Spatio-Semantic Enhanced Cascade Swin Detection Network for Infrared Maritime Targets

To address the issues of high missed detection rates in maritime infrared small and weak target detection łcaused by imbalanced target scales, strong ocean wave interference, and low signal-to-clutter ratiołthis paper proposes a convolutional neural network based on cascade fusion and local feature enhancement. Built on the CSPDarknet53 backbone network, the proposed network incorporates a structural tensor spatiotemporal information enhancement module and a dynamic context-aware attention semantic enhancement module to improve the representational capabilities of shallow spatial features and deep semantic features. Meanwhile, a hierarchical feature bidirectional symmetric fusion network and a scale-aware upsampling operator are designed to achieve effective interaction and reconstruction of multi-scale features. Additionally, Swin transformer is adopted as the detection head to enhance the modeling of long-range semantic dependencies.Experimental results demonstrate that on the self-constructed infrared maritime target dataset, the proposed network achieves a Precision of 0.865, a Recall of 0.875, and an F1-score of 0.870. These performance metrics outperform those of mainstream methods such as RLCM, PSTNN, Faster RCNN, YOLOv5s, YOLOv8s, and AGPCNet. Particularly, the proposed network exhibits high robustness and real-time performance in complex background environments and small/weak target detection tasks.