基于稀疏表示的SAR图像目标识别研究

发布时间：2018-04-20 13:07

本文选题：SAR目标识别 + 稀疏表示　；参考：《电子科技大学》2014年硕士论文

【摘要】：SAR图像目标识别作为信息获取的关键技术,具有重要的应用价值,一直是国内外目标识别领域的研究热点。近年来稀疏表示理论被广泛应用于各类图像处理领域,并且在人脸识别中已取得了良好的效果。本文着重研究了将稀疏表示理论应用于SAR图像目标识别中的两个关键步骤:冗余字典的构造和稀疏系数的求解。主要研究内容如下:(1)针对原始冗余字典类别差异性不足和规模较大的两个缺陷,利用冗余字典的二维结构提出了字典的纵、横双向改进方法。在纵向改进中,针对SAR图像由确定信息和不确定信息组成的特点,利用小波变换来提取有利于识别的低频确定信息,并经过2DPCA降维处理得到小波域字典。在横向改进中,利用K-近邻算法的样本选择思想,实现了字典原子的横向动态筛选,从而生成基于近邻子空间的动态字典。(2)在完成了冗余字典构造的基础上,对稀疏系数分解算法进行研究。将最小L1范数凸优化算法和OMP算法进行对比分析,验证了后者的类别差异性和分解效率均优于前者。同时,针对OMP算法稀疏度K未知的问题,提出了用类别统计量C来替换稀疏度K作为算法迭代终止条件的改进方法,并通过仿真实验验证了改进后的OMP算法具有更好的识别效果。(3)根据稀疏分解系数的分布特点,总结出最大系数准则、归类系数最大准则两种分类判别准则,并对这两种准则进行仿真对比。仿真结果表明,归类系数最大准则能够取得更高的识别率,故本文利用它来完成分类识别器的设计。(4)基于MSTAR数据库,统计了本文识别算法在各种非理想情况下的识别率,验证了在含有噪声、遮挡及分辨率下降情况的算法鲁棒性。
[Abstract]:As a key technology of information acquisition, SAR image target recognition has important application value and has been a hot research topic in the field of target recognition at home and abroad. In recent years, sparse representation theory has been widely used in various image processing fields, and has achieved good results in face recognition. This paper focuses on two key steps of applying sparse representation theory to target recognition in SAR images: the construction of redundant dictionaries and the solution of sparse coefficients. The main research contents are as follows: (1) aiming at the deficiency of the difference of the original redundant dictionaries and the two defects of large scale, this paper puts forward the vertical and horizontal bidirectional improvement methods of the redundant dictionaries by using the two-dimensional structure of the redundant dictionaries. In the longitudinal improvement, according to the characteristic that SAR image is composed of determinate information and uncertain information, wavelet transform is used to extract the low-frequency deterministic information which is favorable to recognition, and the dictionary in wavelet domain is obtained by 2DPCA dimensionality reduction. In the lateral improvement, by using the idea of sample selection of K- nearest neighbor algorithm, we realize the horizontal dynamic selection of dictionary atoms, so as to generate a dynamic dictionary based on nearest neighbor subspace. (2) based on the construction of redundant dictionary, the redundant dictionary is constructed. The sparse coefficient decomposition algorithm is studied. By comparing the minimum L1 norm convex optimization algorithm with the OMP algorithm, it is verified that the class difference and decomposition efficiency of the latter are better than the former. At the same time, aiming at the problem that the sparsity K of OMP algorithm is unknown, an improved method is proposed to replace the sparse degree K with class statistic C as the iterative termination condition of the algorithm. The simulation results show that the improved OMP algorithm has better recognition effect. According to the distribution characteristics of sparse decomposition coefficient, the maximum coefficient criterion and the maximum classification coefficient criterion are summarized. The two criteria are simulated and compared. The simulation results show that the maximum criterion of classification coefficient can achieve a higher recognition rate, so this paper uses it to complete the design of the classifier. (4) based on the MSTAR database, the recognition rate of the algorithm in this paper is calculated under various non-ideal conditions. The robustness of the algorithm with noise, occlusion and resolution reduction is verified.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN957.52

【参考文献】