Image Super-Resolution Using Gradient Guidance and Generative Adversarial Networks

Abstract

To address the issue of poor visual quality and structural distortion in existing image super-resolution reconstruction models, a new approach utilizing a deep gradient guidance generative adversarial network is introduced. This model incorporates a gradient branch within the generator, which transfers gradient image features and integrates this gradient information with the image branch to maintain edge integrity. Inspired by MSRB, ResNext, and Inception, an enhanced multiscale residual block is developed and integrated into the core modules of both the image and gradient branches, facilitating the capture of multi-scale information. The discriminator is enhanced with WGAN-GP to bolster the stability of network training. When compared to perception-driven algorithms like SRGAN, ESRGAN, and NatSR, this new approach more effectively prevents structural distortions and elevates the quality of the generated images. The computational complexity of this model is 23.7GFLOPs, significantly lower than that of ESRGAN and SPSR by approximately 1/4 and 1/10, respectively