基于幽灵卷积多尺度特征融合的船舶检测算法

Ship target detection algorithm under adverse weather conditions

摘要:

如今船舶目标检测技术在不断发展，但在雾天、雨天和雪天等恶劣天气条件下，船舶目标的可见度低，导致模型特征提取不全面，检测性能会显著下降。针对以上问题，提出了一种轻量化目标检测模型GSPN-YOLO。该模型基于YOLOv8s架构，在主干网络采用幽灵卷积替换普通卷积，使用GhostNCSP模块替代C2f模块，降低计算量的同时提高网络精度。在颈部网络设计了通道重排多尺度特征金字塔，显著提升了多尺度特征的表征能力。经实验测试，GSPN-YOLO在自制数据集上，相较于基准模型YOLOv8s，其AP@50:95提升了2.1%，AP@50提升了2.7%，AP@75提升了2.5%，同时参数量减少了45.5%。此外，其检测性能优于YOLOv10s、YOLOv11s等SOTA目标检测算法，满足了在复杂天气条件下船舶检测的应用需求。

Abstract:

With the continuous development of ship target detection technology, the visibility of ship targets is low under adverse weather conditions such as fog, rain, and snow, which leads to incomplete feature extraction of the model and a significant decline in detection performance. To address the above issues, a lightweight target detection model named GSPN-YOLO is proposed. This model is based on the YOLOv8s architecture, where the ordinary convolution in the backbone network is replaced by Ghost convolution, and the C2f module is replaced by the GhostNCSP module. This approach reduces the computational load while improving network accuracy. In the neck network, a channel-rearranged multi-scale feature pyramid is designed, which significantly enhances the representation ability of multi-scale features. Experimental tests show that on a self-made dataset, compared with the baseline model YOLOv8s, the AP@50:95 of GSPN-YOLO is improved by 2.1%, AP@50 by 2.7%, and AP@75 by 2.5%, while the number of parameters is reduced by 45.5%. Moreover, its detection performance is superior to that of state-of-the-art (SOTA) target detection algorithms such as YOLOv10s and YOLOv11s, meeting the application requirements for ship detection under complex weather conditions.