Super Resolution of Videos using SRGAN
International Journal for Research in Applied Science and Engineering Technology
The primary objective of our project is to generate the Super resolved images. To enhance the quality of images we are using SRGAN techniques. There are various methods of image transformation where the computing system receives some input and transforms it in output image. Previous techniques of up scaling were based on minimizing the mean squares reconstructed error. Generative Adversarial Networks are the deep neural network architectures comprised of two networks Generator and
... GANs are about creating, like drawing a portrait or composing a symphony. SRGAN provides several benefits over other techniques it proposes a perceptual loss function which consists of adversarial loss and content loss. The two main blocks are Generator and Discriminator. The Discriminator discriminates between real HR images from generated super resolved images. The Generator function is to train a propagating model. Adversarial loss function uses discriminator network which is already trained to discriminate between the two images. The content loss function uses perceptual similarity instead of pixel space similarity. One of the best things about SRGANs is that they generate data that is similar to real data. SRGANs learn the internal representations of data to generate the upscale images. The neural network is successful in recovering the photo realistic textures from downgraded images. The SRGAN techniques lack high peak to signal noise ratio but provides high efficiency and visual perception. Combining the perceptual and adversarial loss will generate a high quality super resolved image. During training phase perceptual losses measure image similarities more robustly than per-pixel losses. Perceptual loss functions measures the high level perceptual and semantic differences between the images. Our method uses semantic content losses which has a VGG network and an edge promoting adversarial losses for edges.