Massachusetts Institute of Technology
One of the long-standing challenges in photography is motion blur. Blur artifacts are generated from relative motion between a camera and a scene during exposure. While blur can be reduced by using a shorter exposure, this comes at an unavoidable trade-off with increased noise. Therefore, it is desirable to remove blur computationally.
To remove blur, we need to (i) estimate how the image is blurred (i.e. the blur kernel or the point-spread function) and (ii) restore a natural looking image through deconvolution. Blur kernel estimation is challenging because the algorithm needs to distinguish the correct image--blur pair from incorrect ones that can also adequately explain the blurred image. Deconvolution is also difficult because the algorithm needs to restore high frequency image contents attenuated by blur. In this dissertation, we address a few aspects of these challenges.
We introduce an insight that a blur kernel can be estimated by analyzing edges in a blurred photograph. Edge profiles in a blurred image encode projections of the blur kernel, from which we can recover the blur using the inverse Radon transform. This method is computationally attractive and is well suited to images with many edges. Blurred edge profiles can also serve as additional cues for existing kernel estimation algorithms. We introduce a method to integrate this information into a maximum-a-posteriori kernel estimation framework, and show its benefits.
Deconvolution algorithms restore information attenuated by blur using an image prior that exploits a heavy-tailed gradient profile of natural images. We show, however, that such a sparse prior does not accurately model textures, thereby degrading texture renditions in restored images. To address this issue, we introduce a content-aware image prior that adapts its characteristics to local textures. The adapted image prior improves the quality of textures in restored images. Sometimes even the content-aware image prior may be insufficient for restoring rich textures. This issue can be addressed by matching the restored image's gradient distribution to its original image's gradient distribution, which is estimated directly from the blurred image. This new image deconvolution technique called iterative distribution reweighting (IDR) improves the visual realism of reconstructed images.
Subject motion can also cause blur. Removing subject motion blur is especially challenging because the blur is often spatially variant. In this dissertation, we address a restricted class of subject motion blur: the subject moves at a constant velocity locally. We design a new computational camera that improves the local motion estimation and, at the same time, reduces the image information loss due to blur.
Last update: Aug. 23. 2010