Fine-grained Classification of Substation Meters based on Vision Transformer

With the increasingly heavy task of power inspection, intelligent power inspection robots are used in substations under high-risk working environments, which possess popularization and application value. The use of inspection robots not only saves labor costs but also makes the detection task more efficient. We present a fine-grained classification model of substation meters based on a vision transformer. First, it uses the deformable patch module to adaptively segment the image into patches with different positions and proportions. In this way, each patch can retain the complete semantic information of the target. Then, we use a transformer architecture for fine-grained recognition (TransFG) to improve the ability of the network to capture subtle differences. Finally, we introduce a contrastive loss function to expand the gap between similar category features further. The experimental results show that the method can accurately classify multiple meters commonly used in substations. Compared with the baseline, the fine-grained feature extraction in meters is more accurate and interpretable, which lays a foundation for the subsequent detection of substation meters.