Optimized neighbor embeddings for single-image super-resolution

Abstract

We describe a self-content single-image super-resolution algorithm based on multi-scale neighbor embeddings of small image patches. Given an input low-resolution patch, we gradually expand its size by relying on local geometric similarities of low-and high-resolution patch spaces under small scaling factors. We characterize the local geometry with K-similar patches taken from an exemplar set and we collect exemplar patch pairs from the input image and its appropriately rescaled versions. While ensuring local images compatibility with an optimization on K, we satisfy image smoothness by patch overlapping. We further enforce global consistency through an adaptive back-projection. Our experimental results show better performance on synthesizing natural looking textures and sharp edges with less artifacts when compared to other methods.

Weight Optimization for Multiple Image Integration and Its Applications

“…The proposed SR method is based on the Turkan et al's SR method [14]. We will briefly summarize this algorithm first, and then, we will explain the details of the proposed SR approach.…”

“…This approach is based on the assumption that the small patch geometry can be preserved under small scaling factors [11]. The conventional method [14] calculates the scalar reconstruction weights of the K-NN patches in…”

“…In the conventional method [14], the set of down-scaled images N s t is introduced for searching the K-NN patches by down-scaling the input image l with several pixel-shifting offsets, where s t is the down-scaling factor. Let the input LR patch size be m × m and the iteration number t (we set t = 1 as the initial value) in the Step 1.…”

“…In this experiment, we compared our SR results with Turkan et al's SR method [14]. We set the SR factor s = 2, the initial LR patch size m = 3, and the increase size of patches p = 1.…”

“…SR is the problem of recovering a high-resolution (HR) image from one or multiple lowresolution (LR) image inputs. Turkan et al [14] introduced an effective single-image SR technique based on manifold learning. First, they estimate the optimal weights to approximate an input LR patch from its K nearest neighbors (K-NN) patches searching from the down-scaled images introduced from the input LR image, using locally linear embedding (LLE) [15].…”