Wavelet-based Restoration Method for Swiss-cheese-type Defects in 2D/3D Images

摘要

Swiss-cheese-type defects (SCD) in 2D/3D images refer to corrupted small sets of pixels/voxels contained inside an image. The shape defined by the corrupted set can be arbitrary. Unlike image noise, which is a stochastic process; SCD is rather an image error due to various reasons that contribute to the final digital image. Traditional techniques of image interpolation, image enhancement, image restoration, image recovery, image error concealment, etc., do not address such problems. The proposed solution is to patch SCD with realistic sets of pixels/ voxels. Assuming the neighborhood of the corrupted set is reliable, interpolation based on the immediately local neighborhood is only applicable to relatively small and smooth corrupted sets. Images of natural scenes often contain similar structures that sparked successful image compression methods. This structure also helps to produce realistic image patches. We offer a solution based on Harr wavelet transformation and long range correlation. The arbitrary shape of the SCD posed extra complexities. We use Hilbert traversal to transform a 2D/3D image to a ID signal so that the arbitrary shaped SCD become corrupted ID segments. The restoration is performed in both ID and the image's original 2D/3D space.