Research On Image Recognition Method Of Instrument Representation Based On Deep Learning

As an important tool for data display and monitoring,meters are widely used in industrial fields,especially in high-pressure,high-temperature,and high-radiation environments such as substations.There are a large number of pointer meters and digital display meters in substations.The number of manual readers has shortcomings such as high work intensity,high cost,low efficiency,and large errors.There is an urgent need to study an intelligent and unattended method for identifying numbers..Instrument image processing has problems such as instrument classification,difficulty in detecting dial and pointer areas,and cumbersome instrument image preprocessing.Traditional instrument recognition methods cannot solve these problems.This article is based on deep learning target detection methods for instrument representation image recognition.The main research contents are as follows :Research on the Faster R-CNN-based dashboard positioning method.Aiming at the small proportion of the dashboard and pointer areas in the meter image,this paper proposes an improved Faster R-CNN deep learning target detection method,which combines the regions in Faster R-CNN It is recommended that the network(Region Proposal Networks,RPN)and the Feature Pyramid Networks(FPN)network be integrated to improve the accuracy of the positioning of the dashboard and pointer area.Aiming at the problem of balancing the positive and negative samples of the instrument image,the Focal Loss sample balance loss function is introduced into the RPN network for sample training.The experimental results show that the missed detection rate of the improved method is increased by 8.9% compared with the traditional algorithm,and the false detection rate is increased by 5.1% compared with the traditional algorithm.Research on the recognition method of pointer representation number based on feature pyramid(FPN).In order to improve the accuracy of pointer region detection,this paper proposes an improved FPN image segmentation method,which adds a deconvolution layer on the top layer of FPN and adds a deconvolution layer to the convolution feature map.Multiplexing can achieve accurate and efficient segmentation of the pointer area.In addition,for the problem of instrument image tilt,perspective transformation is used to calibrate the image.Finally,in order to improve the recognition accuracy,the least squares method is used to fit the pointer area,calculate the deflection angle of the pointer,and obtain the indication number of the pointer type instrument.Experiments show that the accuracy rate of the instrument representation number recognition method based on the improved FPN method is increased by 2.38%,the recall rate is increased by 5.96%,and the total time is increased by 21 ms than that based on the Hough transform.Research is based on the Connectionist Text Proposal Network(CTPN)digital display instrument representation number recognition method.Traditional character detection methods are susceptible to the effects of image preprocessing and character segmentation.This paper locates the character area based on CTPN,and then recognizes the number through a densely connected network(Dense Net).This algorithm avoids image preprocessing and character segmentation and simplifies operations.Identification process.Experiments show that the algorithm in this paper improves the accuracy and reliability of the number recognition of the digital display instrument.The research content of this article is applied to the image recognition of the instrument display,and realizes the positioning of the instrument panel,the identification of the pointer type instrument and the identification of the digital display instrument.Experimental results show that the algorithm has higher reading accuracy,stronger universality and generalization ability than traditional algorithms.