A Two-Stage Deep Learning-Based Approach for Automatic Reading of Analog Meters

In many factories in Japan, there are still thousands of analog meters to monitor and control the fabrication environments. Human inspectors must walk around, read the meters, record the values, and input them into the computer for further analysis. This process is prone to human errors. To solve the problem, we may consider to replace all analog meters with digital ones, but this is not practical due to the high replacement cost. In this study, we propose a system to automate the meter reading process. Using this system, an inspector just takes a short video of the meter using a smart phone, sends the video to a server, and the server can read the meter value and record the results automatically. To make the system practically useful, we introduce three techniques. The first one is a two-stage deep learning-based method for detecting the meter and its pointer; the second one is homography transform to normalize the view point for capturing the video; and the third is a multi-frame-based method for improving the reading accuracy. Experimental results with circular meters show that the accuracy of the proposed system is acceptable. This paves a way for us to automate the reading process of other types of analog meters.