改进YOLOX-S的红外舰船目标检测算法  

Infrared ship target detection algorithm based on improved YOLOX-S

作　　者：娄树理 王岩[1] 郭建勤[2] 公维锋 LOU Shuli;WANG Yan;GUO Jianqin;GONG Weifeng(School of Physics and Electronic Information,Yantai University,Yantai 264005,China;Department of Electronics and Communication Engineering,Shandong College of Electronic Technology,Jinan 250200,China;State Key Laboratory of High-end Server and Storage Technology,Jinan 250101,China)

机构地区：[1]烟台大学物理与电子信息学院,山东烟台264005 [2]山东电子职业技术学院电子与通信工程系,山东济南250200 [3]高效能服务器和存储技术国家重点实验室,山东济南250101

出　　处：《应用光学》2023年第5期1054-1060,共7页Journal of Applied Optics

基　　金：山东省自然科学基金(ZR2019LZH016)。

摘　　要：红外舰船目标检测与识别技术是反舰导弹红外成像制导的关键,对于武器装备制导性能具有重大意义。针对在复杂环境下红外舰船目标检测的精度和速度问题,提出了改进YOLOX-S的红外舰船目标检测算法。首先引入深度可分离卷积(depthwise separable convolution)代替FPN(feature pyramid network)及YOLOHead残差结构中的传统卷积,降低模型的参数量;其次引入ECANet通道注意力机制,提高网络的注意力,降低舰船目标的虚检率和漏检率;最后使用CIoU损失函数,进一步提高网络的检测准确率。实验表明,改进后算法的检测平均精度(AP)达到98%,检测速度为56帧/s,对比改进前YOLOX-S算法,检测速度与平均精度分别提升6帧/s和3%,且模型更加轻量化。实验结果充分证明本文提出的算法能够有效完成红外舰船目标检测任务。Infrared ship target detection and recognition technology is the key technology of infrared imaging guidance for anti-ship missile,which is of great significance to the guidance performance of weapon equipment.For the accuracy and speed in complex environment,an infrared ship target detection algorithm based on YOLOX-S was proposed.Firstly,the depthwise separable convolution(DSC)was introduced to replace the traditional convolution in the feature pyramid network(FPN)and YOLOHead residual structures,which could reduce the amount of parameters of the model.Secondly,the ECANet channel attention mechanism was introduced to improve the attention of network,which could reduce the false detection rate and missed detection rate of ship targets.Finally,the CIoU loss function was used to further improve the detection accuracy of the network.The experimental results show that the average precision(AP)of the optimized algorithm reaches 98%and the detection speed is 56 frame/s.Compared with the previous YOLOX-S algorithm,the detection speed and the average precision are improved by 6 frame/s and 3%,respectively,and the model is more lightweight.The experimental results fully prove that the proposed algorithm can effectively complete the infrared ship target detection task.

关 键 词：红外舰船 目标检测 注意力机制 损失函数 

分 类 号：TN219[电子电信—物理电子学] TP391[自动化与计算机技术—计算机应用技术]