Rapid adaptive segmentation of images using successive cost processing

Arturo Aureliano Rodriguez, Purdue University

Abstract

The effectiveness of most vision systems depends on their ability to segment input images meaningfully. Image segmentation depends on decisions based on one or more thresholds. If thresholds are preselected, either by empirical analysis or teaching the operating environment, the system becomes constrained to perform in a specific environment. Typically, fast segmentation methods that determine thresholds automatically are confined to limited domains of applicability by their underlying assumptions, whereas methods capable of performing on broader domains suffer from high computational costs. The motivation behind developing novel segmentation methods that execute fast and adapt automatically to any given scene is to increase vision system versatility. Algorithms introduced in this thesis are fast, automatic, and robust. They are considered fast since they first exploit an inexpensive form of segmentation, thresholding, and they identify image regions requiring further processing with costlier techniques. Time-consuming parts are amendable to parallel architecture implementation, and good results can be obtained in low resolution (or coarsely sampled) image data. Automatic operation for a range of applications is attained by self-adjusting to and throughout the current image independent of scene knowledge. The algorithms are considered robust since they can perform in scenes of different contents, illumination, and background, under different backgrounds in the same scene, and do not require object surfaces to be homogeneous. The potential for different applications and validity of the algorithms is demonstrated in scenes exhibiting these properties. After discussing most segmentation methods throughout the literature and problems that prevent them from attaining the proposed objectives, the adaptive thresholding methods and methods for resolving uncertainty regions with costlier techniques are introduced. The methods incorporate several techniques that simulate aspects of human perception. Rectangular image regions are sampled to track background variations. Based on statistical theory, region homogeneity is determined. Background gray-tone distributions are estimated from local statistical measurements. Images are mapped into multiple labels corresponding to pixel brightness relative to background. Uncertainty is resolved using local and spatial information (edge magnitude and orientation). The threshold to generate the edge map is determined automatically from background labelled pixels in homogeneous regions.

Degree

Ph.D.

Advisors

Mitchell, Purdue University.

Subject Area

Electrical engineering

COinS