基于CReLU和FPN改进的SSD舰船目标检测  

Ship object detection based on SSD improved with CReLU and FPN

作　　者：李晖晖[1] 周康鹏 韩太初 Li Huihui;Zhou Kangpeng;Han Taichu(I.School of Automation,Northwestern Polytechnical University,Xi’an 710072,China;Chinese Flight Test Establishment,Xi'an 710089,China)

机构地区：[1]西北工业大学自动化学院,西安710072 [2]中国飞行试验研究院,西安710089

出　　处：《仪器仪表学报》2020年第4期183-190,共8页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金(61333017)项目资助。

摘　　要：在遥感图像中,舰船目标具有目标尺寸较小、形状细长、多个目标紧密排列、类间相似度高等特点,现有的深度学习目标检测算法对舰船小目标的检测精度不高,易发生错检、漏检情况。为了更有效地利用遥感图像信息,提高小目标检测精度,构建了舰船数据集SDNGV,提出基于串行修正线性单元CReLU和特征金字塔网络(FPN)改进的单射探测器(SSD)舰船目标检测识别方法。首先,在SSD网络的浅层添加CReLU,提升其浅层特征的传递效率;然后,采用FPN从网络的深层到浅层逐级融合SSD中用于检测的多尺度特征图,提升网络的定位精度和分类精度。实验表明,所提目标检测算法具有较好的检测精度,改进方法具有明显的效果,在舰船小目标的检测上有10%的检测精度提升。In remote sensing images, ship objects have the characteristics of small size, slender shape, close arrangement of multiple objects and high similarity between classes. The existing deep learning object detection algorithms have low detection accuracy for small ship objects, and are prone to error detections and missed detections. In order to effectively utilize the remote sensing image information and improve the accuracy of small object detection, the SDNGV ship data set is constructed, and an improved single short multiBOX detector(SSD) ship object detection and recognition method based on concatenated rectified linear unit(CReLU) and feature pyramid networks(FPN) is proposed. Firstly, CReLU is added to the shallow layer of the SSD network to improve the transmission efficiency of its shallow layer features. Secondly, FPN is used to fuse the multi-scale feature map used for detection in SSD step by step from the deep layer to the shallow layer of the network to improve the positioning accuracy and classification accuracy of the network. Experiments demonstrate that the proposed detection algorithm has good detection accuracy, the improved method has obvious effect, and the detection accuracy of small ship objects has 10 percent improvement.

关 键 词：目标检测 舰船检测 深度学习 卷积神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TH39[自动化与计算机技术—计算机科学与技术]