Independent components extraction from image matrix

The key problem of extracting independent components (ICs) is to learn the demixing matrix from the known training images which can be unfolded to vectors in conventional independent component analysis (ICA). However, the unfolded vectors lead to the small sample size problem (SSS) and the curse of dimensionality. In this paper, a novel independent feature extraction method is proposed to solve these problems by encoding each input image as a matrix. In addition, the row and column directional images of the matrix are introduced to better exploit the spatial and structural information embedded in image during the training phase. Compared with the conventional ICA, the proposed method directly evaluates the two correlated demixing matrices from the image matrix without matrix-to-vector transformation, greatly alleviates the SSS and the curse of dimensionality, reduces the computational complexity, and simultaneously exploits the spatial and structural information embedded in image. Extensive experiments show that the proposed method is superior to the standard ICA method and some unsupervised methods.