A FAST PARALLEL ALGORITHM FOR BLIND ESTIMATION OF NOISE VARIANCE

The only information available to a blind noise variance estimation algorithm is the corrupted image and the white nature of the zero mean Gaussian noise. The proposed algorithm recovers the variance of the noise in two steps. First, the sample variances are computed for square cells tessellating the noisy image. Several tessellations are applied with the size of the cells increasing fourfold for consecutive tessellations. The four smallest sample variance values (the outcomes of the first four order statistics) are retained for each tessellation and combined through an outlier analysis into one estimate. The different tessellations thus yield a variance estimate sequence. In the second part of the algorithm, the value of the noise variance is determined from this variance estimate sequence. We have applied the blind noise variance algorithm to 500 noisy 256 × 256 images derived from seven prototypes of classes often employed in computer vision and image processing. In 98 percent of the cases the relative estimation error was less than 0.2 with an average error of 0.06. Application of the algorithm to differently sized images is also discussed. All the operations in the algorithm are parallel and if they are implemented on an image pyramid, the variance of the noise is recovered in 0[log (image__size)] processing time. © 1990 IEEE