Image Deblurring Using Generative Adversarial Networks (GANs)

Image inpainting and image blur restoration are two essential image processing tasks in computer vision wherein missing or corrupt areas in images are filled in and blurry images corrected. Due to the characteristics of these two problems, they are both extremely difficult; it is necessary to predict the right kind of content while taking care not to distort or lose the general shape, texture, and contextual relevance of the image. GAN has come in handy for these cases, as they are quite efficient when it comes to both painting images and deblurring them. Images are formed using a generator and discriminator, both being neural networks, which function in a non-simultaneous manner; the former is responsible for generating images while the latter determines whether the images are authentic or not. By this mean approach, the generator is able to optimize its outputs gradually, and so is the discriminator in that it can better tell between real and fake images. In the paper, we show that the acknowledged capabilities of GANs can be effectively employed in image inpainting as well as blur restoration. Intensive evaluations on a number of datasets comprising of various levels of synthetic and real blur have proved that the proposed GAN-based method of detail restoration and reconstruction is superior even when the images are very blurry. This work reveals that GANs can be deployed in very different applications in a very straightforward and convincing way.