Image and video segmentation using unsupervised classification in a Bayesian set up

Srinivas Sista, Purdue University

Abstract

Segmentation of images and video requires obtaining meaningful regions and their descriptions from the available data. In this thesis, the segmentation problem is posed as an unsupervised classification problem, i.e., pixels have to be assigned to various classes without a priori knowledge about which class the individual pixels arise from. Different segments that have the same description are considered as arising from the same class even if they are far apart. A solution is proposed that iteratively computes the MAP estimates of both the class description parameters and the data partition in a small number of iterations. The proposed method starts with an initial partition, estimates the class description parameters and refines the partition using the estimated parameters. These two steps are repeated until there can be no further refinement in the partition. The chosen formulation leads to a parameter estimation scheme that computes the estimates directly from the data and the partition information. Also, a new, refined partition is obtained from the estimated parameters by using a simple decision rule that uses a statistic of the pixel data. This iterative procedure is shown to converge in a finite number of steps. Experimental results on synthetic and real data are presented. A generalized version of the unsupervised classification method for multidimensional data is also developed. Discussion is given on how this method can be applied to problems arising in other fields and illustrated with some numerical examples.

Degree

Ph.D.

Advisors

Kashyap, Purdue University.

Subject Area

Electrical engineering

COinS