SPARSE GEOMETRY FOR SUPER RESOLUTION VIDEO PROCESSING

摘要

In a method of analyzing an input video sequence, pixels of synthesized images of an output video sequence are associated with respective directions of regularity belonging to a predefined set of directions. A first subset of candidate directions is determined from the predefined set of directions for a region of a first image of the output sequence. For a corresponding region of a second synthesized image of the output sequence following the first image, a second subset of candidate directions is determined from the predefined set of directions, based on images of the input sequence and the first subset of candidate directions. The directions of regularity for pixels of this region of the second synthesized image are detected from the second subset of candidate directions. The recursive determination of the subsets of candidate directions provides a sparse geometry for efficiently analyzing the video sequence. The method is well suited for super-resolution video applications such as deinterlacing, frame rate conversion and noise reduction.