Automatic Meter Pointer Reading Based on Knowledge Distillation

With the rapid development of industrial automation, automatic reading of pointer meters has become a trend of data monitoring and efficient measurement in the industrial field. In the process of industrial meter inspection, the automatic reading of the meter is an important means to reduce costs and labor. Improving the accuracy and speed of automatic reading is always an important research topic in this field. In recent years, deep learning has been applied to meter reading. However, due to the large model parameters, it is difficult for mobile devices to carry, and the meter tilt results in inaccurate reading, which leads to poor performance of commonly used detection and recognition algorithms. To solve these problems, a Knowledge Distillation and Keypoint Detection Network (KKN) is designed for automatic meter reading. This network is divided into teacher model and student model. In the teacher model, feature filtering and offset correction are used to improve the ability of keypoint detection. In the student model, parameter information of the teacher model is learned to imitate the learning ability of the teacher model. Finally, the readings are calculated using the coordinates of the three keypoints detected. In a series of comparative experiments, our method achieves higher accuracy than other methods.