基于稀疏理论的SAR图像目标识别研究

发布时间：2018-02-23 18:29

本文关键词： 合成孔径雷达自动目标识别稀疏性特征提取结构化稀疏联合稀疏深度置信网络卷积神经网络　出处：《西安电子科技大学》2015年博士论文　论文类型：学位论文

【摘要】：合成孔径雷达(SAR)是一种重要的对地观测手段。利用SAR图像进行目标识别在战场环境中具有非常重要的意义。本论文围绕基于稀疏理论的SAR图像目标识别进行了深入的研究：重点针对存在遮挡情况下的SAR目标识别问题和从数据中自动进行特征提取的问题。各部分的主要内容概括如下：第一部分,针对合成孔径雷达(SAR,)图像目标识别中存在物体遮挡的情况,提出一种基于非负稀疏表示的分类方法。通过分析L0范数和L1范数最小化在求解非负稀疏表示问题上的区别,证明在一定条件下,Ll范数最小化方法除了保持解的稀疏性还能得到与输入信号更加相似的原子集合,因此也更加适用于分类问题中。在运动和静止目标获取与识别(MSTAR)数据集上的识别实验结果表明,采用L1范数的非负稀疏表示分类方法能达到较好的识别性能,并且相对传统方法对存在遮挡情况下的识别问题更稳健。第二部分,虽然基于稀疏表示或者非负稀疏表示的模型在遮挡情况下的目标识别表现出一定的稳健性能,但是其模型中对于遮挡部分采用的都是像素级假设,即：遮挡效应引起的是目标图像上少量像素点的变化,并且假定受影响的像素点在空间位置上是独立出现的。然而真实情况是,对于尺度小于成像单元大小的遮挡物,其遮挡效应在目标图像上的表现通常与相干斑造成的效果相近。而对于尺度大于成像单元大小的遮挡物,其使得目标图像上发生强度变化的区域将会是一块连续的区域。因此我们可以利用遮挡效应的这一结构化特点,对遮挡部分进行单独建模,提出了结构化稀疏遮挡模型。该模型尝试将测试数据中的遮挡部分以及在训练样本集上的稀疏表示部分分离开来。在识别时,仅仅通过稀疏表示部分进行目标分类,从而避免了遮挡的影响。仿真实验表明,基于结构化稀疏遮挡模型的方法不仅对于遮挡区域的大小,形状,块数以及散射起伏都具有较好的稳健性。第三部分,地面目标的SAR图像中除了包含目标散射回波形成的区域,还包括由目标遮挡地面形成的阴影区域。但是由于这两种区域中的图像特性不相同,所以传统的SAR图像自动目标识别主要利用目标区域信息进行目标识别,或者单独使用阴影区域进行识别。该文提出一种阴影区域与目标区域图像联合的稀疏表示模型。通过使用LI\L2范数最小化方法求解该模型得到联合的稀疏表示,然后根据联合重构误差最小准则进行SAR图像目标识别。在运动和静止目标获取与识别(MSTAR)数据集上的识别实验结果表明,通过联合稀疏表示模型可以有效的将目标区域与阴影区域信息进行融合,相对于采用单独区域图像的稀疏表示识别方法性能更好。第四部分,对于传统的字典学习方法,如KSVD,其目标函数是最小化重构误差。本章通过在字典学习目标函数中增加对稀疏表示系数之间的相似性约束得到具有判决能力的字典。该约束使得不同类样本的稀疏表示间的相似性趋向于0,即最不相似。因此在稀疏表示系数这样的特征空间中不同类别之间的差异更大,更容易找到好的分类面。实验显示,添加了相似性约束后学习到的字典及其稀疏表示相比传统的字典学习方法可以更好的区分各类SAR目标。第五部分,特征提取是合成孔径雷达(SAR)图像目标识别的关键环节。SAR图像中存在的相干斑点和非光滑特性使得传统针对光学图像的特征提取方法变得很难应用。虽然可以采用深度置信网络(Deep belief network, DBN)自动地进行特征学习,但是该方法属于无监督学习方法,这使得学习到的特征与具体的任务是无关的。本文提出了一种叫做相似性约束的限制玻尔兹曼机模型。该模型在学习过程中通过约束特征向量之间的相似性达到引入监督信息的目的。另外,可以将多个相似性约束的限制玻尔兹曼机堆叠成一种新的深度模型,我们称其为相似性约束的深度置信网络模型。实验结果表明在SAR图像目标识别应用中,本文方法相比主成分分析(PCA)以及原始DBN具有更好的识别性能。第六部分,针对现实SAR ATR应该解决的几个问题：需要具备目标平移不变性,对于相干斑噪声随机性不敏感以及能够容忍训练数据集中一定程度的姿态图像缺失,首先研究了通过已有图像合成未知方位角下的图像以弥补训练集中姿态图像缺失的可能性。受稀疏表示模型启发下,提出了一种姿态图像合成模型。实验显示,通过姿态图像合成的方式可以有效的提升识别性能。随后通过姿态图像合成扩充、平移扩充以及相干斑加噪扩充的方式增大训练样本集合,并通过扩充后的数据对卷积神经网络进行训练学习。大量实验表明,该方法训练得到的模型可以有效的克服测试样本中出现的平移问题、相干斑变化以及姿态变化的情况。
[Abstract]:Synthetic aperture radar (SAR) is one of the important means of earth observation. Using SAR image target recognition is very important in the battlefield environment. SAR image target recognition based on sparse theory is studied in this thesis: focusing on the existing block SAR target recognition problems and from the data in the automatic feature extraction problems. The main contents of each part are as follows: in the first part, the synthetic aperture radar (SAR) objects occlusion image target recognition, a non negative sparse representation classification method based on. Through the analysis of the differences on the issues that the L0 norm and L1 norm minimization in to solve the non negative sparse, prove that under certain conditions, the Ll norm minimization method in addition to maintain the sparsity of the solution can get with the input signal is more similar to atomic collection, so it is more suitable In the classification problem. On a moving and stationary target acquisition and recognition (MSTAR) recognition experiments on the data set. The results show that the non negative sparse representation classification using L1 norm method can achieve better recognition performance, and compared with the traditional method of occlusion identification problem under the condition of more robust. The second part, although based on sparse representation or a non negative sparse representation model under occlusion target recognition showed robust performance, but the model for the occlusion are based on pixel level assumptions, namely: shielding effect caused by the small target image pixels change, and assume that pixels affected are independent in appearance the space position. But the reality is that the scale is less than the size of the imaging unit covering, the shielding effect in the target image usually causes and speckle effect is similar While the scale is larger than the size of the imaging unit covering the target image, regional intensity changes will be a continuous area. So we can use the structural features of the blockage of the occlusion were modeled separately, put forward a structured sparse block model. This model attempts to test sparse occlusion in the data in the training sample set and the said part separated. In recognition, only through the sparse representation of