Lightweight Maritime Surveillance System Based on ResNeXt-18 and YOLO-HPSD

Marine intelligent monitoring is of great significance for safety early warning. However, due to complex marine conditions (dynamic waves, sudden changes in illumination, and scale differences), existing methods are insufficient in small target detection and false judgment control, and relevant research remains limited. This paper proposes a two-stage lightweight marine monitoring framework for computing-constrained devices, with the core process of "abnormal frame detection - ship recognition" and the integration of closed-loop optimization. In the abnormal frame detection stage, the ResNeXt-18 backbone network is improved and a spatiotemporal attention module is embedded to effectively suppress background interference such as ocean waves; meanwhile, a hybrid-kernel LSSVM is proposed to adapt to the non-uniform distribution of marine surface features. In the ship recognition stage, YOLO-HPSD is improved by introducing dynamic anchor box clustering and WIoU loss, which significantly enhances the detection accuracy and positioning stability of long-distance small targets. Additionally, a cross-stage closed-loop optimization mechanism is constructed to convert detection errors into lightweight parameter corrections for presequence modules, enabling adaptive enhancement of the system. Experimental results show that the proposed method achieves an F1-score of 93.8% in abnormal frame detection, a mAP@0.5 of 98.45% in ship detection, and an end-side inference speed of 137 FPS, outperforming various lightweight baseline methods. Ablation experiments verify the independent contributions and synergistic effects of each module improvement. Overall, this study provides a solution that balances high accuracy and real-time performance for marine surface monitoring in resource-constrained environments.