Single-image reflection removal using conditional GANs

Removing undesired reflections from a single-image is inherently ill-posed problem since defining its exact physical model is almost impossible. Most previous works tackle this problem through the use of multiple images or hand-crafted features. These methods are still quite limited in terms of the generality and photo-realistic result. In this paper, we propose a conditional Generative Adversarial Networks based deep neural network structure to render realistic image and formulate our problem in a simple objective function. Specifically, we use gradient information to elaborate this formulation to preserve both low and high frequency details. Our proposed network does not rely on any physical prior information and performs effectively with a single-image. Experimental results demonstrate that proposed algorithm conducts favorably against existing algorithms from human perceptual aspect.