Recently, many computational saliency models have been introduced 2, 5, 7, 13, 23 to transform a given image into a scalar-valued map that represents visual saliency of the input image. These approaches, however, generally assume the given image is clean. Fortunately, most methods implicitly suppress the noise before calculating the saliency by blurring and downsampling the input image, and therefore tend to be apparently rather insensitive to noise. 11 However, a fundamental and explicit treatment of saliency in noisy images is missing from the literature. Indeed, as we will show, the price for this apparent insensitivity to noise is that the overall performance over a large range of noise strengths is diminished. Accordingly, the question is how to compute saliency in a reliable way when a noise-corrupted image is given. To address this problem, we propose a novel and statistically sound method for estimating saliency based on a non-parametric regression framework. The proposed estimate of the saliency at a pixel is a data-dependent weighted average of dissimilarities between a center patch and its surrounding patches. This aggregation of the dissimilarities is simple and more stable despite the presence of noise. For comparison's sake, we apply a state of the art denoising approach before attempting to calculate the saliency map, which obviously produces much more stable results for noisy images. Despite the advantage of preprocessing, we still found that our method consistently outperforms the other state-of-the-art 2, 13 methods over a large range of noise strengths.