May 29, 2014
UNIVERSITY OF UTAH
ELECTRICAL AND COMPUTER ENGINEERING DEPARTMENT
DISSERTATION DEFENSE FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
Advisor: Tolga Tasdizen
Scene Labeling with Supervised Contextual Models
Scene labeling is the problem of assigning an object label to each pixel of a given image. It is the primary step towards image understanding and unifies object recognition and image segmentation in a single framework. A perfect scene labeling framework detects and densely labels every region and every object that exists in an image. This task is of substantial importance in a wide range of applications in computer vision. Contextual information plays an important role in scene labeling frameworks. A contextual model utilizes the relationships among the objects in a scene to facilitate object detection and image segmentation. Using contextual information in an effective way is one of the main questions that should be answered in any scene labeling framework.
In this dissertation, we develop two scene labeling frameworks that rely heavily on contextual information to improve the performance over state-of-the-art methods. The first model, called the multi-class multi-scale contextual model (MCMS), uses contextual information from multiple objects and at different scales for learning discriminative models in a supervised setting. The MCMS model incorporates cross-object and interobject information into one probabilistic framework and thus is able to capture geometrical relationships and dependencies among multiple objects in addition to local information from each single object present in an image. The second model, called the contextual hierarchical model (CHM), learns contextual information in a hierarchy for scene labeling. At each level of the hierarchy, a classifier is trained based on downsampled input images and outputs of previous levels. The CHM then incorporates the resulting multi-resolution contextual information into a classifier to segment the input image at original resolution. This training strategy allows for optimization of a joint posterior probability at multiple resolutions through the hierarchy. We demonstrate the performance of CHM on different challenging tasks such as outdoor scene labeling and edge detection in natural images and membrane detection in electron microscopy images.
We also introduce two novel classification methods. WNS-AdaBoost speeds up the training of AdaBoost by providing a compact representation of a training set. Disjunctive normal random forest (DNRF) is an ensemble method that is able to learn complex decision boundaries and achieves low generalization error by optimizing a single objective function for each weak classifier in the ensemble.
Finally, a segmentation framework is introduced that exploits both shape information and regional statistics to segment irregularly shaped intracellular structures such as mitochondria in electron microscopy images.
Tuesday, June 3, 2014
Evans Conference Room
3780 WEB (3rd floor)
The public is invited