next up previous
Next: 9 Discussion and future Up: Inform. theory approach to Previous: 7 An Information Infinitesimal

8 Implementation

The algorithm has two main parts, the computation of a scale-space image expansion, and the application of the Kullback contrast measure to that expansion. Most of the implementation is a straightforward coding of the math. The part that needs special focus to achieve a fast algorithm is the convolution and two dimensional moving average used in several of the formulas. We approximate the convolutions by several moving averages. Our 2D moving average separates into two 1D, and can be efficiently computed on a parallel processor. Details of our implementation can be found in [18]. Running the algorithm on a tex2html_wrap_inline651 image takes about 20s on a Sparc 1 and just over 1s on an 8 processor Sparc center 2000.