Reflection Removal With Generative Adversarial Networks

When we take a photo through transparent glass,the image we obtain often contains both the desired scene that beyond the glass(the target image)and the undesired scene that reflected by the glass(the reflection image).In fact,the image we obtain is a linear superposition of the target image and the reflection image.Decomposing the single image into two separate images is a highly ill-posed problem: in the absence of prior knowledge about the screen,mathematically,there are infinite number of valid decompositions.Researchers have proposed different priors to make the problem more constrained.However,most of these priors are lowlevel and because of that,they are not robust with respect to non-trivial images.In this paper we propose a novel deep learning method for single image reflection removal.To the best of our knowledge,this is the first work that shows how to tackle reflection removal problem with convolutional neural networks and generative adversarial networks.Our approach is to train a feedforward reflection removal network that is able to decompose a given superposed image into two output images using just one feedforward operation.The reflection removal network is a deep convolutional neural network that are able to extract high-level features from the input image after training on a large image dataset.This ability is crucial for our reflection removal algorithm in that the low-level structures of the input superposed image is often corrupted during the superposition process so that they are not stable,while high-level features are stable in spite of the process.In order to overcome the challenge of training the reflection removal network,which is caused by the symmetric roles of the target image and the reflection image,we make use of generative adversarial network framework: by introducing a critic network,we are able to work around the problem,so that we can successfully train the reflection removal network.Our reflection removal network demonstrates state-of-the-art decomposition quality and achieves fast speed.