Wednesday, August 05, 2009

Salience as Local Entropy

Based on a suggestion by Romann Weber, regarding an alternative definition of visual salience.

  1. There is an image I comprised of integers in the range R
  2. The probability of a specific value c from R occurring in I, PrI(c), equals the number of pixels with the value c in I divided by the total number of pixels in I
  3. There is an image S that is a sub region within I
  4. The entropy of S with respect to I is equal to the negative sum of PrI(p) * log(PrI(c)) for every pixel p in S
Using this metric, an entire image may be analyzed by moving the sub region S about the image. Also consider different sizes of S, and the possibility of use both a moving S and I within the actual image.

No comments: