GPU random walkers for iterative image segmentation.

摘要

Image segmentation is the act of partitioning an image into distinct regions based on properties that the pixels in those regions share, such as luminance, texture, or color. Image segmentation finds application in fields ranging from medical imaging to computer vision, all of which require the ability to distinguish contiguous regions. Based on user-specified foreground and background "seed" pixels, the random walkers segmentation algorithm calculates probabilities for pixel-placed random walkers "walking" across additional pixels that they are connected to and arriving at one of the seeds. The probability is calculated based on the variance of the shared property.;This thesis presents an algorithm to expand the usefulness of random walkers that provides the ability for interactive image segmentation and refinement. This approach minimizes delays in visual feedback during segmentation through the use of iterative processes. Starting with lower convergence thresholds leads to lower initial probabilities with less definitive segmentations. As more seed points are added and more is known about the desired segmentation, the system maintains interactivity for further refinements through dynamic convergence thresholding. To aid in this computationally intensive process, highly parallel graphics processing units are employed. The implementation is developed as an Adobe R Photoshop R plug-in to enable comparison with other currently available image segmentation techniques.