Real-World Super-Resolution via Multi-degradation Based Texture Feature Reconstruction

In the field of real-world image super-resolution(SR), collecting paired training data is extremely challenging. As a result, some SR methods are trained on synthetic datasets. But there remains a gap when applied to real-world data. To address this issue, this paper proposes a robust multi-degradation real-world reconstruction method for real-world SR based on a texture degradation reconstruction network. This method cleverly overcomes the inadaptability issue associated with solely using synthetic datasets. Specifically, we first train the encoder, decoder, and the proposed SR network using real data. The encoder learns to extract the genuine texture features of input images through training, while the decoder significantly enhances the model's adaptability by incorporating the multi-degradation information. Then, leveraging the trained encoder, decoder, and real LR data, we retrain the SR network. Through this approach, our method can be trained in the absence of paired data and can improve the SR network's adaptability to LR images with unknown degradation. Experimental results indicate that this method outperforms existing real-world SR methods in generating SR outputs with finer textures.