HP INCFollow


This disclosure defines a new strategy to enhance image transformation approaches for image scale-up and super-resolution (SR) using a federated learning architecture to augment image resolution. The main idea relies on using FL which allows the tunning of learning algorithms in a decentralized way without the necessity of any sensitive data leaving the client’s device. Hence, it is possible to train models to adjust image resolution directly on customer devices, using locally available information to learn the patterns most relevant to the desired task in an self-supervised and privacy-guaranteed way. FL can optimize models indirectly by using the info from high-resolution users (i.e., users that contain high-resolution images) without the need for labeling and storage on external servers. Moreover, FL reduces the need for specialized hardware for training and maintaining models while improving the result using accurate data related to the application. Labels should be self-created from the high-resolution data available. This way, low-resolution clients (i.e., users that contain low-resolution images) can benefit from the shared models trained in other clients' devices for super-resolution in the same task. Yet, high-resolution clients can maintain good accuracy using the models when adjusting the local data to a lower resolution (e.g., due to changes in the network or low battery status).

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution-Share Alike 4.0 License.