Multilevel intuitive attention neural network for airborne LiDAR point cloud semantic segmentation

Abstract

Three-dimensional laser scanning technology is widely employed in various fields due to its advantage in rapid acquisition of geographic scene structures. Achieving high precision and automated semantic segmentation of three-dimensional point cloud data remains a vital challenge in point cloud recognition. This study introduces a Multilevel Intuitive Attention Network (MIA-Net) designed for point cloud segmentation. MIA-Net consists of three key components: local trigonometric function encoding, feature sampling, and intuitive attention interaction. Initially, trigonometric encoding captures fine-grained local semantics within disordered point clouds. Subsequently, a multilayer perceptron addresses point-cloud feature pyramid construction, and feature sampling is performed using the point offset mechanism in the different levels. Finally, the multilevel intuitive attention(MIA) mechanism facilitates feature interactions across different layers, enabling the capture of both local attention features and global structure. The point-offset attention scheme introduced in this study significantly reduces computational complexity compared to traditional attention mechanisms, enhancing computational efficiency while preserving the advantages of attention mechanisms. To evaluate the results of MIA-Net, the ISPRS Vaihingen benchmark, LASDU and GML airborne datasets were tested. Experiments show that our network can achieve state-of-art performance in terms of Overall Accuracy(OA) and average F1-score(e.g., reaching 96.2% and 66.7% for GML datasets, respectively).