GAN based single-image reflectance removal using depth of field guidance

Eliminating reflections on a single-image has been a challenging issue in image processing and computer vision, because defining an elaborate physical model to separate irregular reflections is almost impossible. In fact, while human vision can automatically focus on the transmitted object, basic deep neural networks even have a limitation to learn the attentive mechanism. In this paper, to solve this problem, a Generative Adversarial Networks guided by using Depth of Field (DoF) is proposed. The DoF is formulated by using image statistics and indicates the focused region of image. Thus, by adding this information to both generative and discriminative networks, the generator focuses on the transmitted layer and the discriminator will be able to estimate the local consistency of the restored areas. Since it is intractable to obtain the ground-truth transmitted layer in real images, a dataset with synthetic reflection is considered for quantitative evaluation. The experimental results demonstrate that the proposed method outperforms the existing approaches in both PSNR and SSIM. The visual outputs indicate that the proposed network convincingly eliminates the reflection and produce sufficient transmitted layers as compared to the previous methods.