Appearance Memory Metrics for Online Electrical Equipment Segmentation with Structural Encoder-Decoder

The rapid development of intelligent detection of electrical equipment based on deep learning effectively ensures the normal operation of the power system. However, existing methods hardly separate electrical equipment from the background efficiently for high-precision state recognition. In this paper, we propose an appearance memory metric for online electrical equipment segmentation with structural encoder-decoder. First, the appearance memory bank is designed to address uncertainty changes of the object. High-quality memory samples are collected according to an update strategy. On this basis, the memory similarity metric network is proposed which utilizes temporal information to sense the pixel-level position distribution of electrical equipment. Second, pairwise structural encoder-decoder is proposed. The encoder jointly encodes the captured structural features along with the depth features from the backbone and the position distribution via the attention mechanism to learn a suitable feature representation for the current detection scene. The decoder fuses coded features and structural features to reconstruct the mask representation of the electric equipment. Finally, the proposed method is evaluated via real application scenarios and popular benchmark to prove the practicality of the proposed method.