Blind image separation by combining neighborhood information

This paper proposes a novel approach to blind image separation. By using the property that the near pixels have close correlation, blind image separation is formulated to diagonalize several difference correlation matrices. Two algorithms are proposed based on generalized eigen-decomposition and a joint approximate diagonaliztion method respectively. The key point of our method is the formulation of difference correlation matrices. Since the mixed images are two dimensional, two mixed signals are formulated in columnwise order or in rowwise order respectively. Based on the close correlation of near pixels, the matrices to be diagonalized are constructed. The separation matrix for the mixed images can be obtained by generalized eigen-decomposition or joint approximate diagonalization. Compared with the ´Non-Negative PCA´ algorithm, the original signals in our algorithm are not required to be well-grounded, which means that they have a non-zero pdf in the region of zeros. In contrast to many second order methods in recent literatures, the two dimensional signals are used. Simulation results on mixed images are employed to further illustrate the advantages of our approach.