基于改进YOLOv3的遥感小目标检测网络  

Remote Sensing Small Target Detection Network Based on Improved YOLOv3

作　　者：陈成琳 鲍春 曹杰[1,2] 郝群[1,2] CHEN Cheng-lin;BAO Chun;CAO Jie;HAO Qun(School of Optics and Photonics,Beijing Institute of Technology,Beijing 100081,China;Yangtze Delta Region Academy,Beijing Institute of Technology,Jiaxing Zhejiang 314003,China)

机构地区：[1]北京理工大学光电学院,北京100081 [2]北京理工大学长三角研究院(嘉兴),浙江嘉兴314003

出　　处：《计算机仿真》2023年第8期30-35,共6页Computer Simulation

基　　金：北京理工大学长三角研究院(嘉兴)研究生交叉创新专项计划(GIIP2021-005);基础加强技术领域基金(2019-JCJQ-JJ-273);北京市自然科学基金(4222017)。

摘　　要：目标检测在遥感图像解译中起着至关重要的作用,但由于遥感影像场景的复杂性以及拥有大量小而密集的、杂乱和旋转的目标使得检测难度大大增加。针对YOLOv3目标检测算法在遥感图像小目标检测方面精度较低的缺点,提出一种改进的YOLOv3检测网络,首先引入注意力机制,设计了一种新的特征提取模块,融合背景感知,强调小目标特征;其次对YOLOv3网络结构进行优化,增加了小目标检测层且优化了损失函数,进一步解决小目标识别精度低的问题;最后在大型光学遥感数据集DIOR和NWPU VHR-10上展开实验研究。实验结果表明,改进的模型与YOLOv3相比mAP提升8%~14%,且泛化性较好,检测速度达66FPS,满足实时性要求。Object detection plays a significant role in remote sensing image interpretation.However,due to the complexity of remote sensing image scenes and a large number of small,dense,chaotic and rotating targets,the difficulty of detection is greatly increased.Aiming at the low accuracy of the YOLOv3 object detection algorithm in small target detection,an improved YOLOv3 detection network is proposed.Firstly,the idea of attention mechanism is introduced,and a new feature extraction module is designed,which integrates background perception and emphasizes small target features.Secondly,the structure of YOLOv3 is optimized and the small target detection layer is added to further solve the issue of low accuracy of small target recognition.Finally,experiments were carried out on large optical remote sensing data sets DIOR and NWPU VHR-10.The experimental results show that compared with YOLOv3,the mAP@0.5 of the improved model is improved by 8% ~ 14%,and the generalization is better.The detection speed is 66 FPS,satisfying the real-time requirements.

关 键 词：深度学习 小目标检测 遥感影像 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]