【作者】 郭武；

【导师】 王润生；

【作者基本信息】 国防科学技术大学 ， 信息与通信工程， 2008， 博士

【摘要】 独立分量分析(Independent Component Analysis, ICA)是一种新近发展起来的信号处理技术,在许多领域有着广泛应用。论文围绕ICA算法及其在图像处理中的应用进行了深入系统的研究。在论述ICA基本方法的基础上,论文深入分析了广为应用的快速ICA算法(FastICA),提出一种改进算法。通过在下降方向增加一维搜索,减少了迭代次数,改善了算法的收敛性能。在ICA用于纹理图像分类的研究中,提出了一种基于独立子空间分析(Independent Subspace Analysis, ISA)和独立谱表述(Independent Spectral Representation, ISR)的纹理特征提取方法。所提供的特征滤波器与传统的滤波器相比,可以有效挑选出数量更少、分类性能更优的纹理特征。在ICA用于图像去噪的研究中,从变换和分离两个不同的技术建立了去噪算法。一、从变换的角度出发,提出了一种改进的ICA变换去噪算法,通过建立收缩补偿的概念和一种新的收缩函数,可以有效地避免软阈值收缩函数在去噪过程中造成的图像边缘特征损失。二、从分离的角度出发,将噪声图像作为源图像和噪声混合叠加而成的混合信号,尝试了全新的处理模式。据此,我们分别提出用重构相空间(RPS)和增加虚拟观测两个方法进行图像去噪,初步的实验结果是令人满意的。在ICA用于图像分离的研究中,结合脊波变换和复杂性寻踪,建立了两个混合图像分离算法。一、从脊波变换出发,提出了基于脊波变换的ICA混合图像盲分离算法。该算法保留了小波ICA的特点,同时由于脊波可以有效表述直线的特性,提高了具有明显线特征的混合图像的分离精度。该方法适用于源图像相互统计独立的混合图像分离。二、从复杂性寻踪出发,将图像分离过程表述为复杂性寻踪过程,提出一种基于复杂性寻踪的不动点迭代算法,成功分离了传统盲分离算法不能分离的非独立源混合图像。在ICA用于运动目标检测的研究中,在提出一种结合Infomax和FastICA的改进梯度学习算法基础上,形成了一个改进的运动目标检测算法。它可以精确检测出序列图像中的运动目标,具有较强的抗噪声性能。更多还原

【Abstract】 Independent component analysis (ICA) is a new signal processing technique developed in the last decades and has been used in a broad range of applications. This thesis focuses on the ICA algorithm and its applications in image processing.Based on discussion of ICA basic theory, the thesis analyses the fast fixed point algorithm (FastICA),which is widely applied. Then a modified FastICA algorithm is offered by imposed one-dimensional search on the iterative direction. With the modified algorithm, number of iteration is reduced and convergence performance is improved.In texture image classification, a method to select the texture feature based on independent subspace analysis (ISA) and independent spectral representation is proposed. We can achieve better classification performance by the feature filters comparable to other traditional filter schemes while resulting in considerably smaller filters.In image denoising, two different denoising techniques are developed. The first is based on ICA transform. After obtained the independent components of the corrupted image by ICA, a new compensating operation and an improved shrinkage function are proposed to effectively combat the loss of image details in ordinary soft-thresholding shrinkage. The second is based on BSS. In this method the noisy image is regarded as a mixture by the source image and the noise, and fire-new processing patterns are tried. Based on it, two approaches to separate source image and the noise are developed, which use the theory of reconstructive phase space (RPS) and adding a dummy image as another observation, respectively. The experimental results is satisfied.In image separation, two algorithms are proposed. The first is based on ridgelet transform, it inherits the advantage of wavelet ICA and can improved the separation performance for the mixed images with notable line feature. The method is suitble to separate the mixed images, in which the source images are statistically independent each other. The second is based on complexity pursuit, it describes the process of separation as a process to find out the interesting projective directions, and a fixed point iterated algorithm is developed. It can separate the mixed images successfully, in which the source images are not statistically independent each other.In moving targets detection in the series of images, a modified algorithm of detecting moving targets is presented based on a new gradient leaning algorithm combined Informax and FastICA. It can detect the moving targets in image series accurately and be robust to the noise.更多还原