Fast Super-Resolution via Dense Local Training and Inverse Regressor Search

Abstract

Abstract. Regression-based Super-Resolution (SR) addresses the upscaling problem by learning a mapping function (i.e. regressor) from the low-resolution to the high-resolution manifold. Under the locally linear assumption, this complex non-linear mapping can be properly modeled by a set of linear regressors distributed across the manifold. In such methods, most of the testing time is spent searching for the right regressor within this trained set. In this paper we propose a novel inverse-search approach for regression-based SR. Instead of performing a search from the image to the dictionary of regressors, the search is done inversely from the regressors' dictionary to the image patches. We approximate this framework by applying spherical hashing to both image and regressors, which reduces the inverse search into computing a trained function. Additionally, we propose an improved training scheme for SR linear regressors which improves perceived and objective quality. By merging both contributions we improve speed and quality compared to the stateof-the-art.

Antipodally Invariant Metrics for Fast Regression-Based Super-Resolution

“…KRR of [15]) results in prohibitive training complexity, which is usually mitigated by reducing the training sets under certain assumptions. In testing time, although nonlinear regression SR methods are reasonably fast, they still compare modestly with fast state of the art methods such as [13], [14], [21].…”

“…Following this observation, in [14] we proposed a different approach when training linear regressors for SR: Using sparse representations as anchor points to the manifold, but forming the neighborhoods with raw manifold samples (e.g. features, patches).…”

“…It is recurrent in literature the usage of Euclidean distance for this purpose. We show that for the characteristics of the features used in our proposed algorithm (which are also shared in other SR algorithms [10], [13], [23], [14]), Euclidean distance is suboptimal as it fails to manage antipodal points (i.e. two points on the surface of a sphere are antipodal if they are diametrically opposite).…”

“…We obtain maximum quality by setting our dictionary size to 8192 elements and, afterwards, fixing a number of hyperspheres which gives us the desired speed. We aim to obtain the same speed as the original work of [14] where they used 1024 atoms and 6 hyperspheres. We obtain a very similar time figure (while obtaining substantially improved PSNR quality) with 7 spheres and 8192 atoms.…”

“…The key observation of performing the neighbor embedding just for a predefined set of anchor points allowed to preprocess them during training time, thus lessening substantially the testing time complexity. Our recent work [14] puts light in how to properly create the neighborhoods in such methods, obtaining sizable benefits in terms of reconstruction quality. In addition to that, we also explored sublinear search structures for the regressor nearest neighbor search, as this takes a significant quota of the running time from within the processing of the whole SR pipeline.…”