一种改进YOLOv8的水下声呐图像目标检测方法  

An improved target detection method for YOLOv8 underwater sonar images

作　　者：刘凡诚 邢传玺[1,2] 魏光春 崔晶 董赛蒙 LIU Fancheng;XING Chuanxi;WEI Guangchun;CUI Jing;DONG Saimeng(School of Electrical and Information Technology,Yunnan Minzu University,Kunming 650500,China;Yunnan Key Laboratory of Unmanned Autonomous System,Kunming 650500,China)

机构地区：[1]云南民族大学电气信息工程学院,云南昆明650500 [2]云南省无人自主系统重点实验室,云南昆明650500

出　　处：《应用科技》2025年第1期34-40,共7页Applied Science and Technology

基　　金：国家自然科学基金项目(61761048);云南省基础研究专项面上项目(202101AT070132).

摘　　要：为解决水下声呐图像中目标形状小、信息少等识别精度低带来的漏检、误检问题,本文提出一种改进YOLOv8水下声呐图像目标检测方法(YOLOv8-Underwater Sonar Image,YOLOv8-USI)。首先对水下声呐图像进行图像增强、图像降噪等预处理,并利用生成对抗网络对水下声呐图像训练集进行扩充,防止过拟合现象;其次,引入GhostNet模块解决YOLOv8网络结构参数量多的问题,从而提高水下目标识别速度;接着根据预处理后声呐图像的特征,提取水下声呐图像中的目标特征信息。最后,根据识别到的目标物体置信度,验证声呐图像中目标物体的漏检与误检情况。实验结果表明,输出结果图的目标识别效果与整个检测过程速度均有所提高,时间加快0.08 s,因此YOLOv8-USI网络结构可有效提高水下声呐图像目标检测精度与速度。In order to solve the problems of missed detection and false detection caused by low recognition accuracy such as small shape and little information in underwater sonar images,this study proposed an improved target detection method for underwater sonar images of YOLOv8(YOLOv8-USI).Firstly,the underwater sonar image was preprocessed with image enhancement and image noise reduction,and the training set of underwater sonar images was expanded by using the generative adversarial network,so as to prevent overfitting phenomenon.Secondly,the GhostNet module was introduced to solve the problem of a large number of parameters in the structure of YOLOv8 network,so as to increase the speed of underwater target recognition.And then the target feature information in the underwater sonar image was extracted according to the features of the preprocessed sonar image.Finally,according to the confidence level of the identified target object,the missed detection and false detection of the target object in the sonar image were verified.Experimental results showed that the target recognition effect of the final output map was improved and the speed of the whole detection process was increased with 0.08 second,proving that the YOLOv8-USI network structure can effectively improve the accuracy and speed in the target detection of underwater sonar images.

关 键 词：侧扫声呐图像 图像降噪 目标检测 YOLOv8-USI 过拟合 数据增强 生成对抗网络 GhostNet模块 

分 类 号：TB566[交通运输工程—水声工程]