Image segmentation is a ubiquitous problem in computer vision and image processing. In some applications, such as biomedical imaging, the problem may become very complex, especially for noisy video sequences captured with low resolution and insufficient contrast. However, the highly structured nature of such data often provides additional information. When image deformations have just a few underlying causes, such as continuously captured cardiac MRI, the captured images lie on a low-dimensional, non-linear manifold. The manifold structure of such image sets can be extracted by automated methods for manifold learning, and used as new constraints for tracking and segmentation of relevant image regions. In this, we explore mechanisms to integrate automated manifold learning tools as a pre-processing step to provide new multi-image constraints to be used in segmentation and tracking. We demonstrate that substantial improvements can be achieved for tradition segmentation and tracking techniques in challenging conditions.
展开▼