ConSeaNet: A Dynamic SAR Ship Detection Model via Unsupervised Contrastive Learning of Representative Features

In recent years, Synthetic Aperture Radar(SAR) ship detection technology based on deep learning has developed vigorously. At present, most of the mainstream SAR ship detection models rely on supervised learning, and the models based on unsupervised learning are rare. SAR ship images have the following characteristics: 1) Similarity: under the same intensity of noise interference, the images have similarity. 2) Difference: under different intensities of noise interference, the image difference is obvious. First, based on the similarity features, we propose a queue-based representative feature contrastive learning method (QRC). QRC is a method based on unsupervised learning, which significantly enhances the feature extraction ability through dynamic clustering feature comparison. Secondly, based on the differential features, we propose the Lambda Dynamic Detection Head (LDyhead). LDyhead utilizes a set of adaptive vectors to dynamically distinguish between weak and strong noise images, thereby selecting different detection paths and significantly enhancing detection performance. Finally, we propose ConSeaNet, a dynamic SAR detection model based on unsupervised learning, and verify the effectiveness of the proposed method through experiments. ConSeaNet achieves the best results compared to state-of-the-art methods. Source code will be released upon publication.