Lightning Arrester Target Segmentation Algorithm Based on Improved DeepLabv3+ and GrabCut

Abstract

Since the camera is severely disturbed by sunlight during the surge arrester replacement task of the distribution network robot, and the operation background is complex, it is important to achieve accurate lightning arrester target segmentation. Aiming at the problem that the GrabCut algorithm requires user initialization in lightning arrester image segmentation and has a poor effect on segmentation of images with similar foreground and background colors, a lightning arrester segmentation algorithm based on improved DeepLabv3+ and GrabCut is proposed in this paper. Firstly, the acquisition images disturbed by sunlight are processed by the EnlightenGAN network for low light enhancement; then the lightning arrester target is segmented for the first time with the improved DeepLabv3+ network, then the segmented regions are made into dilation and erosion operations respectively, and the obtained results are used as the input images and initialization regions of GrabCut separately, and through two segmentations, the lightning arrester target and The separation of the operation background. The experimental results show that the algorithm proposed in this paper can not only realize the automatic segmentation of GrabCut but also can segment the target lightning arrester very well.