GAN-Based Super-Resolution With Enhanced Multi-Scale Laplacian Pyramid and Frequency Domain Loss

Super-resolution techniques play an important role in the fields of image processing and computer vision. However, existing super-resolution methods based on generative adversarial networks still exhibit significant shortcomings in recovering high-frequency details and effectively utilising multi-scale information. To address these issues, this paper proposes an improved generative adversarial network. Specifically, an enhanced multi-scale Laplacian pyramid structure is designed to capture and process image details at different scales. Then, convolutional operations are added to each layer of the pyramid to further improve the recovery of multi-scale details. Additionally, a frequency domain loss is introduced, where the generated and real images are transformed into the frequency domain using Fourier transforms for comparison. This method enhances the reconstruction of high-frequency details. The experiments are validated on four publicly available datasets and the results show that the proposed network significantly outperforms existing methods in both reconstruction quality and visual performance.