Task-driven lens design.

Area of Science:

• Optical Engineering
• Computer Vision
• Computational Imaging

Background:

• Classical lens design prioritizes aberration minimization for sharp images, but operates independently of computer vision tasks.
• Current end-to-end optical design methods integrate optics and networks but face training instability.
• A gap exists in designing optical systems optimized for specific downstream machine vision applications.

Purpose of the Study:

• Introduce task-driven lens design as a novel optimization philosophy for integrated optics-network systems.
• Develop a method for designing lenses that are optimized for the feature preferences of pretrained vision models.
• Explore a broader design space for optical lenses through automated, human-intervention-free optimization.

Main Methods:

• Implement a network-frozen optimization strategy, keeping the pretrained vision model fixed.
• Optimize only the lens parameters to align image formation with the vision model's feature preferences.
• Conduct comparative experiments using multiple computer vision tasks to evaluate lens performance.

Main Results:

• Task-driven lenses (TaskLenses) demonstrate superior performance compared to classical ImagingLenses, even with fewer optical elements.
• The optimization process in a network-frozen setting is low-dimensional and stable, allowing design from scratch.
• Learned optics exhibit long-tailed point spread functions, effectively preserving structural cues during aberration correction.

Conclusions:

• Task-driven lens design offers a practical approach for creating optical lenses compatible with modern vision models.
• This method inspires new optical design objectives beyond traditional aberration minimization.
• Automated, task-specific lens design opens new avenues for optimizing imaging systems for AI applications.