Removing water fluctuation via motion field-based kernel regression

A new approach is proposed in this paper to recover the original image of an underwater scene from a given sequence distorted by water fluctuation. The contribution of our method is to use a motion field-based kernel regression to reconstruct an undistorted image. It first utilizes the image registration method to generate registered frames and the corresponding motion fields, and then detect a set of stationary patches from registered frames by calculating the movement energy of local regions. Finally, a temporal kernel regression is carried out to combine the detected stationary patches to reconstruct an undistorted stationary image. Experiments illustrate that this approach can effectively alleviate distortions and significantly improve visual quality of the distorted images.