A Novel Framework for Denoised High Resolution Generative Adversarial Network – DHRGAN

The advent of deeper convolutional neural networks and related methodologies have made significant achievements in the area of single image super-resolution (SISR). However, none of these techniques are equipped to handle noisy images, i.e., denoising the image as well as enhancing its spatial resolution. In this paper, we propose a denoised high resolution generative adversarial network (DHRGAN), capable of handling noise removal from given sample images while trying to super-resolve it to the desired magnification. As per our knowledge, this is the first GAN framework equipped to remove noise while simultaneously trying to magnify images. To accomplish this, we propose a two-layered generator network in addition to the discriminator network. In the upper layer of the generator network (also called UDCNN), we make use of the mean square error function for training samples while in the lower layer of generator network (called LSRResNet) we make use of a perceptual loss function (comprising of content loss function and adversarial loss function) for the same. We train the discriminator network, which is responsible for separating super-resolved images and ground truth images, using the adversarial loss function, thereby making the generator more robust. The content loss function that we employ in the generator actuates perceptual similarity rather than pixel similarity, which further strengthens its robustness. The resulting DHRGAN network can recover more realistic textures from heavily downsampled noisy images. We have used speckle noise, a common noise observed in natural scenes captured by airborne acquisition devices, to check the performance of DHRGAN. Experiments performed on standard image database, both visually and quantitatively, justify the superior performance of DHRGAN over existing similar networks.