Image restoration methods for simple optical systems based on deep learning.

Area of Science:

• Optical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Traditional optical design relies on MTF and spot diagrams, often resulting in complex, heavy lens systems.
• Simple optical systems offer size and weight benefits but suffer from significant aberrations and poor imaging quality.
• Deep learning presents a potential solution for aberration correction without adding structural complexity.

Purpose of the Study:

• To develop a lightweight, high-quality optical imaging system using deep learning.
• To compensate for aberrations in simple optical systems to improve imaging performance.

Main Methods:

• Construction of a MIMO-UNet restoration framework incorporating coordinate attention and deformable convolution.
• Application of the framework to a 100 mm doublet optical system for aberration correction.
• Comparative analysis against a Double Gauss system to evaluate performance.

Main Results:

• The enhanced MIMO-UNet model achieved an average Peak Signal-to-Noise Ratio (PSNR) of 35.42 dB in a 100 mm doublet system.
• The deep learning method significantly improved imaging quality in a simple system with fewer lenses.
• Imaging results from the simplified system approached the quality of a complex Double Gauss system.

Conclusions:

• The proposed deep learning framework effectively addresses aberration issues in simple optical systems.
• This method enables high-quality imaging with lightweight optical designs, outperforming traditional approaches.
• The MIMO-UNet restoration framework offers a promising solution for advanced optical system design.