Tuesday, April 28, 2009
Colleen O'Hagan
Chair: Worthy Martin; Kamin Whitehouse
Advisor: Jason Lawrence
OLSSON 228E, 15:30:00
A Master's Thesis Presentation
Visualizing Multiparameter Images
ABSTRACT
The work describes a technique for fusing an arbitrary number of aligned image bands into a single displayable visualization. This problem is relevant in many fields including medicine, astronomy, and particularly remote sensing. In each case images are captured with a range of modalities including visible spectrum representations from traditional photography, IR, UV, and thermal response images. With a stack of images domain scientists can infer information about a region they may not be able to interpret from an image taken in the visible spectrum alone. Studying each image in the stack and synthesizing the responses by sight is an unrealistic means of processing the information, especially in cases where there are hundreds of images. Fusing the stack of images into a single color image allows for an easier classification and can uncover patterns that may go unnoticed by merely examining the input images. The dimensionality reduction of information is performed in the context of a constrained multidimensional scaling problem designed to preserve consistency in distances between pixel values in the image stack and those in the visualization. The embedding is achieved by majorizing an objective function known as stress. Constraints on the stress formulation and adaptive scaling of input distances allow for user control and detail exaggeration, respectively. The presence of such constraints allows domain scientists to tune output images to correspond with familiar visualizations. Optimizations on the algorithm offer an efficient solution and it is shown that this technique has superior performance and accuracy as compared to existing methods. To illustrate this, results of the algorithm are presented for various applications within the scientific visualization domain.