Modeling and Analysis of Optical Blur for Everyday Photography

Optical blur obscures image detail due to the camera's optics. The blur appearance depends on the scene's depth, wavelength, as well as the camera. The inference of such information --- optical calibration, image deblurring, and depth estimation -- is of great interest to computer vision. Existing approaches make impractical assumptions about the camera and the scene, and are therefore restricted to lab settings. In this thesis, we propose novel optical blur models towards solving the above inference problems in everyday settings. First, we propose a practical model for lenses with monochromatic optical aberrations, assuming the aberrations can be fully described by Seidel aberrations. Although the model produces realistic 2D blur kernels of complicated appearance, it has very few parameters and thus can be estimated using very limited data. We show specifically that just three blur kernels contain enough information to estimate the lens parameters up to an aperture-dependent scale factor. Second, we study image deblurring for RGB-Infrared imaging. This problem is especially challenging since RGB-Infrared cameras suffer from inevitable chromatic aberration, channel crosstalk, and pixel multiplexing due to the color-filter-array. The coupling of the distortions are critical to deblurring quality but is non-trivial to invert efficiently. We propose a practical image formation model that captures all three distortions while enables fast restoration. Third, we develop an approach to two-frame Depth from Defocus (DFD) in everyday settings. The main challenge here is that the inter-frame motion needs to be estimated in addition to depth from just two images. We propose conditions to control the camera focus to guarantee near-optimal DFD estimation and a method to estimate depth and the motion jointly in local 9x9 patches. This allows us to capture sparse depth map even in non-rigid scenes. Fourth, we propose a method to fuse local DFD results into dense depth and flow maps. The noise, sparsity and ambiguities inherent in DFD makes this problem particularly difficult. We introduce a splined-based prior to enable fusion over great image distances and irregular shaped regions, and to allow explicit reasoning about figure-ground relationships. Our final results exhibits high quality results with crisp boundaries.

Department dc:contributor.department

Computer Science

Year dc:date.issued

2018

Author dc:creator

• Tang, Huixuan

Advisor dc:contributor.advisor

• Kutulakos, Kiriakos N.

• aberration
• calibration
• depth from defocus
• image restoration
• optical blur

Handle dc:identifier.uri

http://hdl.handle.net/1807/89662

OAI identifier oai:identifier

oai:utoronto.scholaris.ca:1807/89662