Promoting Inshore Ship Detection in SAR Images: A Fourier-Based Scene Transformation and Semantic Enhancement Framework

Deep learning-based synthetic aperture radar (SAR) ship detection approaches have made significant progress, but still face challenges in inshore scenarios. Inshore regions are cluttered with docks, port facilities, and buildings, which exhibit high visual similarity to actual ship targets in SAR images and are major factors that reduce detection accuracy. Mainstream methods typically focus on directly extracting ship target features from inshore scene images. However, such cluttered backgrounds hamper the effectiveness of feature extraction and ultimately decrease detection accuracy. To alleviate this dilemma, we revisit farshore and inshore images in the frequency domain space from a new perspective, and propose a novel Fourier-based scene transformation and semantic enhancement (FSTSE) framework that utilizes farshore target as an explicit prior information to improve inshore SAR ship detection. Specifically, inspired by Fourier properties, a farshore-inshore scene transformation module (Far-In-STM) is proposed. The Far-In-STM enriches the diversity of SAR image background environments by performing scene transformations between farshore and inshore scene images. The resulting transformed image is called the pseudoinshore image. Then, a dual consistency supervision mechanism (Dual-CSM) is designed, which imposes consistency constraints on gradients and prediction distributions between the farshore image and its corresponding pseudoinshore image. The Dual-CSM allows the discriminative semantics of ship targets to be fully captured and explored in complex backgrounds by exploiting the semantic correlation between images. Extensive experiments on several public SAR ship detection datasets indicate that the proposed FSTSE significantly outperforms other methods in inshore ship detection.