Unsupervised Domain Adaptation for SAR Ship Detection Based on Multitask Decoupling

Remote sensing ship detection is crucial for maritime surveillance, especially with the rise of Synthetic Aperture Radar (SAR) images, known for their all-weather capabilities. Yet, annotating SAR images is resource-intensive and requires specialized expertise. Facing the challenge, we propose an unsupervised domain adaptation method for SAR ship detection based on multi-task decoupling, which leverages data annotated with optical images instead of SAR annotations. In domain adaptation object detection, conflict arises between transferability required for adaptation and discriminability required for detection. Similarly, within object detection, conflict in feature preferences between classification and regression tasks occurs. We address these conflicts by decoupling the adaptation module from the detector module and decoupling the adaptation module into category, and bounding box adaptation modules. Our detector module integrates a feature enhancement network (FEN) to improve the classification and localization ability under the challenge of dense small targets with low visual contrast in SAR images. The adaptation module employs adversarial adaptation and cross-domain category alignment (CCA) to pull up the feature representations of different domains of ships for domain alignment. Additionally, the cross-domain image synthesis module (CDIS) generates transitional images covering SAR domain style and optical domain content to bridge the huge gap between the domains. Through the separation and coordination of different modules, our multi-task decoupling framework strategically balances feature transferability and discriminability. Extensive experiments on classic remote sensing datasets validate the effectiveness of our method, showing better performance than state-of-the-art methodologies.