Weakly Supervised Learning for Oriented Ship Detection in SAR Images Using Multiscale Feature Enhancement and Angle Encoding

Abstract

This letter proposes a weakly supervised learning (WSL) framework for oriented ship detection in synthetic aperture radar (SAR) images, which tackles the challenges of detecting ships with arbitrary orientations using only horizontal bounding box (HBB) annotations. The proposed approach integrates three key modules: the Swin transformer with local context augmentation (STLCA) for enhanced multiscale feature extraction, recursive and hierarchical feature reconstruction (RHFR) for improved small-ship detection through iterative feature fusion, and the angle correct module with constraint (ACMC) to stabilize orientation prediction. The STLCA block enhances local feature extraction by dynamically adjusting attention windows, while RHFR performs multiscale feature fusion recursively to boost detection accuracy in complex environments. ACMC ensures precise angle prediction by applying constraints and leveraging periodic ambiguity properties. Comprehensive experiments on the SSDD and HRSID datasets demonstrate that our method achieves competitive performance with fully supervised methods, reaching 89.2% and $90.9\% \text {AP}_{50}$ , confirming its effectiveness and practical value for SAR-based applications.