【作者】 邱宇；

【导师】 邹斌；

【作者基本信息】 哈尔滨工业大学 ， 信息与通信工程， 2020， 硕士

【摘要】 目前,具有高分辨率SAR成像功能的卫星逐渐增多,这使得高分辨率和超高分辨率的SAR图像数据越来越丰富,依靠SAR卫星全天时、全天候的优势,针对SAR图像的解译研究也逐渐深入。在海洋运输高度发达、海洋周边环境逐步复杂、大型港口建设层出不穷的今天,对海面目标和港口内目标的监视和识别具有强烈的应用需求。对SAR图像的海面目标监测和识别成为了具有重要实际应用意义的研究课题,本文面向高分辨率SAR图像,主要进行了近海岸舰船目标的检测和分类研究,完成了对不同环境下船只的检测工作以及对船只大致类别的区分。文章的主要内容如下:首先,对国内外学者对舰船特性分析、SAR图像舰船目标检测和SAR图像舰船目标分类的研究现状进行了总结,明确了本文的研究方向和研究意义。接下来,本文对SAR图像舰船目标多种特征的提取进行了研究,其中,主要针对Pol SAR图像进行了特征提取的研究。本文还对SAR图像的船海特性进行了详尽的分析,主要研究了SAR图像中舰船目标的特点,针对不同的极化通道,研究了船海特性在不同极化条件下的差异,还对几种比较典型的舰船目标的几何特性进行了分析,为后续舰船目标的目视解译以及舰船目标的特征提取和类别区分奠定了基础。然后,本文在对舰船目标进行了极化特征、纹理特征的提取以及几何特性分析后,重点研究了如何有效利用大量的特征提取结果。本文利用Relief F算法对大量的舰船目标特征进行了衡量,筛选并构建出精简高效的舰船特征空间。利用精简过的特征空间完成了基于Relief F-SVM的舰船目标检测方法,并对单极化高分辨率SAR图像中的舰船进行了检测,还对检测结果结果进行了分析,比较了极化特征的引入对于舰船目标检测效果的提升。最后,在舰船目标检测结果的基础上,本文对检测目标进行了切片处理,构建了包含军船和商船两大类,细分为四个小类的数据集。然后利用深度卷积神经网络VGGNet结合模型迁移方法,完成了对本文构建数据集的分类,并采用传统的多类别SVM分类方法进行了对比实验,最终实现了舰船目标的高精度分类。更多还原

【Abstract】 At present,with an advantage of all-time and all-weather detection,satellites with high-resolution SAR imaging function are gradually increasing,with which provide abundant SAR image data with high and ultra-high resolution.Under the circumstances of heavily loaded marine transportation,the gradual complexity of the marine surrounding environment,and the emergence of large-scale seaport construction,a strong demand has been derived for the monitoring and identification of targets at sea and in ports,among which sea surface target monitoring and recognition for SAR images has become a practical research topic of crucial significance.This paper mainly focuses on high-resolution SAR images for the detection and classification of near-shore ship targets.The main content of the article is as follows:Firstly,the current research status of ship characteristics analysis has been summarized.SAR image ship target detection and SAR image ship target classification and recognition at home and abroad are summarized,which help to determine the research direction and significance of this paper.Next,feature extraction of ship targets has been analyzed in SAR images,among which Pol SAR images containing polarization informat ion are studied in particular.This paper also conducts a detailed analysis of the ship-sea characteristics of SAR images.The characteristics of ship targets in SAR images,the differences of ship-sea characteristics under different polarization conditions and the geometric characteristics of several typical ship targets were analyzed to lay the foundation for visual interpre tation of the ship,feature extraction and classification of ship targets.Then,after an extraction of polarization and texture features and geometric characteristics analysis of ship targets,effective utilization of massive feature extraction results has been studied.The Relief F algor ithm has been applied to weigh the generated ship target features,in order to construct a a simplified feature space.A ship target detection method based on Relief F-SVM has been inplemented on the simplified feature space.Ships have also been detected with single-polarized SAR image,which demonstrates the priority of the polarized features on detection accuracy.Finally,based on the results of ship target detection,this paper has sliced the detection targets and constructed a data set which contains four major categories,namely large military ships,small military ships,large commercial ships and small commercial ships.Deep convolutional neural network VGGNet and model migration have been exploited to complete the classification of the constructed data set,and traditional SVM classification method has been applied to make an comparison,which proves inferior to the proposed method with a high-precision classification result of ship targets.更多还原