Applications of Level Set Methods for Image and Video Processing

In this paper, the level set algorithm is examined with a view to applications in image and video processing, and machine vision. Conventional parametric deformable models, which can be viewed as Lagrangian geometric models, where the boundary of the model has the ability to align itself with certain image features, are initially presented. This paper describes some of the problems associated with this approach and introduces the level set method as a step towards a solution. The main idea is to take the Lagrangian form of the parametric deformable model, and replace it with an Eulerian, initial value partial differential equation. This involves recasting the problem in a higher dimension, but it can be shown that this actually leads to a reduction in complexity. Various improvements to the basic form are presented, as well as several different formulations of the guiding speed function. For this paper, the level set method was applied to two areas; finding and tracking a human face in video-conferencing sequences, and extracting subtle contours from medical images. In applications where off-line computation is acceptable, the level set algorithm competes well with established techniques, and can be demonstrated to have several advantages over them. A disadvantage of the level set method over active contours has been found to be the high computational cost.