3) Natural image occlusions

This dataset contains an image of underbrush in a forest (taken bridge.jpg, which has been used for denoising bench-marking [33]), which is rich with occluding branches and twigs. From this original noise-free image, we cut a 110 × 110 pixel occlusion-rich section and scaled it up to 256 × 256 pixels to use in our dataset. To compose the dataset, we cut the 256 × 256 image (with pixel values ranging from (0, 255)) into N = 61009 overlapping image patches of D = 9 × 9 pixels, then add independent Gaussian noise with σ = 5. In the corresponding publication, the data is shown in Figure8A-B. [33] Mairal, J., Bach, F., Ponce, J., Sapiro, G., and Zisserman, A. (2009): Non-local sparse models for image<br>restoration. International Conference on Computer Vision 25.