All-day thin-lens computational imaging with scene-specific learning recovery.

Area of Science:

• Optics and Photonics
• Computer Vision
• Computational Imaging

Background:

• Traditional imaging optics achieve high quality but result in bulky devices, hindering portability.
• Advancements in neural networks offer potential for thin-lens imaging, but challenges like artifacts and chromatic aberration persist.

Purpose of the Study:

• To develop a photorealistic thin-lens imaging system for practical applications.
• To address limitations of current thin-lens imaging techniques, focusing on artifact reduction and all-day usability.

Main Methods:

• A scene-specific generative adversarial network (GAN) learning strategy was developed.
• An integrated automatic acquisition and processing pipeline was created.
• A chromatic aberration pre-correction technique was implemented to minimize color fringe artifacts.

Main Results:

• The proposed method significantly reduces color fringe artifacts.
• The system demonstrates superior visual perception compared to existing thin-lens imaging approaches.
• Effective performance was achieved in both normal and low-light photography scenarios.

Conclusions:

• The developed thin-lens imaging strategy enables photorealistic image capture with a compact form factor.
• The scene-specific GAN and integrated pipeline offer a robust solution for portable, high-quality imaging.
• This work advances the feasibility of thin-lens imaging for real-world applications.