Research On Reading Recognition Of Industrial Pointer Instrument Based On Deep Learning

Because of its good stability and durability,pointer meter plays an irreplaceable role in the complex environment of strong electromagnetic,such as petrochemical industry,military aerospace,electric power and so on.Meter inspection in industry mostly depends on labor,which not only consumes lots human resource costs,but also is vulnerable to environmental impact and has low work efficiency.Therefore,improving the intelligent level of meter inspection has vital research value and broad prospects.The research on pointer meter recognition and reading method combining deep learning and image processing was carried out for solving the problem The research contents are as follows:1)An improved meter detection algorithm based on yolov3 network was proposed.Firstly,the meter dataset was expanded,and the k-means++ algorithm was used to cluster the initial anchor to improve the detection speed;Then,two channel attention mechanism and residual block were added to yolov3 network to improve the detection accuracy.The results show that the average accuracy of the training and testing of the three types of meter on the dataset is 90.8%,which is about 2.1% higher than that of the original yolov3.2)Aiming at the problem of range recognition,an improved range digital detection and recognition algorithm based on yolov4 was proposed.Leading out a large feature map from backbone network and adding a detection head to improve the accuracy of digital small target detection.Integrating depthwise separable convolutions reduced model computation and improved detection speed.Numerical merging was realized based on relative position relationship of numbers.3)The indication reading includes two steps of aiming and reading.In the aiming link,the pointer was extracted through image preprocessing,rough extraction,fine extraction,thinning,line fitting and other steps.In the reading link,it was proposed to determine the pointer deflection angle and reading by the position relationship of range number,image center and pointer line.The experimental results show that the relative error of reading is less than 5%.Figure 69;Table 5;Reference 67...