PConvSRGAN: Real-world super-resolution reconstruction with pure convolutional networks

Abstract

Image super-resolution (SR) reconstruction technology faces numerous challenges in real-world applications: image degradation types are diverse, complex, and unknown; the diversity of imaging devices increases the complexity of image degradation in the super-resolution reconstruction process; SR requires substantial computational resources, especially with the latest significantly effective Transformer-based SR methods. To address these issues, we improved the ESRGAN model by implementing the following: first, a probabilistic degradation model was added to simulate the degradation process, preventing overfitting to specific degradations; second, BiFPN was introduced in the generator to fuse multi-scale features; lastly, inspired by the ConvNeXt network, the discriminator was redesigned as a pure convolutional network built entirely from standard CNN modules, which matches Transformer performance across various aspects. Experimental results demonstrate that our approach achieves the best PI and LPIPS performance compared to state-of-the-art SR methods, with PSNR,SSIM and NIQE being on par. Visualization results show that our method not only generates natural SR images but also excels in restoring structures.