part of the target classification, so as to avoid the occlusion effect. Simulation results show that the method of structured sparse block model based on not only for the occlusion area size, shape, number and scattering fluctuations are better the robustness of SAR. In the third part, the image of ground targets in addition to containing the target scattering echo area, including a ground target occlusion form The shaded area. But because of the image characteristics of the two regions are not the same, so SAR image automatic target recognition using the traditional main target area information for target recognition, or used alone shadow region recognition. This paper proposes a shadow region and object region image joint sparse representation model. By using the LIL2 norm minimization the method of solving the model combined with the sparse representation, then SAR image target recognition based on joint reconstruction error criterion. On a moving and stationary target acquisition and recognition (MSTAR) recognition experiment data sets. The results show that the joint sparse representation model can effectively integrate the target region and the shadow region information, compared with the sparse a separate regional representation of the image recognition method of performance better. In the fourth part, for the traditional dictionary learning methods, such as KSVD, its target Function is the reconstruction error minimization. This chapter increased by learning objective function in the dictionary that similarity coefficient between the constraint has the ability of judgment. This makes the dictionary sparse constraint sparse samples of different classes of similarity between said tends to 0, is the least similar. So in the sparse feature space coefficient in such the difference between different categories of the larger, more easy to find good classification. Experimental results show that adding the dictionary and sparse similarity constraints to learn that compared to the traditional dictionary learning method can better distinguish between various types of SAR. The fifth part, feature extraction is a synthetic aperture radar (SAR) image target key.SAR the identification of existing speckle and non smooth characteristics of the traditional feature extraction method for optical image becomes very difficult. Although the application can use the deep belief network (Deep belief network, DBN) automatically feature learning, but this method belongs to unsupervised learning method, which makes the features learned with a particular task is irrelevant. This paper presents a restricted Boltzmann machine model called similarity constraints. The model through the similarity between feature vectors to the introduction of supervision and restraint the purpose of information in the learning process. In addition, the stack limit Boltzmann machine multiple similarity constraints into a new depth model, we call it the similarity constraint deep belief network model. The experimental results show that the SAR image target recognition application, this method compared with principal component analysis (PCA) and the original DBN has better recognition performance. In the sixth part, several problems should be solved according to the reality of SAR ATR: need to have the goal of translation invariance, for speckle noise with the machine is not sensitive Sense and can tolerate attitude of image missing training data to a certain extent, first study the existing image by image synthesis of unknown azimuth to compensate for the possibility of the training set. The lack of motion image sparse representation model inspired, presents a synthesis model of attitude image. Experimental results show that can enhance the recognition performance through effective way the attitude of image synthesis. Then through gesture image synthesis expansion, speckle noise and translational expansion expansion increases the training set, and through the expansion of the data after the convolution neural network training. The experiments show that the method obtained from the training model can effectively overcome the problem of translation appeared in the test sample. Speckle variation and the change of attitude.

【学位授予单位】：西安电子科技大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TN957.52

【参考文献】