Velocity-Dependent Shutter Sequences for Motion Deblurring

We extend the flutter shutter method of Raskar et al. [1] to fast-moving objects by first demonstrating that no coded exposure sequence yields an invertible point spread function for all velocities. Thus, the shutter sequence must depend on object velocity, and we propose a method to compute such velocity-dependent sequences. We demonstrate improved image quality from velocity dependent sequences on fast moving objects. Increasing Velocity D ebl ur re d C ap tu re d D ebl ur re d C ap tu re d Velocity-Dependence The objective of exposure coding is to choose a shutter timing sequence that produces an invertible Point Spread Function (PSF), i.e. one whose MTF does not go to zero. But, the effective PSF depends on both the shutter sequence and the object’s velocity. For a given sequence, how does reconstructed image quality vary with object velocity? Lemma: If a given flutter shutter sequence S produces an invertible PSF for velocity V, inversion of the PSF produced by S under velocity ≥ 2V will be ill-posed at the Nyquist frequency. Proof: In the paper. Implication: Shutter sequences need to be designed for specific object velocities. M TF SummaryWe characterize the performance of shutter timingsequences for coded exposure de-blurring, and show that aunique timing sequence is needed for different objectvelocities. We also demonstrate and characterize a previouslyundocumented noise effect in coded exposure withelectronic shutters. Based on these findings, we present analgorithm to find velocity-dependent sequences anddemonstrate their utility on real images. Key References[1] Ramesh Raskar, Amit Agrawal, and Jack Tumblin. CodedExposure Photography: Motion Deblurring using a FlutteredShutter. In Proceedings of SIGGRAPH 2006.Experimental Results OursRealPSFRaskaretal.[1].Shutter Sequence Design: Noise When using an electronic shutter (e.g., Point GreyFlea/Dragonfly) to capture a coded exposure image, read-outnoise is proportional to the number of open shutter periods(blue curve). This noise is amplified by de-blurring (red curve).Implication: When designing sequences for electronicshutter cameras, favor those with fewer open shutter periods. Shutter Sequence Design: AlgorithmWe represent a shutter sequence as a combination of:(1) A number of open shutter durations and(2) the start times of each open shutter duration. We note that (1)imposes an upperbound on the MTF. Algorithm Overview:Step 1: Enumerate all partitions of exposure time into openshutter durations.Step 2: Rank each partition by the upper bound of the MTF.Step 3: Rank order search, determining start times until...End condition: The upper bound MTF of the next bestpartition is worse than the current best MTF.Full algorithm details can be found in the paper.