Author(s): Rui Zhou; Qianyang Sun; Li Zhou; Runxin Xu; Pei Xu
Linked Author(s):
Keywords: Polar sea ice; Image dehazing; Semantic segmentation; Cascade perception model
Abstract: With global warming, the navigable period for Arctic shipping routes is lengthening, and polar shipping activities are becoming increasingly frequent. Accurate sea ice information has become an important prerequisite for navigation safety in polar waters. However, shipborne visual images are often affected by sea fog in actual navigation environments. Target edges are blurred. Contrast is reduced. Scale features are weakened. Accurate segmentation of sea ice targets is thus affected. To address these problems, a dehazing-segmentation cascade perception model that integrates image dehazing, geometric correction, and pixel-level sea ice segmentation is proposed. Clarity and visual quality of shipborne images are improved by a dehazing model. Interference from complex meteorological conditions is reduced. On this basis, geometric correction is realized by a spatial transformer network. An improved DeepLabV3+ model is introduced for sea ice semantic segmentation. A complete visual perception pipeline for shipborne polar scenes is thus constructed. The results show that the optimized dehazing model improves the contrast, brightness, and color saturation of polar shipborne images. Peak signal-to-noise ratio is increased by 3.87dB. Structural similarity reaches 0.95. A pixel accuracy of 91.89% is achieved in sea ice segmentation. A mean intersection over union of 78.97% is achieved. Proposed cascade perception model can effectively improve sea ice segmentation in shipborne images under complex polar conditions. Technical support can be provided for navigation safety assurance and environmental perception of polar vessels.
Year: 2026