400 likes | 413 Views
Explore the world of superresolution imaging, covering Nyquist sampling, sub-Nyquist sampling, multichannel blind deconvolution, resolution enhancement, and more. Learn about realistic superresolution techniques, acquisition models, optimization methods, space-variant cases, camera-motion blur, and infrared video super-resolution. Discover how superresolution can be applied to handheld IR cameras, smartphones using Wiener filtering, and collaboration with esteemed researchers. Join us in advancing the field of imaging technology!
E N D
Superresolution Imaging from Equations to Mobile Applications Filip Šroubek sroubekf@utia.cas.cz Institute of Information Theory and Automation Academy of Sciences of the Czech Republic www.utia.cas.cz NIPS Workshop 2011 Sierra Nevada
Nyquist Sampling • Nyquist sampling an "ideal" image, no frequencies are missed. • Superresolution = scaling + interpolation
Superresolution • Sub-Nyquist sampling aliasing high frequencies are lost or modified • Superresolution can recover aliased data interpolation SR
pixel interpolation superresolution Traditional superresolution sub-pixel shift
Realistic Superresolution Deconvolution
noise channel K channel 2 + channel 1 acquired images D[u* hk](x) + nk(x) = zk(x) Multichannel Acquisition Model original image u(x)
Realistic superresolution • Acquisition model: • Noise • Blur • Shift • Downsampling • Reconstruction method: • Multichannel blind deconvolution • Shift compensation • Resolution enhancement
Image regularization term L1 norm Blur Regularization term positivity Data term L2 norm Superresolution & Blind Deconv. • Acquisition model • Optimization problem
Alternating Minimization • u-step • h-step
Variable Splitting • u-step • h-step
Augmented Lagrangian Method • u-step • h-step
rough registration Optical zoom (ground truth) Superresolved image (2x) Superresolution SIAM Imaging Science 2010
Space-variant Case • Video with local motion • Masking • Slowly changing PSFs and/or misregistered images • Patch-wise approach
interpolated SR
interpolated SR t-2 t-1 t t+1 t+2
interpolated SR SR + masking t-2 t-1 t t+1 t+2
Close-up Original Input LR Space-invariant Reconstruction Space-variant Reconstruction
Infrared video super-resolution • Handheld IR camera • 160 x 120, 9 fps • Real-time super-resolution • with factor 2 (320 x 240) • computed directly inside the camera using DSP
LR input SR output
“Blind” Deconvolution on Smartphones Using accelerometers and/or gyroscopes Rotation and translation of the phone
Thank You… collaborators: Jan Flusser, Michal Šorel, Jan Kamenický, Peyman Milanfar
z1 = u h1 h2 u = z2 * * z1 h2 = u h1 h2 h1 h2 u = h1 z2 * * * * * * 0 PSF Regularization u
Incorporating a between-image shift original image PSF degraded image