Mixed Noise Removal Using Adaptive Median Based Non-Local Rank Minimization

Abstract

In this paper, we present an innovative mechanism for image restoration problems in which the image is corrupted by a mixture of additive white Gaussian noise (AWGN) and impulse noise (IN). Mixed noise removal is much more challenging problem in contrast to the problems where either only one type of noise model (either Gaussian or impulse) is involved. Several well-known and efficient algorithms exist to effectively remove either Gaussian noise or Impulse noise, independently. However, in practice, noise may occur as a mixture of such noise models. Thus, the existing techniques devised to handle individual types of noise may not perform well. Moreover, the complexity of the problem hinges on the fact that the removal of either type of noise from the given image affects the noise statistics in the residual image. Therefore, a rigorous mechanism is required which not only infers altered noise statistics but also removes the residual noise in an effective manner. In this regard, an innovative approach is introduced to restore the underlying image in three key steps. Firstly, the intensity values, affected by impulsive noise, are identified by analyzing noise statistics with the help of adaptive median filtering. The identified intensity values are then aggregated by exploiting nonlocal data redundancy prior. Thus the first step enables the remaining noise to follow the zero mean Gaussian distribution in the median filtered image. Secondly, we estimate Gaussian noise in the resulting image, which acts as a key parameter in the subsequent singular value thresholding process for rank minimization. Finally, a reduced rank optimization applied to the pre-processed image obtained in the first step. The experimental results indicate that the proposed AMNLRA (Adaptive Median based Non-local Low Rank Approximation) approach can effectively tackle mixed noise complexity as compared to numerous state of the art algorithms.

Keywords

• Image denoising
• low rank approximation
• mixed noise
• nuclear norm
• rank minimization
• similarity measure