Aiming at the problem that it is difficult to collect arrester image data in outdoor environments and is easily affected by strong light

resulting in poor segmentation effect

a mask region-based convolutional neural network（Mask R-CNN）+PointRend network based on transfer learning and feature channel improvements is proposed to solve this problem.The original network lacks the fusion between data feature channels.This method adds the squeeze-and-excitation network(SENet)module to the backbone residual network(ResNet)

so that the network can learn the importance information of the feature channels and better segment the arresters.At the same time

the synthetic dataset is used for transfer learning

improves the generalization ability and segmentation average precision(AP)of the improved network.Experimental results show that the segmentation AP of the proposed method is 3.742 higher than the Mask R-CNN+PointRend method

about 12.6%.The AP of the network segmentation after transfer learning is 2.507 about 7.5%

higher than that of the network trained in real dataset only

which effectively solve the problems of difficulty in real arrester data collecting and poor segmentation edges.