Super-Resolution Generative Adversarial Network with Modified Architecture for Single Image Super-Resolution

Abstract

Recently, Single image Super-Resolution (SISR) has become an attractive research area in Image processing which generates a High-Resolution (HR) image by using Single Low-Resolution (LR) image. Deep learningbased SISR approaches have achieved better Super-Resolved (SR) results by using mean squared error (MSE) as an objective function that increases the quality of SR results over performance metrics like peak-signal-to-noise-ratio (PSNR) and structural similarity index (SSIM). Nevertheless, MSE based approaches lead to generate over smoothed images with less high-frequency texture information at larger upscaling factors. Recent experiments have proved that Generative Adversarial Networks (GAN) generates perceptually convincing SR images through efficient extraction of high-frequency information from single LR image. In this paper, we propose a GAN based approach for SISR with modified deep-residual network architecture. In our proposed technique, we introduce the bottle-neck convolutional (CN) layer in the network structure of the Generator. Adding bottle-neck layers improves the network performance through 1 x 1 convolutional layers which extract complex features from the input and also reduces the computational complexity compared to 3 x 3 convolution layers. We further improve the model performance by removing batch normalization layer from the entire generator to overcome the unpleasant artifacts and improves GPU usage while training.