In this talk I will introduce an algorithm for computing the covariance matrix for all image sub-patches from a covariance matrix built from larger patches. This method allows for the linear statistics of all patches smaller than a given size to be efficiently computed on a very large collection of images while visiting each image once. I will show the structure of the principal components of 32x32 patches over a very large collection of images, and also introduce two possible applications for these multi-scale statistics: image super-resolution and image salience.
Last modified: Monday, 24-Sep-2012 12:17:25 NZST
This page is maintained by the seminar list administrator.