基于判别性字典学习和稀疏表示的SAR地面目标识别

发布时间：2019-03-28 12:10

【摘要】：合成孔径雷达(Synthetic Aperture Radar,SAR)成像技术不受到距离、光照、天气等因素的影响,具备全天时、全天候、高分辨率成像的特点。目前,SAR成像技术已经被广泛应用于军事探测以及民用测绘的多个领域。但是,在SAR目标自动识别领域,由于SAR图像的相干成像原理,致使SAR图像被大量复杂的相干斑噪声污染,并且由于对SAR辐射特性以及成像原理相关知识的缺乏,导致提取有效的SAR目标特征进行目标识别十分困难,从而制约了SAR自动目标识别技术的发展。基于此,本文绕过特征提取这一难题,利用训练样本的类别信息,通过学习判别性字典来对样本进行稀疏编码,利用编码系数进行SAR目标识别,最终的实验数据证明了此类方法在SAR目标识别上具有可行性,且识别效果可观。本文的主要工作内容可概括为以下几点:(1)针对SAR图像的特性,进行了相关的预处理工作,主要涉及了噪声抑制、目标检测等相关算法的研究。(2)在现有的Fisher判别性字典学习方法的基础上对目标函数进行了改进,并将该方法成功应用在了SAR目标识别上,并取得了一定的识别效果。(3)为了验证本文改进的算法在SAR目标识别上的性能,本文将该算法与现有的经典稀疏分类方法SRC、K-SVD、D-KSVD、以及基于Fisher判别性字典学习方法的目标识别方法进行了对比,并总结了本文改进算法在SAR目标识别上的特性。(4)通过在不同大小训练样本集的实验,分析了字典大小对识别效果的影响,字典越大,本文算法识别率越高;另外,实验结果验证了当训练样本各子类数量差异较大时,数量较少的子类对应的识别率降低。
[Abstract]:Synthetic Aperture Radar (Synthetic Aperture Radar,SAR) imaging technology is not affected by range, illumination, weather and other factors. It has the characteristics of all-weather, all-weather and high-resolution imaging. At present, SAR imaging technology has been widely used in many fields of military detection and civil mapping. However, in the field of automatic recognition of SAR targets, due to the coherent imaging principle of SAR images, SAR images are polluted by a large number of complex speckle noises, and the lack of knowledge about SAR radiation characteristics and imaging principles is also involved. It is very difficult to extract effective SAR target features for target recognition, which restricts the development of SAR automatic target recognition technology. Based on this, this paper bypasses the difficult problem of feature extraction, makes use of the classification information of the training samples, uses the learning discriminative dictionary to carry out sparse coding of the samples, and uses the coding coefficients to recognize the SAR targets. The experimental results show that this method is feasible and effective in SAR target recognition. The main contents of this paper can be summarized as follows: (1) according to the characteristics of SAR images, the related pretreatment work is carried out, which mainly involves noise suppression. Research on related algorithms such as target detection. (2) based on the existing Fisher discriminative dictionary learning method, the objective function is improved, and the method is successfully applied to SAR target recognition. Some recognition results have been obtained. (3) in order to verify the performance of the improved algorithm in SAR target recognition, this paper compares this algorithm with the classical sparse classification method SRC,K-SVD,D-KSVD,. The target recognition methods based on Fisher discriminative dictionary learning method are compared, and the characteristics of the improved algorithm in SAR target recognition are summarized. (4) the experiment of training sample sets in different sizes is carried out. The influence of dictionary size on the recognition effect is analyzed. The larger the dictionary, the higher the recognition rate of this algorithm. In addition, the experimental results show that when the number of sub-classes of training samples varies greatly, the recognition rate of the smaller number of subclasses decreases.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN958

【参考文献】