A Mask RCNN based Automatic Reading Method for Pointer Meter

Abstract

Automatic reading of pointer instrument is of great significance to realize intelligent monitoring of substation. Many scholars have proposed kinds of schemes based on the traditional image processing methods. However, the environmentally sensitive issues have not been effectively solved, which prevents the promotion of automatic reading under complex environmental conditions. In recent years, convolutional neural networks (CNN) have been proved to be suitable for image recognition tasks. The results of image object detection by deep learning methods are greatly improved compared with the conventional methods. Based on the premise that deep learning methodology is developing rapidly, this paper proposes an automatic reading method for pointer meter with keypoint detection. The main idea is to first apply improved Mask R-CNN to detect the key points of the scale and pointer on the meter, then utilize the detected key points to fit the circle composed of the tick marks and the straight line of the pointer, and finally calculate the reading value based on the deflection angle of the pointer relative to the scales. The experimental results demonstrate that the proposed method is better than the traditional Hough Transform based algorithms in terms of accuracy and robustness.