Stern Target Detection Algorithm Based on Lightweight YOLO

To meet the dual demands of high accuracy and real-time performance in stern target detection during ship berthing operations, this paper proposes TriLight-YOLO, a lightweight and efficient object detection algorithm based on YOLOv8n. The proposed model addresses the challenges of high computational cost and poor robustness in complex maritime environments through three key innovations. First, the C2f-DGBlock module integrates GhostNet and dynamic convolution to significantly reduce parameters and computational load while preserving feature representation. Second, a BiFPN feature fusion structure enhances multi-scale detection capability by facilitating efficient bidirectional cross-scale connections. Third, an Efficient Multiscale Attention (EMA) mechanism is introduced to stabilize dynamic convolution outputs and improve detection robustness under occlusion and low-light conditions. Experiments conducted on a self-built shore-based ship dataset demonstrate that TriLight-YOLO achieves a [email protected]:0.95 of ${7 3 . 5 \%}$ while reducing FLOPs by $23.17 \%$ and model size by $22.8 \%$, showing strong potential for deployment in real-world port environments.