Instance Segmentation Model for Substation Equipment Based on Mask R-CNN*

Abstract

Accurate instance segmentation of substation equipment scene image is beneficial to eliminating background interference and completing more efficient fault detection tasks. However, it is difficult to segment complex substation scenes with a large number of substation equipment. In this paper, we propose a substation equipment image dataset. On this dataset, we train and evaluate substation equipment segmentation models based on mask-RCNN. The experimental results show that our model has more than 69.1% mAp in the verification set, and has good segmentation effect in different scenes and lighting conditions. We also try to introduce the automatic data augmentation into the model training to expand the dataset and further improve the model performance, but the experimental results show that using more data augmentation methods cannot improve the model’s mAP. In addition, based on a smaller bimodal dataset of visible light and temperature map, we compare the effect of the instance segmentation models based on visible light and temperature map. The experimental results show that the segmentation model based on visible light is more accurate than the temperature map model.