Cross-Modal Ship Detection From Optical to SAR Images Based on Pixel- and Feature-Level Progressive Transfer

Synthetic aperture radar (SAR) ship detection has advanced through deep learning but often requires many accurately annotated SAR images, which are time-consuming and labor-intensive. To address this, we propose an unsupervised cross-modal adaptation method based on pixel- and feature-level progressive transfer (CD-PFPT). CD-PFPT effectively transfers knowledge from optical to SAR domains, enhancing ship detection performance in unlabeled SAR images. Our approach consists of two components: pixel- and feature-level cross-domain adaptation. At the pixel level, a texture-enhanced style generation network called Edge_CycleGAN is designed. This network integrates the Haar wavelet transform to convert spatial-domain into frequency-domain information, ensuring that the generated images emphasize more on the texture and structural information while effectively reducing common artifacts and distortions in cross-modal image translation. In addition, the proposed novel edge loss function ensures that Edge_CycleGAN generates samples matching the SAR image distribution. At the feature level, we propose a multiscale feature enhancement and denoising (MFED) module, which uses spatial attention to suppress background noise and multiscale convolutional extraction to capture diverse ship sizes. Furthermore, a cross-domain loss is proposed to calculate the mean embedding distance in the reproducing kernel Hilbert space (RKHS) using multiple kernel functions, thereby reducing distribution differences between optical and SAR images. In cross-domain tasks from the optical dataset of HRRSD to the SAR dataset of SAR-Ship, our method improves the ${F}1$ -score by up to 3.9% and the mAP0.5 by up to 6.1% compared to the next best semi-supervised method.