Prediction Driven Image Segmentation

Abstract

In any computer vision and image analysis system, segmentation is an integral part and is broadly defined as the partitioning of an image into separate regions. Each resulting region corresponds to a different object or area of interest. Typically, the partitioning is derived from specific constraints and the segmentation process uses these constraints to construct homogenous regions and smooth boundaries. Over the past 40 years, a multitude of methods and algorithms have been developed. Each of them attempts to solve the segmentation problem based on image properties, constraints derived from the application domain, or a combination. Most approaches are developed for a specific application and cannot be generalized for all images. In fact, no single algorithm can be considered good for all images, nor are all algorithms good for a particular image. The fundamental limit in solving the segmentation problem is the fact that segmentation is a problem of psycho-physical perception and therefore not susceptible to a purely analytical solution. Each algorithm’s utility is limited by its specific characteristics that make it applicable for particular kind of images. The fundamental challenge in image segmentation is then to provide a generalized framework that is capable of choosing a suitable algorithm from many candidates given a particular image. In this talk, I will presents a probabilistic framework that allows for the selection of an appropriate image segmentation algorithm based on the characteristic properties of the image to be segmented and the algorithm’s behavioral properties. Within the developed framework, the ability to perform this evaluation is learned using a training set of images. Based on this knowledge, the evaluation or prediction of each candidate algorithm’s capability of segmenting the input image is done without actually running any of the algorithms. Segmentation is performed using only the algorithm predicted to achieve the best outcome. I will report on the utility of the developed framework and present results of segmentation in natural scenes.

Speaker Biography

Shishir Shah is Assistant Professor of Computer Science at the University of Houston. He received his M.S. and Ph.D. degrees in Electrical and Computer Engineering from The University of Texas at Austin. He received his B.S. degree in Mechanical Engineering from The University of Texas at Austin. His current research includes computer vision and pattern recognition with applications in biomedical image analysis and distributed multi-modality sensing. He has co-edited one book, and authored papers on object recognition, sensor fusion, statistical pattern analysis, and bioinformatics.