Computer Vision News - May 2019

This phenomenon proved to be consistent, so the authors repeated the experiment on 100 randomly selected pairs of natural images from the BSD100 dataset -- and in the case of natural images rather than patterns, there was an even higher disparity in MSE loss values between the mixed image and its component images. Method: The figure above demonstrates the Double-DIP Framework: two Deep-Image- Prior networks (DIP1 & DIP2) decompose the input image I into its layers (y1 & y2), then those layers are recomposed according to a learned mask m, reconstructing an image ≈ . What constitutes a good image decomposition? There are many ways to decompose an image into basic layers, however, the authors propose the following characteristics as defining a meaningful decomposition: 1. The recovered layers, when recombined, reconstruct the input image. 2. Each layer should be as “simple” as possible, that is, it should have a strong internal self-similarity of image elements. 3. There should not be dependence between the recovered layers, or there should be the least correlation possible between them. 7 Research Computer Vision News Double-DIP