MU-Net: An efficient algorithm for foreign objects detection in electric equipment
Jiaxin Liu, Xuke Zhong, Yuzhong Zhong, Xuchen Lu, Shengchuan Li, Rui Guo, Songyi Dian · 2022
The detection of foreign objects inside the electric equipment is an important task to maintain the normal operation of the power grid. Aiming at the problems of low efficiency and easily missed detection in the current manual inspection, this paper proposes an improved multi-channel electric equipment foreign objects segmentation method MU-Net based on U-Net. It fuses information from multimodal foreign objects images and uses this information to segment foreign objects. Two different networks of MU-Net perform shallow feature extraction and feature fusion on the original image and the filtered image respectively; at the same time, we use Transformer for deep feature extraction. This attention mechanism can enhance the feature learning of foreign pixels, make the feature extraction more perfect, thereby improving the accuracy of foreign objects segmentation in electric equipment. The experimental results show that MU-Net achieves 92.5 in F1 and 85.24 in mIou on our dataset. Compared with other excellent segmentation networks, the segmentation accuracy has been improved, which fully verifies that MU-Net can effectively Perform foreign objects detection in electric equipment.