BEB-Net: Boundary Extraction Based Semantic Segmentation Network for Indoor Scenes in Smart Power Plants

Point cloud segmentation can help power inspection robots to recognize and detect power equipment. Due to the complexity of indoor power point cloud scenes, the segmentation effect of the boundary is often not accurate enough. To address this problem, this paper proposes a semantic segmentation method for indoor scenes that fully utilizes the multidimensional information of the original point cloud (BEB-Net). Compared with existing deep learning methods, BEB-Net pays more attention to the boundary information features of the original point cloud. The explicit feature representation of the initial point cloud can be enhanced by calculating the local feature variance of the point cloud while using the geometric information of the input point cloud as independent features. The method maintains point cloud feature better and reduces the loss of information during the coding process. In addition, BEB-Net, based on inverted residual scaling blocks, can demonstrate better performance in the S3DIS indoor scenes dataset, with an mIoU of 69%. At the same time, there has been some improvement in the boundary segmentation effect in indoor scenes.