A Real-Time Semantic Segmentation Method Guided by Boundary Information for Substation Scene

Automatic inspection robots are widely used in the management of unattended substations. For higher-level tasks or understanding its road scenes, the robot needs to extract semantic information from environment, such as roads, buildings, and equipment. However, small devices in substations and real-time requirements are quite a challenging task in semantic segmentation. To meet these demands, we propose a new semantic segmentation network with high real-time. The proposed method contains a contextual branch and a boundary-guided branch, which continuously interacts with another branch through an attention mechanism to improve the accuracy of boundary segmentation between neighbouring objects. Our approach is an excellent balance of accuracy and speed on a new dataset from a transformer substation. It can overcome the problems of missing segmentation of small targets and blurry object boundaries in semantic segmentation, and has good real-time performance. Additionally, compared with the state-of-the-art methods, the proposed method gets advanced accuracy and speed on a publicly available road scene dataset—Cityscapes.