Dual Stage Semantic Information Based Generative Adversarial Network For Image Super-Resolution

Deep learning methods for the super-resolution problem are showing great performance compared to other traditional techniques. However, these methods are unable to learn complex spatial structures and high frequency details; which leads to over-smooth results. In the present paper, a novel Generative Adversarial Network based architecture named as Residue and Semantic feature based Dual Subpixel Generative Adversarial Network has been proposed for generator and discriminator networks to solve super-resolution problem. The generator network is residue and semantic feature based dual subpixel generative architecture. This architecture is divided into two stages: premier residual stage and deuxieme residual stage. These two stages are concatenated together to form a two stage upsamping process, which enhances the feature learning capability of our model. Inter and intra residual connections are made within these two stages; helping us to sustain the high texture details of images. Semantic based information is implanted in generator to enhance the quality of objects in an image. For embedding semantic information in generator, feature maps extracted from pre-trained model are merged with the input image. To stabilize the training process, we introduced spectral normalization in the discriminator. Visual perception and mean opinion score shows that proposed method outperforms the other state-of-the-art methods.