基于深度学习的海面目标检测方法研究  

Research on sea target detection methods based on deep learning

作　　者：李坤垚 郗润平[1,2,3] 连春艳 寇学锋 LI Kunyao;XI Runping;LIAN Chunyan;KOU Xuefeng(School of Computer Science and Technology,Northwestern Polytechnical University,Xi’an 710129,China;National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean Big Data Application Technology,Xi’an 710129,China;Shaanxi Provincial Key Laboratory of Speech&Image Information Processing,Xi’an 710129,China;Shan’xi Baocheng Aviation Instrument C o.Ltd,Baoji 721000,China)

机构地区：[1]西北工业大学计算机学院,西安710129 [2]空天地海一体化大数据应用技术国家工程实验室,西安710129 [3]陕西省语音与图像信息处理重点实验室,西安710129 [4]陕西宝成航空仪表有限责任公司,宝鸡721000

出　　处：《中国体视学与图像分析》2019年第4期352-361,共10页Chinese Journal of Stereology and Image Analysis

基　　金：国家自然科学基金(No.61572405)。

摘　　要：海面目标的检测和识别在军用和民用方面都有着重要的意义,传统的目标检测方法在海面目标检测和识别中受到尾浪、海杂波等问题影响,容易出现虚警和漏检的情况。本文主要研究通过深度学习的方法对海面目标进行检测与识别。本文首先分析并讨论了深度学习在目标检测中的应用及在海面目标检测中的可行性,并构建了海面目标图像数据集,然后研究基于回归思想的YOLO v2网络模型,用于对海面目标的定位和识别,实现了海面目标实时检测,最后提出了通过在线难例挖掘和Faster R-CNN网络相结合的方法对海面目标进行检测,解决正负样本不均衡及复杂样本检测效果差的问题。实验结果表明该方法能够有效地对海面目标进行检测且效果优于传统方法。The detection and identification of sea targets are of great importance in both military and civil-ian applications.Traditional object detection methods are slow and affected by tail waves,sea clutter,and etc.They are prone to give false alarms or miss inspections.This paper mainly studies the detection and recognition of sea surface targets by deep learning methods.In this paper,we firstly analyze and discuss the application of deep learning in object detection and its feasibility in detecting sea targets.The n,we study YOLO v2 based on regression ideas.We construct an image dataset of sea surface targets.The network model is used to locate and identify targets on sea surface in real-time.We also study the method of combining the online hard example mining and the Faster R-CNN network to solve the problems of imbalanced positive/negative samples and the problem of poor detection of complex samples in detecting sea surface targets.Experiment results show that our methods can effectively detect sea surface targets and the effet is better than traditional methods.

关 键 词：海面目标检测 无人机图像 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]