Research On Detection And Reading Method Of Pointer Meter Based On Improved YOLOX

Due to its low cost,reliable structure,anti-electromagnetic interference and other characteristics,industrial pointer meters are widely used in aerospace,factory manufacturing,largescale industrial platforms and other fields as an important measurement tool for data acquisition,which is very important to ensure normal industrial production.However,a large number of pointer meters still require manual reading,which has the problems of low efficiency,poor real-time performance,and high error,and some pointer meters have a harsh working environment,making it difficult to manually collect data,especially in some high temperature,high pressure,high corrosion and other environments,safety accidents occur frequently,which cannot meet the needs of modern industrial production and development.Therefore,there is an urgent need to design an automated and intelligent pointer meter reading method for efficient and accurate data acquisition.At present,the mainstream pointer meter reading methods use equipment such as surveillance cameras or inspection robots to collect pointer meter images.However,in the actual complex working environment,there are difficulties in reading caused by factors such as dim light,tilted dials,and blurred images.In view of the above problems,this thesis combines image processing and deep learning algorithm,and proposes an improved YOLOX-Fusion pointer meter detection and reading method based on YOLOX,which realizes high-accuracy meter information reading without any prior information,and improves the the universality of the algorithm.The main research contents of this thesis are as follows:(1)Aiming at the problem that the meter has low detection and positioning accuracy in the actual complex working environment,a meter object detection algorithm based on YOLOXFusion is proposed.The algorithm mainly completes the detection and positioning tasks of the meter panel,pointer and scale numbers.In order to improve the detection and positioning accuracy of small objects,a network structure improvement method based on CA attention mechanism and ASFF(Adaptively Spatial Feature Fusion)is proposed.The experimental results show that the average accuracy of YOLOX-Fusion in the task of meter object detection is 4.1% higher than that of YOLOX-X,which effectively improves the detection and positioning efficiency of meter images in the actual complex working environment.(2)Aiming at the problem that the meter reading needs a priori range information,a scale digit recognition algorithm based on LeNet-Mono is proposed.The algorithm mainly completes the task of identifying the scale numbers of the meter,and at the same time associates the corresponding coordinates to realize the judgment of the graduation value of the meter.In order to improve the accuracy of meter scale digit recognition,a network structure improvement method based on VGG structure and GAP(Global Average Pooling)is proposed.The experimental results show that the average accuracy of LeNet-Mono is 6.14% higher than that of LeNet-5 in digital recognition tasks,providing more accurate recognition results for subsequent meter readings.(3)Aiming at the problem of low accuracy of pointer pointing determination and pointer rotation center determination during meter reading calculation,which leads to large reading errors,a straight line fitting method based on binarized pixel traversal is proposed to improve the accuracy of pointer pointing determination;The rotation center fitting method based on vertical traversal can improve the judgment accuracy of the pointer rotation center.The experimental results show that the accuracy of the method proposed in this thesis is improved by 1.15% and 2.1% respectively compared with the traditional method,which effectively reduces the meter reading error.