A Pointer Instrument Reading Approach Based On Mask R-CNN Key Points Detection
Xiaolong Wang, Junyan Chen, Huyong Wang · 2021
In the field of automatic reading of instruments, the research direction is roughly divided into two categories. One is to study a kind of intelligent instrument which realizes automatic report reading from hardware. However, the cost of hardware optimization is too high. The other is to realize the reading function of the instrument by studying a suitable approach, so as to simplify the tedious steps of artificial operation. Algorithms already exist to implement pointer instrument readings, but these use either binarization or object detection. The binarization method is very sensitive to the background noise, while the object detection method still needs to binarize the image to detect the pointer after detecting the pointer instrument. So it is of great significance to find a robust and efficient approach to solve this problem. This paper proposes a pointer instrument reading approach based on Mask R-CNN to extract key points of the pointer instrument. The approach firstly uses Mask R-CNN to detect the prior bounding box of the pointer instrument, and then extracts key points of the pointer from the frame, the key points of the pointer are the center of the pointer instrument and the tip of the pointer. Compaired with other existing methods, our method is more precise and more flexible.