120 likes | 265 Views
High Dynamic Range Image Reconstruction from Hand-held Cameras. Pei-Ying Lu Tz-Huan Huang Meng -Sung Wu Yi-Ting Cheng Yung-Yu Chuang National Taiwan University CVPR ’09 Reporter : A nnie Lin. Outline . Introduction Algorithm Results and comparison . Introduction .
E N D
High Dynamic Range Image Reconstruction from Hand-held Cameras Pei-Ying Lu Tz-Huan Huang Meng-Sung Wu Yi-Ting Cheng Yung-Yu Chuang National Taiwan University CVPR ’09 Reporter : Annie Lin
Outline • Introduction • Algorithm • Results and comparison
Introduction • Reconstructing a high-quality high dynamic range (HDR) image from a set of differently exposed and possibly blurred images taken with a hand-held camera • Input : a series captured images • Bayesian framework to formulate the problem and apply a maximum likelihood approach to iteratively perform blur kernel estimation ,HDR reconstruction and camera curve recovery • Goal: to find the motion blur kernel, irradianceand response function
Related work • Camera pipeline • HDR • Image deblurring • Assuming the blur kernel is shift-invariant
Algorithm • Image alignment • The Bayesian framework • Optimization • Tikhonov regularization • H. W. Engl, M. Hanke, and A. Neubauer. Regularization of Inverse Problems. Kluwer Academic, 2000. (book) • Three parameters: • K (motion blur kernel) • E (irradiance ) • g ( response function)
The Bayesian framework • Z: captured image • f: camera response function (unknown) • E: irradiance (unknown) • K: motion blur kernel (unknown) • Overdetermined • the problem is then turned into "a maximum likelihood” problem
Optimization • Initialization • Ki = δ • E [2]Recovering high dynamic range radiance maps from photographs. • G a linear mapping from the pixel values to the irradiance values • Tikhonov regularization (ill-posed) • Optimizing Ki • Optimizing E • Optimizing g • To iteratively update the data : • [2] Landweber method “H. W. Engl Regularization of Inverse Problems”
Optimizing Ki • The iteration stops when the change between two steps is sufficiently small
Optimizing E • After optimization, scaling up the reconstructed irradiance E to keep its values in a similar scale as the initialized irradiance • scaling factor :the ratio of mean values, mean(E initial)=mean(E). • Optimizing g • Add smoothness turn to ensure the function g is smooth • λ3 weighted: