Detail-attention convolutional neural network for meter recognition

Abstract

The classical meter recognition method is severely affected by light and noise interference in the recognition of meters, which limits its generalization ability. Recent methods based on convolutional neural networks cannot handle the diversity of meters via detection of the key points (center of circle, maximum and minimum range). In this study, we investigate the recognition of a mechanical meter. According to the manual reading paradigm, the scale and pointer are first positioned, and the reading value is then determined based on the relative position of the pointer and scale, which is compatible with different meters. To extract the tick marks of meters in a stable and precise scheme, Meter scan net (MSN) is designed based on the low-occupancy features of the meter's tick area to predict the heatmap. Utilizing our proposed Detail-Attention module, MSN can greatly retain detailed features and focus on the tick area, which is helpful for the final segmentation. Post-processing includes using the output of MSN to perform circle fitting to obtain the scale arc, sampling along the normal phase direction of the scale arc, and locating the scale and pointer position. To handle cases of missing tick marks, a frequency-based analysis method is proposed to restore the missing tick marks, which enlarges the application scope of our method. Finally, extensive experiments are conducted that reveal that the proposed method has high accuracy and is compatible with different types of meters.