Camera Lens Super-Resolution

Existing methods for single image super-resolution (SR) are typically evaluated with synthetic degradation models such as bicubic or Gaussian downsampling. In this paper, we investigate SR from the perspective of camera lenses, named as CameraSR, which aims to alleviate the intrinsic tradeoff between resolution (R) and field-of-view (V) in realistic imaging systems. Specifically, we view the R-V degradation as a latent model in the SR process and learn to reverse it with realistic low- and high-resolution image pairs. To obtain the paired images, we propose two novel data acquisition strategies for two representative imaging systems (i.e., DSLR and smartphone cameras), respectively. Based on the obtained City100 dataset, we quantitatively analyze the performance of commonly-used synthetic degradation models, and demonstrate the superiority of CameraSR as a practical solution to boost the performance of existing SR methods. Moreover, CameraSR can be readily generalized to different content and devices, which serves as an advanced digital zoom tool in realistic imaging systems.

1. Introduction

Single image super-resolution (SR) is a typical inverse problem in computer vision. Generally, SR methods assume bicubic or Gaussian downsampling as the degradation model [33]. Based on this assumption, continuous progress has been achieved to restore a better high-resolution (HR) image from its low-resolution (LR) version, in terms of reconstruction accuracy [9],[13],[15],[17],[23],[25],[27],[31],[32],[35],[36] or perceptual quality [2], [3], [5], [12], [16], [22],[28]. However, these synthetic degradation models may deviate from the ones in realistic imaging systems, which results in a significant deterioration on the SR performance [20]. To better simulate the challenging real-world conditions, additional factors including noise, motion blur, and compression artifacts are integrated to characterize the LR images in either a synthetic [26] or a data-driven [4] manner. These modified degradation models promote the SR performance of learning-based methods when the LR images indeed have corresponding degradations.