Image restoration under mixed noise using globally convex segmentation

The total variation based regularization method has been proven to be quite efficient for image restoration. However, the noise in the image is assumed to be Gaussian in the overwhelming majority of researches. In this paper, an extended ROF model is presented to restore image with non-Gaussian noise, in which the locations of the blurred pixels with high level noise are detected by a function and two estimated parameters of noise, while the fidelity and smoothness terms can be adaptively adjusted by updating these parameters. In contrast to the previous method, our model can give a much better restoration in some particular cases, such as the blurred image corrupted by impulsive noise and mixed noise. Moreover, the proposed minimization problem is solved by the split Bregman iteration, which makes our algorithm very fast. We provide some experiments and comparisons with other methods to illustrate the high efficiency of our method.

Research highlights
► We propose an unified model for image restoration and noise detection.
► We give a new adaptive data-fidelity term.
► The implementation is fast due to applying the split Bregman method.
► Improving the quality of the restored images under mixed noise.