基于YOLOv7的轻量化水下目标检测算法

A YOLOv7 Based Lightweight Underwater Target Detection Algorithm

水下目标检测在海洋科学、环境保护、资源开发、军事防御、文化遗产保护等领域具有重要意义。然而,水下环境复杂,水下图像质量较差和小生物聚集导致水下目标检测时出现漏检、误检等问题,需要提高检测精度;检测的实时性也需要设计更加快速的网络结构;水下设备的存储和计算能力有限,需要在保证准确性的同时保持较低的计算开销,为此提出了基于YOLOv7的改进型算法YOLOv7-PSS。首先,利用PConv卷积代替骨干网络中的部分普通卷积,减少算法的参数量,加快算法的训练与预测;然后,融合SE注意力机制,增强特征的提取能力;随后,利用SIoU损失函数加速网络收敛,优化算法训练过程。实验结果表明,在URPC2021水下目标检测数据集上,所提算法的mAP达到87.3%,比原算法提高了7.5%,而算法的参数量比原算法减少了11.9%,为水下设备的部署奠定了基础。

Underwater target detection is of great significance in marine science,environmental protection,resource development,military defense,cultural heritage protection and other fields.However,the complex underwater environment,poor underwater image quality and small biological aggregation may lead to missed detection and false detection in underwater target detection,so it is necessary to improve the detection accuracy.The real-time detection needs to design a faster network structure.Underwater devices have limited storage and computing power and need to maintain low computational overhead while ensuring accuracy.In view of these difficulties,an improved network YOLOv7-PSS is proposed based on YOLOv7.Firstly,PConv convolution is used to replace some ordinary convolutions in the backbone network to reduce parameter quantity of the model and speed up training and prediction of the model.Then,the SE attention mechanism is added to enhance the feature extraction ability,and SIoU loss function is adopted to accelerate network convergence and optimize model training process.Experimental results show that on the URPC2021 underwater target detection dataset,the proposed algorithm has a mAP of 87.3%,which is 7.5% higher than that of the original model,and the parameter quantity is reduced by 11.9%,which lays a foundation for the deployment of underwater equipment.