Title: Advancing Blind Lens Aberration Correction through Large-Scale LensLib Pre-training and Discrete Degradation Priors

Abstract: The advent of deep-learning-based lens library pre-training (LensLib-PT) pipelines presents a novel pathway for correcting lens aberrations without prior knowledge. By training a universal neural network, these approaches have shown robust potential in managing a wide array of unknown optical degradations. This study introduces FoundCAC, a foundational universal framework designed to overcome two primary obstacles that limit the generalization of current pipelines: the scarcity of scalable training data and the lack of prior information that characterizes optical degradation.

To address data scalability, we broaden the design specifications to enhance the diversity of degradations. Consequently, we developed AODLibpro, a large-scale, unbiased lens library constructed using a uniform sampling strategy. This library effectively quantifies both spatial-variation patterns and degradation severity. Regarding model architecture, we aim to utilize Point Spread Functions (PSFs) as guidance while preserving the blind restoration paradigm. To achieve this, we propose a multi-stage vector-quantized representation learning scheme. This approach is tailored to create a Latent PSF Representation (LPR), which explicitly encodes complex, continuous PSFs into a discrete degradation prior. This prior serves to regularize the highly ill-posed restoration process.

By employing a straightforward yet effective codebook-freezing strategy, our framework utilizes this discrete prior to boost restoration performance across full-shot scenarios and facilitate highly efficient few-shot adaptation for previously unseen lenses. Experimental results on various aberrations from both synthetic LensLib datasets and real-world lenses confirm that our framework achieves state-of-the-art zero-shot generalization capabilities. Furthermore, it enables efficient few-shot adaptation for specific lenses. The source code and datasets will be publicly accessible at https://github.com/zju-jiangqi/FoundCAC.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC