| Digital meter reading recognition for industrial inspections generally uses manual identification and recording of meter reading information.However,manual recording of readings is an important factor affecting factory production efficiency.When the factory uses manual meter reading,it is inevitable to record wrong readings and take a long time.Therefore,the inspection equipment is used to automatically optimize the recording process,improve production efficiency,and provide accurate protection for recorded readings.This paper analyzes the influencing factors of digital meter reading recognition in industrial inspection,and designs a scheme that can realize real-time recognition of industrial digital meter readings according to the function and performance requirements of industrial inspection.This article uses the inspection terminal camera to image the digital instrument on the industrial site as a data source,and then preprocesses the instrument image to obtain the image format that meets the requirements of the industrial digital instrument reading visual inspection model,and uses the inspection model to determine the meter reading area in the image Perform detection to obtain the regional information of the reading,and then design the visual recognition algorithm for industrial digital meter reading to read the reading.The main research contents of this paper are as follows(1)Based on the characteristics of industrial digital instrument data sets and the requirements for the accuracy of intelligent recognition of digital instrument readings,this paper designs a digital instrument reading visual inspection method based on An Efficient and Accurate Scene Text Detector(EAST).After using the test set to test and verify this article,it is found that the detection method based on EAST cannot solve the large and long readings well.Based on this,this paper designs a digital meter reading visual detection method based on improved EAST.This method divides the reading area so that the edge area pays attention to the vertex of the neighbors,and trains the improved EAST method.Experiments have verified that the improved EAST detection model has significantly improved accuracy in the case of large and long readings.(2)Based on the data characteristics of industrial digital instrument readings with variable length and rich reading colors,and the requirements for the accuracy of digital instrument reading recognition,this paper designs a digital instrument reading visual recognition method for industrial inspection.This method first extracts the features of the readings,then transposes the features and Long Short-Term Memory(LSTM)serialization to obtain the reading feature sequence,and finally uses the Connectionist temporal classification(CTC)to decode the feature sequence to obtain the reading number.Experiments have proved that it meets the functional and performance requirements of the industry(3)Based on the real-time requirements of the intelligent recognition of digital meter readings in the industrial field,this article compresses the industrial digital reading detection and recognition model designed in this article.This paper uses the pruning method to focus on and learn model connections,and delete connections below the weight threshold to achieve model compression.Experiments show that this method can improve the real-time performance of industrial digital reading detection and recognition models.(4)Based on the reflection problem in the meter reading area in the industrial inspection environment,this paper designs an algorithm that uses the focus information of the camera to achieve image reflection suppression,and compares and analyzes the reflection suppression methods studied by previous researchers.Experiments verify that the algorithm can effectively improve the industry's performance requirements for accuracy.After verification,this paper can accurately realize the detection and recognition of indefinite length readings of digital instruments for industrial inspections in real time,and can achieve the expected reading recognition accuracy of 90%and real-time 1 second program design goals on the prototype platform of this paper.Industrial performance requirements. |
PDF Full Text Request