Context-aware single image super-resolution using sparse representation and cross-scale similarity

• A “denoised” sparse coding model is used to establish a primitive structure prior.
• Context is utilized to determine where the example-based method should be used.
• The proposed method gives the user flexibility in specifying the exemplar images.

Methods for single image super-resolution (SR) are broadly divided into two categories: (i) edge-focused SR and (ii) example-based SR. Edge-focused methods are focused on producing sharp edges with minimal artifacts; however, these methods cannot readily introduce rich texture details. Alternatively, example-based methods extrapolate new details by constructing a high-resolution (HR) image from example images. However, the quality of the constructed HR image is dependent on the suitability of the example images. This study aims to take advantage of both edge-focused and example-based SR in order to generate HR images with high-quality edges and rich details. Specifically, we use a sparse representation method to model a primitive structure prior. With this simple but very effective prior, the proposed method generates an edge-preserving HR image from the input low-resolution image. In addition, we propose a context-aware detection method that aims at determining image regions where details are effectively extrapolated, and then we exploit cross-scale self-similarity to determine the best examples for generating a detail-extrapolating HR image. The final output HR image is generated by integrating these two generated HR images in conjunction with the context-aware weight map. Experimental results demonstrate the effectiveness of the proposed algorithm.