H dli N i i Si l I D bl i i Di ti l FiltHandling Noise in Single Image Deblurring using Directional FiltersHandling Noise in Single Image Deblurring using Directional FiltersHandling Noise in Single Image Deblurring using Directional FiltersHandling Noise in Single Image Deblurring using Directional Filtersg g g g g(CVPR 2013)(CVPR 2013)(CVPR 2013)(CVPR 2013)( )

i Zh 1 S h Ch 2 Di i i M 1 S l i P i 2 J W 2Lin Zhong1 Sunghyun Cho2 Dimitris Metaxas1 Sylvain Paris2 Jue Wang2Lin Zhong1 Sunghyun Cho2 Dimitris Metaxas1 Sylvain Paris2 Jue Wang2Lin Zhong Sunghyun Cho Dimitris Metaxas Sylvain Paris Jue Wang1CBIM Rutgers University 2Adobe Research1CBIM Rutgers University 2Adobe ResearchCBIM, Rutgers University Adobe Research, g y

Introduction Di ti l Filt E i tIntroduction Directional Filter ExperimentsIntroduction Directional Filter ExperimentsDirectional Filter Experiments

S th ti d tG l Synthetic data Comparison with other methodsGoal: Estimate a high q alit bl r kernel and latent Synthetic data Comparison with other methodsGoal: Estimate a high-quality blur kernel and latent Synthetic data pGoal: Estimate a high quality blur kernel and latent yg q yC i ith T i d Li [CVPR 2012]image from a blurry and noisy input image Comparison with Tai and Lin [CVPR 2012]image from a blurry and noisy input image Comparison with Tai and Lin [CVPR 2012]image from a blurry and noisy input image.g y y p g

M ti tiMotivation: handheld camera + low lightMotivation: handheld camera + low-lightMotivation: handheld camera + low lightg blurry and noisy images blurry and noisy images blurry and noisy imagesy y g

Directional filters reduce noise while keeping blur information Abbey (5% noise) Chalet (5% noise) Aque (5% noise)Directional filters reduce noise while keeping blur information Abbey (5% noise) Chalet (5% noise) Aque (5% noise)Directional filters reduce noise while keeping blur information intact in their orthogonal directionsintact in their orthogonal directionsintact in their orthogonal directions

N i i D bl iNoise in Deblurring Kernel EstimationNoise in Deblurring Kernel EstimationNoise in Deblurring Kernel EstimationgS1 Di ti l P i th d iti t i S1: Directional Previous methods are sensitive to noise S1: Directional Previous methods are sensitive to noise filtS1e ous et ods a e se s t e to o se

Noisy Input filterS1Noisy Input filterAbb (10% i )Abbey (10% noise) Chalet (10% noise) Aque (10% noise)bbey ( 0% o se) Chalet (10% noise) Aque (10% noise)

Real examplesS2 S2 1D k l Real examplesS2 S2: 1D kernel Real examplesS2 S2: 1D kernel ea e a p esestimationestimationestimation

Bl I t [Ch &L 2009] [L i t l 2011] O th dBlurry Input [Cho&Lee 2009] [Levin et al. 2011] Our methody p(5% i )

[ ] [ ](5% noise) S3 S3: 2D kernel( ) S3 S3: 2D kernel

Denoising as preprocessing destroys blur information Denoising as preprocessing destroys blur information reconstruction Denoising as preprocessing destroys blur information reconstruction(i d t f )(inverse radon transform)(inverse radon transform)( )

Fi l bli d d l tiFinal non blind deconvolutionFinal non-blind deconvolutionFinal non blind deconvolution G t hi h lit lt ith i ti t d Generate a high quality result with a given estimated Generate a high-quality result with a given estimated Generate a high quality result with a given estimated k l (T t it ti ti i ti )kernel (Two-step iterative optimization)kernel. (Two-step iterative optimization)( p p )

* NLM() l l d i i* NLM(): non-local means denoising NLM(): non-local means denoising() g