The term super-resolution refers to the process of creating a high resolution image from a sequence of sampling limited low resolution observations. The attractiveness of the simple idea, that collecting undersampled frames and applying sampling theory one can obtain a decent image quality video, has led to explosive development of multi-frame super-resolution processing. Resulting algorithms have found their applications in remote sensing where certain bandwidth limitations and pixel size restrictions are present; in security and surveillance imaging where more information regarding particular scene details is required on demand; and in medical imaging where desire to reduce irradiation dosage is paramount.
However, further advances in super-resolution processing will require dealing more effectively with the same problems it has been battling since its birth. The most fundamental of these tasks is the problem of accurate motion estimation between the images. There has been a constant stream of new motion estimation algorithms. However, their accuracy when applied to real life images (especially impaired by aliasing because of undersampling), doesn’t meet the necessary sub-pixel registration levels. In part, this is due to the inherent ill-posed nature of the optical flow problem. State-of-the art strategies to overcome this deficiency include, but not limited to, Bayesian approaches using simultaneous estimation of motion, blur, and noise; optical flow uncertainty measure estimation; or foregoing explicit motion estimation altogether and instead employing block matching algorithms such as collaborative sparsity inducing filtering; steering kernel regression; and nonlocal means. Another super-resolution problem is adequate representation of image collection process by an optical system which usually involves spatially varying point spread function and noise. The problem of spatially variant blur kernel has only been considered for the most benign cases.
This Research Topic covers addressing these super-resolution challenges which would broaden the scope of algorithm application to various types of real life imagery and video. Perhaps even more interesting are innovative extensions of the general super-resolution approach beyond employing similar low resolution frames. Super-resolution has already demonstrated its ability to upsample hyperspectral data, to perform panchromatic sharpening, as well as, has shown promise in demosaicking of color filter arrays. Therefore this Research Topic also considers algorithmic expansion towards the use of disparate images collected for example by cameras with dissimilar properties or by diverse sensor modalities such as spectral or polarization. Finally, there is new interest in leveraging recent developments in compressive sensing to solve problems in super-resolution imaging.
The term super-resolution refers to the process of creating a high resolution image from a sequence of sampling limited low resolution observations. The attractiveness of the simple idea, that collecting undersampled frames and applying sampling theory one can obtain a decent image quality video, has led to explosive development of multi-frame super-resolution processing. Resulting algorithms have found their applications in remote sensing where certain bandwidth limitations and pixel size restrictions are present; in security and surveillance imaging where more information regarding particular scene details is required on demand; and in medical imaging where desire to reduce irradiation dosage is paramount.
However, further advances in super-resolution processing will require dealing more effectively with the same problems it has been battling since its birth. The most fundamental of these tasks is the problem of accurate motion estimation between the images. There has been a constant stream of new motion estimation algorithms. However, their accuracy when applied to real life images (especially impaired by aliasing because of undersampling), doesn’t meet the necessary sub-pixel registration levels. In part, this is due to the inherent ill-posed nature of the optical flow problem. State-of-the art strategies to overcome this deficiency include, but not limited to, Bayesian approaches using simultaneous estimation of motion, blur, and noise; optical flow uncertainty measure estimation; or foregoing explicit motion estimation altogether and instead employing block matching algorithms such as collaborative sparsity inducing filtering; steering kernel regression; and nonlocal means. Another super-resolution problem is adequate representation of image collection process by an optical system which usually involves spatially varying point spread function and noise. The problem of spatially variant blur kernel has only been considered for the most benign cases.
This Research Topic covers addressing these super-resolution challenges which would broaden the scope of algorithm application to various types of real life imagery and video. Perhaps even more interesting are innovative extensions of the general super-resolution approach beyond employing similar low resolution frames. Super-resolution has already demonstrated its ability to upsample hyperspectral data, to perform panchromatic sharpening, as well as, has shown promise in demosaicking of color filter arrays. Therefore this Research Topic also considers algorithmic expansion towards the use of disparate images collected for example by cameras with dissimilar properties or by diverse sensor modalities such as spectral or polarization. Finally, there is new interest in leveraging recent developments in compressive sensing to solve problems in super-resolution imaging.