基于YOLOv5的光学遥感图像舰船目标检测算法  

Ship target detection algorithm of optical remote sensing image based on YOLOv5

作　　者：成倩 李佳 杜娟[1] CHENG Qian;LI Jia;DU Juan(Fundamentals Department,Air Force Engineering University,Xi’an 710038,China)

机构地区：[1]空军工程大学基础部,陕西西安710038

出　　处：《系统工程与电子技术》2023年第5期1270-1276,共7页Systems Engineering and Electronics

基　　金：国家自然科学基金(62105258);空军工程大学基础部科研基金(JK2022106)资助课题。

摘　　要：针对YOLO(you only look once)v5算法在应用于光学遥感图像舰船目标检测任务时所面临的小目标误检率、漏检率较高的情况,提出一种基于YOLOv5改进的光学遥感图像舰船目标检测方法。首先对路径聚合网络结构进行改进,设计语义信息增强模块提取更富语义信息的浅层特征,增强对小目标特征的表达能力;然后使用Swish函数作为激活函数,提高网络对数据非线性特征的表征能力,加快模型的收敛速度;最后针对舰船目标的尺寸特点优化检测端结构,移除大目标检测头以减少推理计算量。测试集上的实验表明,该方法相较改进前将检测精度提高了5.2%且推理时间有所减少,在保证检测实时性的同时增强了模型的小目标辨别能力。When applied to the ship target detection task of optical remote sensing images,you only look once(YOLO)v5 algorithm suffered false positive and false negative a lot on small targets.An improved method based on YOLOv5 is proposed to deal with this situation.Firstly,a module named semantic information enhancement module aiming at extracting shallow features with more semantic information is designed and added into the path aggregation network,so as to enhance the expression ability of small target features.Then,the Swish function is used as the activation function to improve the network’s ability to characterize the nonlinear characteristics of the data and accelerate the convergence speed of the model.Finally,according to the size characteristics of the ship,the large target detection head is removed to reduce the amount of inference calculation and optimize the effect of the detection end.The experiment on the test set shows that compared with the method before the improvement,the proposed method improves the detection accuracy by 5.2%and the inference time is reduced.At the same time of ensuring the real-time detection,the small target discrimination ability of the model is increased.

关 键 词：光学遥感图像 舰船目标检测 YOLOv5 语义信息增强 激活函数 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]