Example-based super-resolution

Abstract

Image-based models for computer graphics lack resolution independence: they cannot be zoomed much beyond the pixel resolution they were sampled at without a degradation of quality. Interpolating images usually results in a blurring of edges and image details. We describe image interpolation algorithms which use a database of training images to create plausible high-frequency details in zoomed images. Image pre-processing steps allow the use of image detail from regions of the training images which may look quite different from the image to be processed. These methods preserve fine details, such as edges, generate believable textures, and can give good results even after zooming multiple octaves.

Pairwise Operator Learning for Patch-Based Single-Image Super-Resolution

“…Single-image super-resolution [10], also named examplebased super-resolution [1], is a problem of estimating a high-resolution image from a given low-resolution image with the help of training samples that consist of low-and high-resolution image pairs. The task of single-image superresolution is modeling the information containing in these training samples for super-resolving a test low-resolution image.…”

“…Single-image super-resolution [1] can be understood as a problem of learning a regression function which maps lowresolution image patches to high-resolution image patches from the perspective of machine learning. Low-and highresolution images are transformed into a training set in which each training sample consists of low-and high-resolution image features extracted from training image patches.…”

Apparent Resolution Enhancement: Structural Similarity Perspective

“…In frequency correspondence approach, high resolution images are obtained through learning the correspondences from low resolution images. Such approaches are termed as Example Based Super-Resolution (EBSR) approaches [7]. In general, the SR methods can be classified into two classes; Reconstruction based methods and Example based methods [1].…”

“…Frequency domain [20], Maximum Likelihood, Non-uniform interpolation (consisting of three stages including registration or motion estimation, interpolation and noise reduction), Regularized Super Resolution and Projection onto Convex Sets [3,8,9] are considered the most valuable algorithms in this regards. Recently, the EBSR methods for Super Resolution got the attention of the researchers by producing remarkable results, for example, starting with the work of [7,16]. The EBSR methods rely on the training data for the prediction of details in LR images, for the SR output.…”

“…The said algorithm results in super resolved background only, in the image. Freeman's [7] algorithm which is used in this work for the EBSR uses the generative model learned from HR images. A high frequency detailed image is obtained by using such model.…”