Mixed noise removal based on a novel non-parametric Bayesian sparse outlier model

We develop a novel non-parametric Bayesian sparse outlier model for the problem of mixed noise removal. Based on the assumptions of sparse data and isolated outliers, the proposed model is considered for decomposing the observed data into three components of ideal data, Gaussian noise and outlier noise. Then the spike-slab prior is employed for outlier noise and sparse coefficients of ideal data. The proposed method can automatically infer noise statistics (e.g., Gaussian noise variance) from the training data without changing model hyper-parameter settings. It is also robust to initialization without using adaptive median filter as in other denoising methods. Experimental results demonstrate proposed model can achieve better objective and subjective performances on mixed noise removal than other state-of-the-art methods.