SGBNet: An Ultra Light-weight Network for Real-time Semantic Segmentation of Land Cover

Designing a lightweight and robust real-time land cover segmentation algorithm is an important task for land resource applications. In recent years, with the development of edge computing and the increasing resolution of remote sensing images, the huge amount of calculations and parameters have restricted the efficiency of real-time semantic segmentation. Therefore, the emergence of lightweight CNN (convolutional neural network) has accelerated the development of real-time semantic segmentation of land cover. However, nowadays, the time and space span of aerial images is getting larger and larger, resulting in the loss of details and blurred edges of lightweight CNN. Therefore, the existing lightweight CNN model has low segmentation accuracy and poor generalization ability in real-time land cover segmentation tasks. In order to solve the problem of lightweight network in this respect, this paper proposes a Semantics Guided Bottleneck Network (SGBNet) to balance accuracy and reasoning speed. First, a basic unit and the overall network structure are redesigned to increase the overall reasoning efficiency of the model. The model can efficiently extract spatial details and contextual semantic information. Then, the model optimizes the lightweight network and realizes the extraction of details and contextual semantic information. Finally, a lightweight attention mechanism is used to restore high-resolution pixel-level features. The results of comparative experiments show that the method in this paper has a higher segmentation accuracy than existing models while achieving lightweight.