Abstract
Fine-grained semantic segmentation of iceberg visual images obtained from low altitude is of great significance for ship navigation and ecological surveys, which can be regarded as the key technology of iceberg management. However, because of the bad weather in Polar regions and resource constraints, the similar texture of different ice images is the major challenge for iceberg identification, which may lead to false navigation for the ship controlling system. The main purpose of this paper is to generate accurate pixel-level iceberg image segmentation and prediction under poor sea ice optical vision acquisition conditions. A new feature extraction network IBNet with multi-scale features and double-layer attention mechanism is proposed for pyramid pooling module layer to realize weighted fusion. This learning-based approach considering six kinds of iceberg backgrounds in the scene of polar regions can be applied to support obstacle avoidance algorithm for Polar region navigation. By comparing PA, MIoU and other parameters with the other methods, IBNet performs obvious advantages in semantic segmentation task of icebergs in polar regions.
