摘要
针对水下低质量成像、水下目标形态大小各异、以及水下目标重叠或遮挡导致水下目标检测精度低的问题,提出一种结合数据增强和改进YOLOv4(you look only once)的水下目标检测算法,在YOLOv4的主干特征提取网络CSPDarknet53中添加卷积块注意力机制(convolutional block attention module,CBAM),以提高网络模型特征提取能力;在路径聚合网络(path aggregation network,PANet)中添加同层跳接和跨层跳接结构,以增强网络模型多尺度特征融合能力;通过数据增强方法PredMix(prediction-mix)模拟水下生物重叠、遮挡等显示不完全的情形,以增强网络模型鲁棒性。实验结果表明,结合数据增强和改进YOLOv4的水下目标检测算法在URPC2018(underwater robot picking control 2018)数据集上的检测精度提升到了78.39%,比YOLOv4高出7.03%,充分证明所提算法的有效性。
Aiming at the problem of low underwater object detection accuracy caused by low-quality underwater imaging,different shapes or sizes of underwater objects,and overlapping or occlusion of underwater objects,an underwater object detection algorithm combining data enhancement and improved YOLOv4 is proposed.By adding CBAM(convolutional block attention module)to the backbone of YOLOv4—CSPDarknet53,the feature extraction ability of network model is improved.In order to enhance the multi-scale feature fusion ability,the same-layer skip connections and cross-layer skip connections are added to PANet(path aggregation network).To enhance the robustness of the network model,the data enhancement method PredMix(prediction mix)is used to simulate the incomplete display of underwater organisms such as overlap or occlusion.The experimental results show that the detection accuracy of the underwater object detection algorithm combining data enhancement and improved YOLOv4 on URPC2018 dataset is improved to 78.39%,7.03%higher than YOLOv4,which fully proves the effectiveness of the proposed algorithm.
作者
史朋飞
韩松
倪建军
杨鑫
Shi Pengfei;Han Song;Ni Jianjun;Yang Xin(College of Internet of Things Engineering,Hohai University,Changzhou 213022,China)
出处
《电子测量与仪器学报》
CSCD
北大核心
2022年第3期113-121,共9页
Journal of Electronic Measurement and Instrumentation
基金
国家自然科学基金(61801169,61873086)
中央高校基本科研业务费(B220202020)项目资助
作者简介
史朋飞,2008年于南京信息工程大学获得学士学位,2011年于河海大学获得硕士学位,2016年于河海大学获得博士学位,现为河海大学副教授,主要研究方向为水下探测与成像、信息获取与处理、机器视觉等。E⁃mail:shipf@hhu.edu.cn;通信作者:倪建军,1999年于中国矿业大学获得学士学位,2002年于中国矿业大学获得硕士学位,2005年于中国矿业大学获得博士学位,现为河海大学教授,主要研究方向为机器学习、模式识别、机器人等。E⁃mail:njjhhuc@gmail.com
