CCD image demosaicing using localized correlations

A new approach to image interpolation using spatial relationships between adjacent pixels is introduced. In its first stage, the localized statistical relationships are studied based on the sparse version of the image. In the second stage, the governing rules of the image are used to build an interpolated version. The proposed interpolation method is suitable for color single-CCD images for demosaicing purposes. The correlation rule is studied first for each color component separately, then difference images (modified hues) are built to eliminate the color correlation, leading to a smoother reconstructed signal. Since in Bayer pattern not all the color components are equally represented, the algorithm deals with the major green component differently from the red and blue, using the green as a basis for the whole image reconstruction. Further statistical tools are added to the algorithm to improve the visual results. We compare our method to presently available demosaicing techniques for single CCD color imaging with the major emphasis on reducing ghost colors and unreal edges. Our conclusion is that the proposed method can significantly improve interpolation and demosaicing tasks in image processing.