Exampled-based Super resolution Presenter: Yu-Wei Fan
Outline Introduction Training set generation Super-resolution algorithms – Idea – Markov Network – One-pass algorithm Results
Outline Introduction Training set generation Super-resolution algorithms – Idea – Markov Network – One-pass algorithm Results
Introduction Why do we need high resolution image? Usually, we cannot get high resolution image easy.
Introduction Aim: High Resolution Image – 1.Reduce the pixel size the amount of light available also decrease generates shot noise – 2.Increase the chip size increase capacitance difficult to speed up a charge transfer rate – 3.Signal processing techniques Low cost
Introduction General Super Resolution – Need multi frames information Exampled-based Super resolution –Need only one frame
Outline Introduction Training set generation Super-resolution algorithms – Idea – Markov Network – One-pass algorithm Results
Training set generation Store the high-resolution patch corresponding to every possible low-resolution image patch. Typically, these patches are 5 × 5 or 7 × 7 pixels.
Outline Introduction Training set generation Super-resolution algorithms – Idea – Markov Network – One-pass algorithm Results
Idea Unfortunately, that approach doesn’t work!
Markov Network
MAP Estimator:
Markov Network Example:
Markov Network Belief Propagation Where is from the previous iteration. The initial are 1. Typically, three or four iterations of the algorithm are sufficient.
One-pass algorithm How do we select a good patch pair? Two constraint: – frequency constraint – spatial constraint
One-pass algorithm
Outline Introduction Training set generation Super-resolution algorithms – Idea – Markov Network – One-pass algorithm Results
α=0
Results α=0.5
Results α=5