基于改进YOLOv7的水下目标检测算法  

Underwater Object Detection Algorithm Based on Improved YOLOv7

作　　者：陈宁[1] 张武 王凯[2] 郑少秋 刘凡[1] CHEN Ning;ZHANG Wu;WANG Kai;ZHENG Shaoqiu;LIU Fan(College of Computer Science and Software,Hohai University,Nanjing 211100,China;Information System Requirement Key Laboratory of CETC,Nanjing 210023,China)

机构地区：[1]河海大学计算机与软件学院,南京211100 [2]中国电子科技集团公司信息系统需求重点实验室,南京210023

出　　处：《指挥信息系统与技术》2024年第5期77-83,共7页Command Information System and Technology

基　　金：江苏省水利科技计划(2021063);装备预研教育部联合基金(8091B032157);信息系统需求重点实验室开放基金(LHZZ2021-M04)资助项目。

摘　　要：水下目标检测已成为计算机视觉的热门研究领域之一,现有的水下场景数据集中的水下目标存在尺度分布跨度大且分布密集的问题。针对现有光学成像水下目标检测技术在挖掘、表征水下目标多尺度特征方面的局限性,提出了基于尺度序列特征金字塔和金字塔分割注意力机制的YOLOv7水下目标检测算法。首先,在YOLOv7算法基础上引入了一种新的尺度序列特征金字塔,以增强模型的多尺度特征提取能力;然后,引入一种高效的金字塔分割注意力机制,以提升模型提取细粒度多尺度空间信息的能力;最后,在2个公共数据集上进行了试验。试验结果表明,该算法在多种水体的不同目标上都能表现出较好的性能。Underwater object detection has become one of the hot research areas in computer vision.However,existing underwater scene datasets have issues such as a wide range of scale distributions and dense distribution of underwater targets.Addressing the limitations of existing optical imaging un⁃derwater object detection technologies in mining and characterizing multi-scale features of underwater objects,a YOLOv7 underwater object detection algorithm based on a scale-sequential feature pyramid and a pyramid partition attention mechanism.Firstly,a new scale-sequential feature pyramid is intro⁃duced on the basis of YOLOv7 to enhance the model's multi-scale feature extraction capabilities.Then,an efficient pyramid partition attention mechanism is introduced to improve the model's ability of extracting more fine-grained multi-scale spatial information.Finally,experiments conducted on two public datasets are carried out.Experiment result shows that the algorithm performs well on various underwater targets across different water bodies.

关 键 词：目标检测 多尺度特征提取 水下目标探测 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP181[自动化与计算机技术—计算机科学与技术]