Enhancing Image Super-Resolution with Deep Gradient Guidance and Multi-Scale Residual Blocks in a GAN Framework

Abstract

To address the issue of poor visual quality and susceptibility to structural distortion in current image super-resolution reconstruction models, a novel approach based on a deep gradient-guided generative adversarial network (GAN) is proposed. In this model, the generator incorporates a gradient branch that transfers the features of the gradient image and integrates them with the image branch. This fusion of gradient information helps preserve the integrity of image edges, preventing distortion. Drawing on techniques from MSRB, ResNeXt, and Inception, an enhanced multi-scale residual block is introduced, which is applied to both the image and gradient branches. This design allows the model to capture multi-scale information more effectively. The discriminator employs the WGAN-GP method to enhance the stability of network training. In comparison to perceptually-driven algorithms like SRGAN, ESRGAN, and NatSR, the proposed algorithm effectively mitigates structural distortion and improves the overall quality of the generated images. Furthermore, the model's computational complexity is 23.7 GFLOPs, which is approximately one-quarter and one-tenth that of ESRGAN and SPSR, respectively.