Precise and fast segmentation of sea ice in high-resolution images based on multiscale and knowledge distillation

The complicated distribution patterns of sea ice pose a challenging issue in terms of establishing effective monitoring and early warning systems. Convolutional neural network (CNN) is used by its powerful learning capabilities for accurate segmentation of sea ice areas. In this paper, the connections between large objects and their surrounding pixels are achieved through CNN. In order to enhance the detection of tiny objects, innovative methods including data augmentation, creative design of the loss function, and multi-scale strategies are implemented. Knowledge distillation is conducted with the primary network to improve the effectiveness for real-time application. Comparison and ablation experiments demonstrate the effectiveness of the proposed methods on a large open dataset of sea-ice segmentation.