MoE-ESRGAN: A Super-Resolution Network for Multi-degraded Chinese Painting Images
Mingxu Wang, Yue Zhou, Kun Xu · 2023
The conventional super-resolution algorithms are unable to fully address the Chinese painting images that contain some artifacts and were taken many years ago. In this paper, we propose a novel SR model named MoE-ESRGAN for restoring multi-degraded Chinese painting images based on the structure of ESRGAN. To simulate the degradation of real-world Chinese painting images, this paper uses multi-degradation model, including bicubic downsampling, blur and noise, to obtain low-resolution images as our dataset. This paper further designs a new generator network by incorporates a set of Mixture-of-Experts (MoE) to enhance the network’s ability for mapping from low-resolution (LR) images to high-resolution (HR) images. Meanwhile, we assign the specific guidance to each expert in MoE to learn different image features. Based on the self-built Chinese painting image dataset, we compare the performance of our model with various models through experiments. Experimental results demonstrate that, compared with ESRGAN, our model has an average increase of 0.7 dB in peak signal-to-noise ratio (PSNR) and an average increase of 0.03 in structural similarity (SSIM). Meanwhile, our MoE-ESRGAN model exhibits high effectiveness in image denoising and deblurring. Furthermore, the reconstructed images of our model are subjectively more in line with human visual standards.