Image Restoration via Multi-prior Collaboration

This paper proposes a novel multi-prior collaboration framework for image restoration. Different from traditional non-reference image restoration methods, a big reference image set is adopted to provide the references and predictions of different popular prior models and accordingly further guide the subsequent multi-prior collaboration. In particular, the collaboration of multi-prior models is mathematically formulated as a ridge regression problem. Due to expensive computation complexity of handling big reference data, scatter-matrix-based kernel ridge regression is proposed, which achieves high accuracy while low complexity. Additionally, an iterative pursuit is further proposed to obtain refined and robust restoration results. Five popular prior methods are applied to evaluate the effectiveness of the proposed multi-prior collaboration framework. Compared with the state-of-the-art image restoration approaches, the proposed framework improves the restoration performance significantly.