| dc.description.abstract | Computer vision-aided scene understanding has drawn much attention, especially in autonomous vehicles and advanced driver assistance systems. Image segmentation forms the core of these systems, which is required for effective planning of maneuvers. Recently, with more navigational routes opening up in the Arctic ocean, computer-aided understanding of the ice conditions in front of a surface vessel has become very important for pilot assistance systems because understanding the severity of ice conditions is essential for safe ice navigation. The floating ice features need to be correctly detected, classified and accurately segmented, and the path needs to be carefully planned to avoid excessive ice loads on the ship hull and the ship besetting in ice. With the increasing popularity of deep learning in recent years, many deep architectures have been successfully proposed for solving a number of segmentation problems. This paper focuses on polar scene segmentation tasks. The goal is to find a model that can accurately classify and locate the following surface ice features: icebergs, deformed ice, level ice, broken ice, ice floes, floebergs, floebits, pancake ice, and brash ice. In this paper, it is shown that this segmentation task can be solved with a fully convolutional neural network (U-Net) model with a pretrained residual neural network architecture (ResNet) as the downsampling part. The performance of ResNet18, ResNet34, ResNet50 and ResNet101 on images collected online and from the cruise to the Fram Strait in 2012 was analyzed. In addition, the model performance coupled with convolutional conditional random fields (CRFs) and fully connected CRFs was analyzed and compared. Preliminary results indicate that the U-Net model with ResNet101 as the backbone in combination with the convolutional CRF-based postprocessing technique performs the best on available optical images of ice cover. | en_US |
