Sparse principal component analysis for image in-painting

Sparse representation, as a powerful statistical image modelling technique, has been proved to be a successful method for image restoration applications. Because of the development of l1-minimization, the optimization of sparse representation can be solved efficiently. Intrinsically, natural images are sparse in some domain, and the image quality largely depends on whether the sparse domain can represent the image well. Considering that the contents can vary significantly across different patches in a single image, sparse represented clear patches are employed to replace missing pixel ones. We propose a sparse principal component analysis approach to adaptive learn dictionary (basis) to sparsely represent the image patches for image in-painting. The in-painting problem and dictionary learning process are combined together and easily solved by soft-thresholding and SVD (singular value decomposition) alternatively. The experimental results on image in-painting validate that sparse principal component analysis method achieves much better performance and faster computation comparing to the state-of-the-art dictionary learning algorithms in both RMSE and visual perception.