LAUSR

Generative adversarial networks (GAN) have been widely used in the field of image-to-image translation. In this paper, we have proposed a novel object extraction and background recovery (OEBR-GAN) model, which can extract objects from an image and then complete the image by inpainting the background of the image. This model has been developed for a solar panel installation project, where the user would like to input an original colored image of the roof, and as output, the user requires an edge detected roof image. However, the condition in user requirement is that any object that is hiding the roof edges should be removed first and the background of that part of the roof image should be recovered so that the user can obtain a complete connected edge detected image of the roof. Therefore, the model also completes the image by connecting the hidden edges of the roof. We could achieve the user objective by building a GAN model with a dual generator and dual discriminator network. The generators have been built using an encoder-decoder network with and without skip connections and the discriminators have been built using deep convolutional neural networks and encoder architecture. Quantitative comparisons in the result section shows that OEBR-GAN performs much better than other adversarial models on our collected dataset.