteristics.Aiming at this problem,this paper proposes an automatic level correction algorithm for the meter.The instrument panel text is detected based on the first shared convolution,and based on the recognized text area,the random sampling consistency algorithm is used to fit the vanishing point,the vanishing line is obtained through the vanishing point,and the text is projected according to the relationship between the vanishing line and the projection transformation matrix.Distortion correction,after that,the text geometric correction is realized by two groups of originally vertical straight lines,and the original quadrilateral of the text area is corrected into an axis-parallel rectangle to realize the horizontal correction of the instrument text.Use the text level correction matrix to correct the instrument panel features,and indirectly complete the automatic level correction of the instrument.Finally,the range and category of the instrument are identified based on the instrument panel text.In addition,in order to extract the pointer pointing when the instrument is in the horizontal state,the feature map after horizontal correction is used as the second shared convolution to realize the re-multiplexing of the feature map.Since the principal component analysis algorithm（PCA）can only fit an undirected straight line and cannot identify a directed instrument pointer straight line,this paper adopts the overlapping degree strategy to optimize the PCA and designs the instrument indicator number interpretation method based on the optimized algorithm（DOO-PCA）.The meter pointer is segmented based on quadratic shared convolution,and the vector fitting of the meter pointer is realized by DOO-PCA.Finally,the size and unit of the number are indicated by the pointer vector,the type of the instrument and the range information to determine the instrument.Through the comparative analysis of experiments,the results show that the method proposed in this paper has high target detection rate,recall rate and reading recognition accuracy,and has a certain reference value for the research of pointer meter reading recognition work. |
