SSANet: A Scale and Shape Awareness Network for Ship Detection in Remote Sensing Scenes

Ship detection within remote sensing images is vital for tasks such as maritime surveillance, national security, and emergency response. Although mainstream object detectors work well on natural photographs, transferring them straight to remote-sensing scenes often leads to inferior performance. This performance gap stems largely from two characteristics of ship targets: pronounced scale variability and very high aspect ratios. To jointly tackle these factors, we propose the Scale and Shape Awareness Network (SSANet). For scale awareness, a standard convolutional pathway runs alongside a Multi-scale Stacked Dilated Convolution (MSDC) branch, harvesting fine-grained local cues and long-range contextual information, respectively. To integrate these heterogeneous representations, we devise a Spatio-Channel Cross-Learning (SCCL) mechanism. Unlike conventional spatial–channel attention modules, which derive both spatial and channel weights from a single feature tensor to reweight itself, SCCL operates on two heterogeneous branches in a cross manner: scene-level cues from the MSDC branch modulate the spatial attention over local features to suppress background clutter, while local evidence from the standard convolution branch sharpens the channel weighting of global features to recover details, yielding complementary fusion across spatial and channel dimensions. For shape awareness, the SCCL-fused features are processed by a deformable convolution block whose sampling pattern adapts to the elongated geometry of ships, enabling precise feature focusing. Experiments on HRSC2016 and SSDD+ report mAP scores of 96.4% and 98.40%, respectively.