Deep Super-Resolution Imaging Technology: Toward Optical Super-Vision

Abstract

Spatially variant optical blur is inevitable in real optical imaging systems with imperfect camera lenses since each object has a different depth in a real scene. Unfortunately, existing super-resolution imaging techniques may not achieve satisfactory performance on images acquired from practical optical imaging systems, where the optical blur is unknown. This is because a predefined global point spread function is adopted for every spatial coordinate without any optical blur estimation in the existing models. To address the challenging issue in recent video technologies, in this article, we propose an optimal super-resolution imaging network based on object-adaptive optical point spread function estimation. In technical, a new object sharpness consistency cost is suggested in the dark channel domain to estimate the point spread function, which is optimal to the super-resolution imaging. In addition, an object image sharpening network is proposed and adaptively combined with the point spread function estimation. The performance of the proposed network is also validated via experiments in terms of various quantitative metrics such as the peak signal-to-noise ratio, S3 index, and BRISQUE values.