A Flexible Framework for Universal Computational Aberration Correction Via Automatic Lens Library Generation and Domain Adaptation
Qi Jiang,Yao Gao,Shaohua Gao,Zhonghua Yi,Lei Sun,Hao Shi,Kailun Yang,Kaiwei Wang,Jian Bai
2024-01-01
Abstract:Emerging universal Computational Aberration Correction (CAC) paradigms provide an inspiring solution to light-weight and high-quality imaging without repeated data preparation and model training to accommodate new lens designs. However, the training databases in these approaches, i.e., the lens libraries (LensLibs), suffer from their limited coverage of real-world aberration behaviors. In this work, we set up an OmniLens framework for universal CAC, considering both the generalization ability and flexibility. OmniLens extends the idea of universal CAC to a broader concept, where a base model is trained for three cases, including zero-shot CAC with the pre-trained model, few-shot CAC with a little lens-specific data for fine-tuning, and domain adaptive CAC using domain adaptation for lens-descriptions-unknown lens. In terms of OmniLens's data foundation, we first propose an Evolution-based Automatic Optical Design (EAOD) pipeline to construct LensLib automatically, coined AODLib, whose diversity is enriched by an evolution framework, with comprehensive constraints and a hybrid optimization strategy for achieving realistic aberration behaviors. For network design, we introduce the guidance of high-quality codebook priors to facilitate zero-shot CAC and few-shot CAC, which enhances the model's generalization ability, while also boosting its convergence in a few-shot case. Furthermore, based on the statistical observation of dark channel priors in optical degradation, we design an unsupervised regularization term to adapt the base model to the target descriptions-unknown lens using its aberration images without ground truth. We validate OmniLens on 4 manually designed low-end lenses with various structures and aberration behaviors. Remarkably, the base model trained on AODLib exhibits strong generalization capabilities, achieving 97 performance in a zero-shot setting.