Global Fused Multiresolution Semantic Interactive Network for Low-Contrast Infrared Iceberg Detection

Ice floes and icebergs pose a deadly threat to ships. Ensuring safe navigation by timely detecting these hazards through shipborne detection systems represents a potent solution. Infrared iceberg detection stands out due to its exceptional visual capability and immunity to light and weather conditions, enhancing detection robustness. However, existing infrared small target detection methods struggle to effectively capture the features of the low-contrast ice target, making it difficult to detect the target. In this paper, we propose a Global Fused Multi-resolution Semantic Interactive Network (GMSINet) for feature extraction of ice targets. Firstly, we design a novel parallel encoder structure for complementary feature extraction. The structure incorporates an additional global branch that effectively retrieves submerged low-contrast features of the targets. The global branch effectively retrieves submerged low-contrast features of the targets. The local branch obtains the semantic feature of the target and reduces the loss of downsampling. Furthermore, we introduce the Multi-resolution Feature Interactive (MFI) module, which augments semantic features through the interaction of features across different resolutions. To further refine these enhanced features, we design the De-redundancy Block (DB), which emphasizes pertinent features while suppressing irrelevant ones. Finally, experimental results on our self-conducted SFIIT dataset and the publicly available SIRST dataset demonstrate that the proposed method performs satisfactorily in infrared target detection.