GFCR-Net: Gradient-Guided and Frequency-Gated Context Refinement for Robust Ship Detection in Complex Optical Remote Sensing Images

Abstract

Robust ship detection in remote sensing imagery is essential for aerial surveillance, yet it remains challenging under adverse environmental conditions such as fog, haze, thin clouds, and sea clutter, as well as in dense and crowded port scenes. Existing methods often suffer from ambiguous receptive fields due to large variations in ship scales and complex surrounding contexts. This weakens target–background separability and results in frequent false alarms, missed detections, and inaccurate localization in complex maritime environments. To address these challenges, we propose the gradient-guided and frequency-gated context refinement network (GFCR-Net), which strengthens backbone feature representations under difficult maritime conditions through carefully designed modules. First, the gradient-guided boundary enhancement (GGBE) module strengthens boundary-consistent activations while suppressing background interference using spatial gating on feature gradients. Second, the dual-stream contextual refinement (DSCR) module synergistically integrates the frequency-gated boosting module (FGBM) and the global relational context modeling (GRCM) module. FGBM decomposes features into low-frequency components and high-frequency details to preserve ship contours while suppressing dominant background textures. Meanwhile, GRCM captures long-range contextual dependencies to better distinguish ships from nearby docks and ship-like clutter, especially in dense port scenes. Together, these components enable refined feature aggregation across both frequency and spatial dimensions. Extensive experiments on SCCOS, HRSC2016, and FGRSCS demonstrate consistent improvements over strong baselines while maintaining competitive parameter counts and FLOPs. Qualitative results further show fewer missed detections for small or partially visible ships under cluttered and low-visibility conditions. The code is publicly available at: https://github.com/tanish1403/GFCR-Net.