DINO-CUT: Contrastive Learning Guided by Decoupled Structure for Optical-SAR Conversion and Ship Detection

Synthetic Aperture Radar (SAR) plays a vital role in ocean monitoring, yet the scarcity of annotated data hinders its development for target detection. While generating SAR images from optical imagery is an effective solution, mainstream unpaired transformation methods that rely on Convolutional Neural Network (CNN) encoders result in generated images exhibiting structural distortions and missing details. This paper therefore proposes the self-DIstillation with NO labels-Contrastive Learning for Unpaired Image-to-Image Translation (DINO-CUT) framework, which decouples the generator and feature extractor. It incorporates the pre-trained, parameter-frozen Distillation with No Labels for Vision Transformer (DINO-ViT) as a structural arbiter, leveraging its global semantic modeling capabilities to provide structural guidance. Experiments demonstrate that DINO-CUT improves training stability and the structural accuracy of generated images. Its synthetic data augmentation achieves superior performance in downstream SAR vessel detection tasks compared to existing methods. This research provides valuable insights into cross-modal remote sensing image generation and data augmentation tasks. DINO-CUT demonstrates superior training stability compared to the original CUT, with Fréchet Inception Distance (FID) scores consistently below 250 and the cosine similarity curve maintaining values above 0.55. When deployed on the You-Only-Look-Once version 11 (YOLOv11) detector, it achieves an [email protected] of 83.45%, representing improvements of 0.7% and 3.1% over the original model and baseline.