Anchored Neighborhood Regression for Fast Example-Based Super-Resolution

Abstract

Recently there have been significant advances in image upscaling or image super-resolution based on a dictionary of low and high resolution exemplars. The running time of the methods is often ignored despite the fact that it is a critical factor for real applications. This paper proposes fast super-resolution methods while making no compromise on quality. First, we support the use of sparse learned dictionaries in combination with neighbor embedding methods. In this case, the nearest neighbors are computed using the correlation with the dictionary atoms rather than the Euclidean distance. Moreover, we show that most of the current approaches reach top performance for the right parameters. Second, we show that using global collaborative coding has considerable speed advantages, reducing the super-resolution mapping to a precomputed projective matrix. Third, we propose the anchored neighborhood regression. That is to anchor the neighborhood embedding of a low resolution patch to the nearest atom in the dictionary and to precompute the corresponding embedding matrix. These proposals are contrasted with current state-ofthe-art methods on standard images. We obtain similar or improved quality and one or two orders of magnitude speed improvements.

Acceleration techniques and evaluation on multi-core CPU, GPU and FPGA for image processing and super-resolution

“…1d ) the results display a perceptible contrast enhancement. Similar results (i.e., aliasing reduction and contrast enhancement) can be observed when SIL-SEABI is applied before DPSR [9] and ANR [2] respectively. Please see Fig.…”

“…These results are consistent with those documented in [6] regarding dictionary-based methods processing a single image. Table 2 displays the average per resolution quality difference between the results obtained by the methods in [32,33,2,35,9,36] when using L-SEABI /SIL-SEABI /L-SEAI and the results obtained by the same methods when using the parameters proposed by their respective authors in place of their construction phase. Providing alternative initialization to the method in [35] results in a slight MSSIM degradation of 0.08 on average.…”

“…Motivated by the result that L-SEABI can increase the quality of common interpolation methods we use L-SEABI to improve the performance of state-of-theart SR. Ergo, this section presents the application of a) L-SEABI, b) SIL-SEABI and c) their construction phase only (noted as L-SEAI ), in place of the initial reconstruction step of the [32,33,2,35,9,36] methods to upsample each low-resolution image by a factor of two in each dimension. The output quality is measured by employing the full-reference MSSIM [37] and the no-reference BRISQUE [38] metrics.…”

“…Comparison to the state-of-the-art shows [6] that L-SEABI is faster than a) dictionary-based methods such as the Sparse Representation with over-complete dictionaries [32], the Nonlocal Similarity Adaptive Regularization [33] (ASDS-AR-NL) and the Anchored Neighborhood Regression (ANR) [2], b) Bayesian prior methods such as the Norm l1 [34], and better than c) IBPbased methods such as the Nonlocal Iterative BackProjection [35] (NLIBP ). More recently, the authors in [36] proposed the concept of a nonlocal autoregressive model (NARM ) i.e.…”

“…Such approaches exploit Super-resolution (SR) techniques that construct high-resolution (HR) images from several low-resolution (LR) observations and they thus trade off computational with hardware/implementation cost [1]. Consequently, current research focuses on improving the complexity and/or the running time performance of these processes [2] [3] [4] [5].…”