雾环境下的船舶目标检测研究

Research on ship object detection in foggy environments

针对船舶在雾环境中因能见度不良易发生碰撞、船舶识别困难和检测精准度较低等问题:首先构建出雾环境下的船舶目标检测数据集;接着在YOLOv5基础上进行改进,在原网络结构基础上,为使深度可分离卷积更接近可分离卷积,使用GSConv模块替换Head部分CBS模块,以提高模型精准度,并引入Slim-Neck范式,进一步提高模型平均精准度,降低模型的计算量。同时,采用多项式损失函数替换原二元交叉熵损失函数,以提高模型的精准度,并引入SIoU Loss消除真实框与预测框方向问题的缺陷,以提高训练速度和推理准确性。实验结果表明,模型在mAP0.5指标上达到95.7%,相较于基础YOLOv5模型,改进后的船舶目标检测模型mAP0.5提高0.9%、mAP0.95提高0.9%,同时FLOPs也降低2.1G。这一结果表明,雾环境下的船舶目标检测模型具有更优的精准度和更轻量的模型结构,在提高雾环境下船舶检测的精准度和效率方面具有很好的应用前景。

To effectively avoid ship collisions and overcome difficulty in ship identification,and low detection accuracy in foggy environments,this paper first builds a dataset for ship detection in foggy environments.Then,improvements are made on the YOLOv5 model.Specifically,the GSConv module is employed to replace the CBS module in the Head section to make the depth separable convolution closer to the separable convolution,improving model accuracy.The Slim-Neck paradigm is introduced to further boost the model’s average accuracy and reduce computational complexity.Additionally,the binary cross-entropy loss function is replaced with a polynomial loss function to enhance the model’s accuracy.The SIoU Loss is introduced to address the deficiency in direction between the real box and predicted box,thereby improving training speed and inference accuracy.Our experimental results show the model reaches 95.7%in mAP0.5 score,0.9%higher than that of the baseline YOLOv5 model while the FLOPs is down by 2.1G.Our study demonstrates the fog-based ship detection model achievesa better accuracy and has a lighter model structure and thus it has great potentials for application in improving the accuracy and efficiency of ship detection in foggy environments.