With the deepening of the polar “new frontier” strategy, it is urgent to improve the auxiliary decision-making ability of ships under polar environmental conditions. Getting the sea ice thickness during the icebreaker’s voyage can help icebreakers grasp the ice situation in real-time, assist the captain in making scientific navigation decisions, and ensure the stability of the entire operation process. Using deep learning to recognize ship-side overturning sea ice can obtain sea ice thickness during the ship’s travels. This article utilizes the images of sea ice overturning on the side of the vessel captured by the on-board camera during the Antarctic voyage of the icebreaker Xuelong 2, we propose a polar sea ice thickness recognition technique based on the improved U-Net model. Based on the U-Net network, the improved EMA attention mechanism and VGG decoder structure are utilized and combined with the weighted loss function to recognize the thickness of sea ice overturned on the ship side. The experimental results prove that the mean pixel accuracy of the polar sea ice thickness recognition algorithm based on the improved U-Net model is 93.26%, and the detection frame rate is 14 frames/s. Compared with the traditional U-Net model, the detection frame rate is only 4 frames/s lower, while a 4.85 percentage point increase in the mean pixel accuracy. Conducted comparative experiments with algorithms such as Deeplabv3+ and PSPNet. For the mean pixel accuracy, the proposed algorithm is 7.94 and 9.02 percentage points better than the Deeplabv3+ and PSPNet algorithms, respectively. For detection speed, the research algorithm is 62.5 frames/s and 112 frames/s lower than the Deeplabv3+ and PSPNet algorithms, respectively. Compared to the mechanisms, it is 4 frames/s lower than KAN, 2 frames/s lower than CBAM, ECA, and 3 frames/s lower than CA. Combining the accuracy and running speed, the improved U-Net model is more suitable for sea ice identification in the navigation of icebreakers. These results provide a reference for the auxiliary navigation of polar icebreakers.
Xing et al. (Thu,) studied this question.