Utilizing Changepoint Detection to Improve Boundary Tracking in Noisy Images

• Activity Number: 409
• Type: Contributed
• Date/Time: Tuesday, July 31, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Section on Government Statistics
• Author(s): Alexander Chen*+
• Companies: SAMSI
• Keywords: Change-point detection ; CUSUM ; Boundary tracking ; Image processing ; High dimensional data

Abstract

This talk introduces improved algorithms for tracking boundaries in an image using change-point detection techniques. Boundary trackers move at the interface between two regions, weaving between them and making course corrections based on a decision function. By considering only local information, boundary tracking algorithms are able to follow boundaries very efficiently, making them quite suitable for work in large or high-dimensional images. As local algorithms, however, boundary tracking methods are extremely susceptible to noisy data or texture. We adapt the CUSUM algorithm to detect region changes on a local level, vastly improving tracking ability. Further improvements can be made by introducing a second change-point detection statistic to detect global off-boundary movement. We discuss the uncertainty involved in the tracking of noisy images and the extent to which false alarms and detection delays in change-point algorithms affect tracking. The adaptation of boundary tracking to the problem of image segmentation via hybrid schemes and applications to hyperspectral data and the tracking of fractal-like structures will also be discussed.