The purpose of image segmentation is to isolate objects in a scene from the background. This is a very important step in any computer vision system since various tasks, such as shape analysis and object recognition, require accurate image segmentation. Image segmentation can also produce tremendous data reduction. Edge-based and region-based segmentation have been examined and two new algorithms based on recent results in random field theory have been developed. The edge-based segmentation algorithm uses the pixel gray level intensity information to allocate object boundaries in two stages: edge enhancement, followed by edge linking. Edge enhancement is accomplished by maximum energy filters used in one-dimensional bandlimited signal analysis. The issue of optimum filter spatial support is analyzed for ideal edge models. Edge linking is performed by quantitative sequential search using the Stack algorithm. Two probabilistic search metrics are introduced and their optimality is proven and demonstrated on test as well as real scenes. Compared to other methods, this algorithm is shown to produce more accurate allocation of object boundaries. Region-based segmentation was modeled as a MAP estimation problem in which the actual (unknown) objects were estimated from the observed (known) image by a recursive classification algorithms. The observed image was modeled by an Autoregressive (AR) model whose parameters were estimated locally, and a Gibbs-Markov random field (GMRF) model was used to model the unknown scene. A computational study was conducted on images having various types of texture images. The issues of parameter estimation, neighborhood selection, and model orders were examined. It is concluded that the MAP approach for region segmentation generally works well on images having a large content of microtextures which can be properly modeled by both AR and GMRF models. On these texture images, second order AR and GMRF models were shown to be adequate.

Date of this Version