BIGPrior is published in IEEE Transactions on Image Processing
Our BIGPrior paper presents a learning-based restoration framework that forms a generalization of various families of classical methods. It is both tightly connected with Bayesian estimation upon which it builds, and also to classical dictionary methods. BIGPrior makes the explicit integration of learned-network priors possible, notably a generative-network prior. Its biggest advantage is that, by decoupling data fidelity and prior hallucination, it structurally provides a per pixel fusion metric that determines the contribution of each. This can be important both for end users and for various downstream applications. We hope this work will foster future learning methods with clearly decoupled network hallucinations, both for interpretability, reliability, and to safeguard against the hazards of black-box restoration.
Helou, M.E., & Süsstrunk, S. (2020). BIGPrior: Towards Decoupling Learned Prior Hallucination and Data Fidelity in Image Restoration. ArXiv, abs/2011.01406.