基于轻量化YOLOv7算法的侧扫声纳图像沉船检测
Side-scan sonar image shipwreck detection based on lightweight YOLOv7 algorithm
王胜平 1刘娉婷 1陈晓红 2陈志高1
作者信息
- 1. 东华理工大学 测绘与空间信息工程学院,江西 南昌 330013
- 2. 交通运输部 长江上海航道处,上海 200010
- 折叠
摘要
针对现有的侧扫声纳图像水下沉船检测方法存在检测速度慢,传统的YOLOv5 算法存在的漏检的问题,提出基于轻量化YOLOv7 算法的水下沉船检测改进方法.首先,通过随机翻转、随机噪声等操作扩充沉船图像的样本数量;然后,引入迁移学习策略,将在COCO数据集上学习到的权重迁移到沉船检测的YOLOv7 网络中;其次,改进模型损失函数中惩罚项的计算方式,提升收敛速度;最后在YOLOv7 网络中引入FasterNet结构,减少模型的参数量和计算复杂度,降低模型对硬件的需求,达到轻量化模型的目的.实验结果表明,改进方法较原始YOLOv7 算法在类平均精度值(mAP值)上提升了 4.75%,检测速度也由原来的 0.0218 秒/帧提升到 0.0179 秒/帧,证明了改进方法的工程应用价值.
Abstract
For the existing side-scan sonar underwater shipwreck detection method,there are deficiencies in the detection speed and leakage detection in YOLOv5.This paper proposes an improved method for underwater wreck detection based on the lightweight YOLOv7 algorithm.First,the sampling numbers of shipwreck images are expanded by random flip,random noise and other operations.Second,a transfer learning strategy is introduced to transfer the weights learned on the COCO dataset to the YOLOv7 network for shipwreck detection.Third,the computation of the penalty term in the loss function of the model is improved to enhance the speed of convergence.Finally,a FasterNet structure is introduced into the YOLOv7 network,which reduces the number of parameters and the computational complexity of the model,and reduces the hardware requirement of the model to achieve lightweight model.The experimental results show that the improved method improves the class mean accuracy value(mAP value)by 4.75%compared with the original YOLOv7 algorithm,and the detection speed is also improved from 0.021 8 fps to 0.017 9 fps,which proves the value of the improved method in this paper for engineering applications.
关键词
侧扫声纳图像/沉船检测/YOLOv7算法/FasterNet结构/迁移学习Key words
side-scan sonar images/shipwreck detection/YOLOv7/FasterNet/transfer leaering引用本文复制引用
基金项目
国家自然科学基金(42266006)
自然资源部海洋环境探测技术与应用重点实验室开放基金(MESTA-2020-A002)
江西省重点研发计划(20212BBE53031)
出版年
2024
国家科技期刊平台
NSTL
万方数据