Abstract: This paper presents a new image blind separation algorithm by using nonparametric entropy estimation. The new algorithm directly estimates m-spacing entropy according to mixture image signal X and avoids explicit probability density estimation. By exhaustive search of all possible rotation matrices, contrast function minimum and corresponding optimum rotation matrix is easily found without the presence of local minima trouble. It is fit for the diversity probability density distribution of image. Experiment tests show that it is a robust image separation algorithm with better performance than traditional ones such as FastICA, natural gradient(NG) and joint approximate diagonalization of eigen matrices(JADE) algorithm.

Key words: Independent component analysis; Nonparametric entropy; Image blind separation

摘要： 提出了一种基于非参数熵的图像盲分离新算法。该方法根据K-L 散度作为信号之间独立性优化准则，不利用概率密度函数知识，由观测向量直接估计m-spacing 熵，通过穷举搜索法寻找目标函数的最小值从而获得最佳旋转矩阵进行盲源分离，适合图像像素分布多样性特点。大量实验证实，该算法鲁棒性好、分离指标高、性能优于传统FASTICA、自然梯度等自适应算法。

关键词: 独立分量分析；非参数熵；图像盲分离