Exploiting Degradation Prior for Personalized Federated Learning in Real-World Image Super-Resolution

In this study, we introduce a novel personalized federated learning (pFL) framework for real-world image super-resolution (SR) tasks, named p(2)FedSR, aiming to protect data privacy and construct the customized model conditioned on the client-specific degradation process. Concretely, we propose a local update with degradation prior at the client-side, which exploits the client-specific information (i.e., image degradation) learned from the customized degradation prior generator as prior knowledge, and along with input to facilitate an efficient local model update. Additionally, we propose a personalized aggregation policy on the server-side, which determines the uploaded model components based on their function. Instead of aggregating the entire models (e.g., FedAvg), the model components that have common representation (e.g., image texture) are shared with the server to augment the representation capability of the global model, while the remaining part with client-related information (e.g., semantic concept) is kept locally to ensure model personalization. Extensive experiments conducted on real-world image SR benchmarks demonstrate the superiority of our model in terms of image quality and model performance. Notably, p(2)FedSR can be seamlessly integrated with various prevalent SR methods, including CNN-based and Transformer-based architectures.