Motion Offset for Blur Modeling

Abstract

Motion blur caused by the relative movement between the camera and the subject is often an undesirable degradation of the image quality. In most conventional deblurring methods, a blur kernel is estimated for image deconvolution. Due to the ill-posed nature, predefined priors are proposed to suppress the ill-posedness. However, these predefined priors can only handle some specific situations. In order to achieve a better deblurring performance on dynamic scene, deep-learning based methods are proposed to learn a mapping function that restore the sharp image from a blurry image. The blur may be implicitly modelled in feature extraction module. However, the blur modelled from the paired dataset cannot be well generalized to some real-world scenes. To summary, an accurate and dynamic blur model that more closely approximates real-world blur is needed.

By revisiting the principle of camera exposure, we can model the blur with the displacements between sharp pixels and the exposed pixel, namely motion offsets. Given specific physical constraints, motion offsets are able to form different exposure trajectories (i.e. linear, quadratic). Compare to conventional blur kernel, our proposed motion offsets are a more rigorous approximation for real-world blur, since they can constitute a non-linear and non-uniform motion field. Through learning from dynamic scene dataset, an accurate and spatial-variant motion offset field is obtained.

With accurate motion information and a compact blur modeling method, we explore the ways of utilizing motion information to facilitate multiple blur-related tasks. By introducing recovered motion offsets, we build up a motion-aware and spatial-variant convolution. For extracting a video clip from a blurry image, motion offsets can provide an explicit (non-)linear motion trajectory for interpolating. We also work towards a better image deblurring performance in real-world scenarios by improving the generalization ability of the deblurring model.