Bayesian approach to alignment-based image hallucination

摘要

Various technologies described herein pertain to enhancing a quality attribute of an input image. The input image can have a first level of the quality attribute. Dense correspondences between the input image and candidate exemplar images, which are included in an image database, can be computed utilizing a dense image alignment technique. The candidate exemplar images can have a second level of the quality attribute. Further, the candidate exemplar images can be warped to align with the input image based upon the dense correspondences. Moreover, patches from the candidate exemplar images as warped can be integrated with the input image to generate an output image. The output image can have the second level of the quality attribute.

主权项

1. A method of enhancing a quality attribute of an input image, comprising:
computing dense correspondences between candidate exemplar images and the input image utilizing a dense image alignment technique, wherein the candidate exemplar images are from an image database, wherein the input image has a first level of the quality attribute, and wherein the candidate exemplar images have a second level of the quality attribute; warping the candidate exemplar images to align with the input image based upon the dense correspondences; and integrating patches from the candidate exemplar images as warped with the input image to generate an output image having the second level of the quality attribute.