Multi-frame super-resolution for mixed Gaussian and impulse noise based on blind inpainting

This paper proposes a robust multi-frame super-resolution algorithm to produce a high-resolution image. In merging the non-redundant information from shifted, rotated, blurred, noise-corrupted, low-resolution observations of the same scene, the approach registers the frame and reduces the impact of the distortions. The method is a generalization of a recently published blind inpainting algorithm to the multi-frame super-resolution case including both Gaussian and Impulse noises. Most multi-frame super-resolution algorithms, only consider blurring and Gaussian noise, ignoring noises that arise in real applications such as when processing time-of-flight camera depth images. Examples on simulated scenarios and real images produce results that compare favorably with other methods and clearly justify the benefits of this imaging model and the reconstruction method presented.