Independent component analysis based on a scatter diagram
Katsuhisa Hirokawa, Taro Yamaguchi, Kazuyoshi Itoh · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 1999
We propose a novel algorithm for the independent component analysis that reshapes a scatter diagram or the joint probability density function of the pixel values of the mixed signals. The detailed procedure of the algorithm is described and its performance is numerically evaluated. The present algorithm geometrically separates original independent images from mixed images by the linear transformation, in which shift, shearing and rotation are performed to the scatter diagram. After the transformations and normalizations, the basis vectors of the independent signal become orthogonal and match the coordinates. In the implementation, it is shown that the algorithm can decompose two independent images from two mixed images. In additions, the proposed algorithm are compared to two conventional algorithms; [G. Burel: Neural Networks 5, 937 (1992)] and [A. Hyvarinen and E. Oja: Neural Comp. 9, 1483 (1997)], in calculation speed and extraction accuracy.