Research on Multi-Scale Ship Target Detection Methods Under Complex Backgrounds

Synthetic Aperture Radar (SAR) imagery presents complex challenges in maritime object identification across diverse marine environments. The detection process spans two primary scenarios: oceanic expanses and coastal regions, each with unique detection complexities. Open maritime domains pose significant obstacles due to environmental dynamics. Ocean phenomena like wave patterns and vortices create intricate background interference, generating heterogeneous scattering zones that elevate false positive risks. Coastal areas further complicate detection through dense natural and artificial structures, producing intense electromagnetic backscattering signals that manifest as highintensity imaging artifacts. Vessel target morphology adds another layer of complexity. Maritime objects can display dramatically different visual characteristics depending on imaging resolution, challenging consistent detection algorithms. To address these multifaceted challenges, our research proposes an enhanced deep learning approach based on the YOLOv8 architecture. Key innovations include: Spatial Depthwise Convolution (SPD-Conv): A feature extraction technique minimizing information loss. Wavelet Transform Convolution (WT-Conv): An approach expanding network receptive field and enhancing small target recognition. We refined the model's localization loss function by integrating Wise-CIoU and Normalized Wasserstein Distance (NWD), enabling more nuanced feature extraction for diminutive and low-resolution maritime targets. Empirical validation using benchmark datasets demonstrates the method's superior performance, with comparative experiments revealing significant improvements in detection accuracy compared to existing algorithms. The research contribution extends beyond incremental enhancement, presenting a robust framework for addressing maritime target detection challenges across varied imaging conditions.