【作者】 罗磊；

【导师】 李跃华；

【作者基本信息】 南京理工大学 ， 信息与通信工程， 2010， 博士

【摘要】 鉴于毫米波探测系统相对微波探测系统和光学探测系统的独特优势,近年来在军事民用等领域得到了广泛应用和发展。随着探测精度的提高,系统对信号处理方法的要求也越来越高。流形学习是2000年出现的一种新的机器学习理论,旨在发现高维数据分布的内在规律,并从中恢复低维流形结构,实现维数约简。本文将流形学习方法应用于毫米波探测器目标识别,并对现有流形学习算法进行了改进和推广。论文的主要研究工作如下：为了减少噪声对目标识别的影响,研究了基于提升9-7小波的信号去噪及其实时实现。对去噪算法中所有相关参数作了近似处理,使其分母皆为2的整数次幂,算法只涉及整数的加法、乘法和移位运算。分析了算法实时实现对硬件平台的要求,在DSP构建的硬件处理平台上,实现了被动毫米波探测器信号的实时去噪。在特征提取方面,将典型的非线性流形学习算法应用于被动毫米波探测器及毫米波高分辨率雷达信号特征提取中,实验结果证明了流形学习算法的有效性。综合线性判别分析算法的优点,在邻域保持投影算法基础上引入了类间散布矩阵,得到了改进算法,邻域保持判别投影。通过在邻域保持投影算法中引入非相关约束,使提取的特征向量具有非相关性,减少冗余信息,得到了改进算法,非相关邻域保持投影。融合邻域保持判别投影及非相关邻域保持投影算法的优点,得到了非相关判别邻域保持投影算法。为了更好的应对非线性问题,通过加核的方法对非相关邻域保持投影和非相关判别邻域保持投影算法进行了非性线扩展,得到了改进算法,核非相关邻域保持投影和核非相关判别邻域保持投影。将本文所改进的算法应用于毫米波探测器目标识别,实验结果证明了算法的优越性能。基于非线性流形学习中局部线性嵌入算法的思想,提出了一种单类分类算法。此分类算法首先计算未知类别样本的重构系数,定义一种误差作为判别标准,根据此误差的大小判断样本的类别归属。将算法应用于被动毫米波探测器目标识别中,实验结果表明,相对目前流行的单类分类算法,具有更好的性能。相似地,基于局部线性嵌入算法的思想提出了一种多类分类算法。此算法考虑的是样本的重构误差及其近邻中非同类样本产生的误差,此误差反映的是样本与其所在低维流形之间的关系。将算法应用于毫米波高分辨率雷达一维距离像的目标识别,实验结果表明,算法能够有效地进行分类识别,与目前流行的多分类算法相比,分类效果较好,且参数估计简单,分类结果受参数影响较小,有效地提高了毫米波高分辨率雷达的探测精确度。更多还原

【Abstract】 Millimeter-wave detection system, widely used in military and civil fields, has many advantages in comparison with microwave detection system and infrared detection system. The better signal processing methods are desired along with the improvement of detection precision. Manifold learning proposed in 2000 is a new theory of machine learning, aiming to find the latent feature of high-dimensionality data, reconstruct the low-dimensionality manifold, and reduce the dimensionality. In this paper, a few manifold learning algorithms are improved and used in the target recognition of millimeter-wave (MMW) detector. The main contents of this paper are stated below.The signal real-time denoising is explored based on lifting 9-7 wavelet for reducing the influence of noises. All parameters of the denoising algorithm are approximated and the denominators are changed to integer power of 2. Then the algorithm only involves multiplication, addition and shift of integer. The demand of hardware is analyzed for realizing the real-time denoising. The signal processing system is constructed based on DSP, and the denoising of passive MMW detector signal is realized.The typical nonlinear manifold learning algorithms are applied to feature extraction of passive MMW detector signal and MMW high range resolution radar signal. And the experimental results show that the methods are adaptable to the signal from MMW detecting system. The improved algorithm, Neighborhood Preserving Discriminant Projections (NPDP), is proposed by generalizing the virtues of Neighborhood Preserving Projections (NPP) and Linear Discriminant Analysis (LDA) and introducing between-class scatter matrix. And the improved algorithm, Uncorrelated Neighborhood Preserving Projections (UNPP), is proposed by introducing an uncorrelated constraint which leads the feature vectors extracted to be uncorrelated and reduces the redundant information. Combining NPDP and UNPP, the improved algorithm, Uncorrelated Discriminant Neighborhood Preserving Projections (UDNPP), is proposed. For adaptation of nonlinear problem, UNPP and UDNPP are extended as Kernel Uncorrelated Neighborhood Preserving Projections (KUNPP) and Kernel Uncorrelated Discriminant Neighborhood Preserving Projections (KUDNPP) by kernel method. The proposed algorithms are used for target recognition of MMW detector and the experimental results indicate their good performance.A new one-class classification algorithm is proposed based on the idea of Locally Linear Embedding (LLE). This algorithm firstly computes the reconstruction weights of unknown samples. Then an error, on which the class of samples can be decided based, is defined as a criterion. The algorithm is applied to target recognition of passive MMW detector and the experimental results indicate its good performance in comparison with current popular one-class classification algorithms.A new multi-class classification algorithm is proposed based on the idea of Locally Linear Embedding (LLE). The algorithm concerns the reconstruction error of samples and the error from different class samples in neighborhood. Actually, the errors reflect the relation between samples and the low-dimensionality manifold. The algorithm is applied to target recognition of MMW high range resolution radar based range profile. The experimental results indicate that the algorithm can classify efficiently. Compared with current popular multi-class classification algorithms, it shows better performance. Moreover, estimation of the parameters is simple and the result is hardly affected by selection of parameters. The detection precision is improved efficiently.更多还原