Statistical Structure of Natural 4 × 4 Image Patches
The direct computation of natural image block statistics is unfeasible due to the huge domain space. In this paper we shall propose a procedure to collect block statistics on compressed versions of natural 4 × 4 patches. If the reconstructed blocks are close enough to the original ones, these statistics can clearly be quite representative of the true natural patch statistics. We shall work with a fractal image compression–inspired codebook scheme, in which we will compute for each block B its contrast σ, brightness μ and a normalized codebook approximation D B of (B - μ)/σ. Entropy and mutual information estimates suggest a first order approximation p(B)≃ p(D B )p(μ)p(σ) of the probabibility p(B) of a given natural block, while a more precise approximation can be written as p(B)≃ p(D B )p(μ)p(σ)Φ(||∇B||). We shall also study the structure of p(σ) and p(D), the more relevant probability components. The first one presents an exponential behavior for non flat patches, while p(D) behaves uniformly with respecto to volume in patch space.
KeywordsMutual Information Natural Image Image Patch Lena Image Joint Entropy
- 1.Fisher, Y. (ed.): Fractal Image Compression: Theory and Application. Springer, New York (1995)Google Scholar
- 4.Lee, A.B., Pedersen, K.S., Mumford, D.: The Complex Statistics of High-Contrast Patches in Natural Images. In: WWW Proceedings of Second International IEEE Workshop on Statistical and Computational Theories of Vision, Vancouver, Canada (July 2001)Google Scholar
- 5.Marr, D.: Vision. W.H. Freeman and Co., New York (1982)Google Scholar
- 10.Weinberger, M., Seroussi, G., Sapiro, G.: LOCO-I: A Low Complexity, Context- Based, Lossless Image Compression Algorithm. In: Proc. of the IEEE Data Compression Conference, Snowbird, Utah (March-April 1996)Google Scholar