Learning Lens Blur Fields.

Area of Science:

• Computer Vision
• Computational Photography
• Optics

Background:

• Optical blur is a fundamental challenge in modern camera systems due to complex optical elements.
• Existing methods struggle to accurately model the variations in blur across different image locations and camera settings.

Purpose of the Study:

• To introduce a novel high-dimensional neural representation, the lens blur field, for modeling optical blur.
• To develop a practical method for acquiring and representing lens blur parametrically as a single, sensor-specific function.
• To enable accurate modeling of blur variations including defocus, diffraction, aberration, and sensor-specific features.

Main Methods:

• Developed a multilayer perceptron (MLP) based lens blur field representation.
• Formulated a generalized non-blind deconvolution problem to optimize MLP weights.
• Utilized focal stacks as input for learning the real-world blur field of a device.
• Created a dataset of 5D blur fields for various camera systems, including smartphones.

Main Results:

• The lens blur field accurately captures 2D point spread function variations with image plane location and focus.
• The representation models combined optical effects (defocus, diffraction, aberration) and sensor features (color filters, micro-lenses).
• Acquired 5D blur fields demonstrated sufficient expressiveness and accuracy to differentiate optical behavior between identical smartphone models.

Conclusions:

• The lens blur field offers a powerful and practical approach to modeling complex optical blur in cameras.
• This representation enables precise characterization of lens performance and sensor interactions.
• The findings reveal previously unobservable optical distinctions in mass-produced devices, opening avenues for quality control and optical design.