Research on reading recognition of pointer meter based on improved U-net network

Abstract

Aiming at the automatic reading task of pointer meter in industrial applications, in order to solve the problems of indistinct meter features, complex background environment and slow recognition speed involved in the reading process of pointer meter, a new type of U-net network is proposed to segment Instrument hands and scales. The network replaces the backbone feature extraction part in the traditional U-net network with the bneck module in MobileNet, and fuses it with the enhanced feature extraction part. First, use web crawlers and data enhancement to expand the dataset; second, use the YOLO network to filter out the complex background information of the collected images and extract the dashboard area. Input the dial area information into the improved version of the U-net network to segment the meter scale and pointer; finally, convert the annular image into a rectangular image, and obtain a one-dimensional array of scale and pointer values through the rectangular image, and then calculate Pointer meter readings. Theoretical analysis and experimental results show that the new U-net can segment the pointer and scale of the meter more accurately, and the segmentation accuracy is increased from 90.79% of the traditional U-net network to 93.63%. The proposed automatic reading method of the pointer meter has good performance. The accuracy and anti-interference ability of the system, the average basic error is 0.48%, which meets the requirements of the application.