Computer Vision News - May 2019

What all these image decompositions have in common is the fact that the distribution within small patches within each layer is “simpler” (more uniform) than for small patches of the “mixed” (original) image, resulting in a strong internal resemblance for each layer. The statistical characteristics (distribution) of small image patches (such as 5x5, 7x7) have been proven to widely repeat in a natural image. This strong internal repetition has been exploited for tackling a wide variety of computer vision tasks. The authors’ approach combines the power of small image patches repeating throughout the image (its power of solving tasks without supervision) with the power of deep learning, and they propose a robust, unsupervised framework, based on DIP networks. A single DIP network was shown to be sufficient to capture the low level statistics of a single natural image, when the input for the DIP network is random noise, and the network learns to reconstruct a single image (the image serving as the sole input for training the network). This network proved to be powerful enough to solve problems such as denoising, super resolution, inpainting, doing all this without supervision. Computer Vision News 5 Research Computer Vision News … the power of small image patches repeating throughout the image with the power of deep learning Double-DIP