< Back to previous page

Publication

Low level vision features for enhancing images and videos

Book - Dissertation

In this work is addressed the problem of image and video enhancing. Enhancement is the process of transforming in a more pleasing version the original image. For this task we take advantage of several reference high quality photographs by employing local feature points that are matched reliably in order to find appropriate corresponding regions. First, we present a novel technique for matching images of widely separated views. Our approach is motivated by the recent comprehensive studies where none of the detector/descriptor combinations performed adequately where the camera viewpoint changed substantially. The method is characterized by two main steps that aim to estimate reliably the neighborhood region around feature points where descriptors are computed. In the beginning, a few kernel correspondences are identified in the images and then, based on their neighbor information, the geometric distortion that relates the surrounding regions of these seed keypoints is estimated iteratively. In the second step the neighbor regions around every keypoint are warped. This is done based on the estimated parameters combined with a rough segmentation, that reduces the searching space of the keypoints descriptors. This strategy shows to increase significantly the number of correct matches for wide separated views of a given 3D scenes. Next, we introduce a new technique for enhancing a low resolution video sequence using a set of high quality reference photographs that have been taken of the same scene. Our technique generates high quality frames by copying information from the photographs in a patch-wise fashion. The copying is guided by a sparse set of reliable correspondences between the video vi frames and photographs. The technique is purely image based, and does not require depth estimation. A robust descriptor is employed for establishing valid matches between the video frames and the photographs. Then, the geometric transformation is estimated between every corresponding patch. With only a few reference photographs, we are able to reduce noise and motion blur, and more important, increase resolution by a significant factor. Finally, we discuss an original approach that removes the undesired blur artifacts from photographs taken by hand-held digital cameras. Our method is based on the observation that in general several consecutive photographs taken by the users share image regions that project the same scene content. Therefore, we took advantage of additional sharp photographs of the same scene. Based on several invariant local feature points, filtered from the given blurred/non-blurred images, our approach matches the keypoints and estimates the blur kernel using additional statistical constraints. The latent image is restored by a simple deconvolution technique that preserves edges while minimizing the ringing effects. The experimental results demonstrate the capability of our technique to infer accurately the blur kernel while reducing significantly the artifacts of the spoilt images.
Publication year:2009
Accessibility:Closed