A large collection of digital images of natural scenes provides a database for
analyzing and modeling small scene patches (e.g., 2 x 2) referred to as natural microimages. A pivotal ¯nding is the stability of the empirical microimage distribution across scene samples and with respect to scaling. With a view toward potential applications (e.g. classi¯cation, clutter modeling, segmentation), we present a hierarchy of microimage probability models which capture essential local image statistics. Tools from information theory, algebraic geometry and of course statistical hypothesis testing are employed to assess the "match" between candidate models and the empirical distribution. Geometric symmetries play a key role in the model selection process.
One central result is that the microimage distribution exhibits reflection and
rotation symmetry and is wellrepresented by a Gibbs law with only pairwise
interactions. However, the acceptance of the updown reflection symmetry hypothesis is borderline and intensity inversion symmetry is rejected. Finally, possible extensions to larger patches via entropy maximization and to patch classification via vector quantization are briefly discussed.
Original language  English 

Qualification  Ph.D. 

Awarding Institution   University of Massachusetts Amherst


Supervisors/Advisors   Geman, Donald, Supervisor, External person


Award date  1 Sept 2000 

Publisher  

Print ISBNs  9780599957473, 0599957476 

Publication status  Published  2000 
