Pointer type instrument reading method based on key point detection

Accurate and generalizable instrument recognition is a requirement of multiple types of pointer instruments in power distribution plants. However, it is difficult to read instruments with limited visibility. A global geometric network (GGNet) key point detection model is proposed for pointer-type instrument recognition by analyzing the geometric structure of these instruments and detecting the scale key points, pointer and dial center. First, a spatial channel fusion (SCF) module is used to obtain the global sensory field, to enhance the detection effect of the pointer tip key points. Second, an ellipse-aware feature aggregator (EFA) module is designed to enhance the overall robustness of the model, adaptively capture the global geometric information of the key points on the instrument scale, and form a GGNet model, thus, the detection effect of the difficult-to-detect scale is improved in the scene of limited visibility, and high precision reading is finally achieved. The experimental results show that the algorithm achieves an accuracy of 94. 0% in key point detection and an average error of only 0. 37% in readings for the pointer gauge recognition task. The results comparing GGNet with four other different network models show that the proposed model performs best for a number of error metrics.

Authors

• Yongjie Zhai
• Wei Xu Georgia Institute of Technology
• Zhenyuan Zhao
• Guotian Yang
• Biqiang Du

Keywords

• Automatic data reading
• Deformable convolution
• Geometric information
• Key points
• Pointer type instrument