LIDAR Image Recovery by Incorporating Heterogeneous Imaging Modalities

摘要

As new imaging modalities arise, the problem of inpainting becomes increasing important. Typical techniques for inpainting are completely determined by the penalization term used in the optimization scheme. These methods range from minimizing over total variation to finding a sparsest solution in a given basis to minimizing the Ginzburg-Landau energy. In this paper, we propose a novel approach to inpainting of remote sensing images, which uses previous measurements taken from heterogeneous image soures in conjunction with these well studied penalization methods. These previous measurements could be images with different illumination or weather conditions, images with spatio-temporal changes, or even all together different imaging modalities. Our approach utilizes manifold learning techniques such as diffusion maps or Laplacian eigenmaps that are applied to each image. This is followed by learning a rotation between the two feature spaces in an effort to place data points from both images in a common feature space. Then, we apply a novel preimage algorithm to the fused data in conjunction with an inpainting penalization method to recreate the missing pixels.