Towards Blind Lens Aberration Correction via Large LensLib Pre-training and Discrete Degradation Priors
Researchers have developed FoundCAC, a new framework for blind lens aberration correction that utilizes large-scale pre-training and discrete degradation priors. The system improves data scalability by constructing a diverse lens library called AODLibpro and employs a multi-stage vector-quantized representation learning scheme to encode Point Spread Functions into a discrete prior. This approach enables state-of-the-art zero-shot generalization and efficient few-shot adaptation for correcting optical degradations in both synthetic and real-world lenses. AI
IMPACT Enhances image restoration capabilities, potentially improving performance in photography and computer vision applications.