Fast super-resolution method
For a variety of applications, we want to be able to increase the resolution of images. The ideal algorithm should be fast, and should add sharpness and detail, both at edges and in regions of texture, without adding artifacts. We have made a new version of our Markov network super-resolution algorithm which does not rely on the iterative belief propagation algorithm. The new algorithm has several speed improvements which result in doubling the resolution of an image from 100x100 to 200x200 in less than 2 seconds.
Background & Objective: For display of images on high resolution display devices, it is desirable to have an algorithm that increases the resolution of the displayed image, so that it has a more pleasing appearance. We would like to achieve higher image quality than our competitors by using a high quality, machine-learning-based image resolution enhancement algorithm.
Technical Discussion: We use a training based approach. We examine many pairs of high resolution, and low resolution versions of the same image data. We divide each image into patches, both high resolution and low resolution patches. We form a training database of 100,000 - 200,000 high and low-resolution patches.
Given a new low-resolution image, we seek to estimate the most probable corresponding high-resolution image. In the training database, there may be several different examples similar to any given input low-resolution patch. However, when we include the boundary conditions corresponding to the neighboring patches that have already been selected by the algorithm, we can find a single best-fitting high-resolution patch for each position in the resolution-enhanced image. We structure the database as a tree, and keep a list of the nearby neighbors to each leaf of the tree. This allows for a fast approximation to the nearest neighbor search that we seek to perform for each patch with the entire dataset. This efficiency results in the dramatic speed up over our previous implementation of the algorithm.
Furthermore, we now convert each input image to a luminance, chrominance color space and only apply super-resolution to the luminance component since this is where the high frequency information mainly resides. This results in a smaller training dictionary by a factor of 1/3, a speed-up by a factor of 3 and fewer artifacts in the resulting super-res image.
| Technical Reports: | |
| Example-based super-resolution | |
Technology Areas:
Computer Vision
Artificial Intelligence
Graphics
Modification Date: July 23, 2003