Home > Events > Oral Candidacy Exam - Jianxu Chen

Oral Candidacy Exam - Jianxu Chen

Start: 12/7/2015 at 3:00PM
End: 12/7/2015 at 5:30PM
Location: 258 Fitztpatrick Hall
Attendees: Faculty and Students are welcome to attend the presentation portion of the defense.
Add to calendar:
iCal vCal

Jianxu Chen

Oral Candiacy Exam

December 7, 2015        258 Fitzpatrick Hall        3:00 pm

Adviser:  Dr. Danny Chen

Committee Members:

           Dr. Mark Alber             Dr. Patrick Flynn        Dr. Walter Scheirer        




Matching two sets of geometric objects is a fundamental problem in various theoretical and applied fields. Different from matching points, polygons, or polylines, when the objects are of arbitrary shapes and topological changes are allowed, the problem is very challenging. I proposed a matching model based on the Earth Mover's Distance (EMD) to find the optimal matching between two sets of objects of arbitrary shapes allowing shape division and fusion. Built upon the EMD matching model, I developed a new hybrid framework for cell segmentation and tracking and applied it to track the motion of Myxococcus xanthus in time-lapse images. The EMD matching model was adapted to a hierarchical matching scheme, a key module in the hybrid framework to establish preliminary cell correspondence across frames. The new framework outperformed the state-of-the-art tracking algorithms by a large margin.

Besides, I studied another two related problems in my preliminary works. First, I proposed anew automated approach for iris recognition based on human interpretable features. The EMD matching model was employed to match features between iris images and compute the similarity. The proposed approach achieved a huge improvement over the previous work on visible feature based iris recognition. Second, I designed a segmentation algorithm for fibrin network in 3D confocal microscopy images, which achieved much higher accuracy in detecting fiber segments and branching points than the best known algorithm. The study in fibrin network segmentation will be the foundation of one potential generalization of the EMD matching model in the future works.

I plan to continue my research in four specific directions. The first direction is to leverage the power of deep learning to extend the EMD matching model to address general cell tracking problems, i.e., different types of cells. The generalizations are meant to handle shaped eformation, special events (e.g., mitosis, apoptosis, or fusion) and overlapping cells. Second, I am interested in generalizing EMD matching to EMD clustering. Considering the advantage of EMD in measuring difference in “patterns”, EMD clustering will be of high importance in both the theory and applications. For instance, EMD can measure the difference between trajectories,and EMD clustering could be an effective means to discovery motion patterns of groups, like huge cell swarms. The third research direction is to elevate the EMD matching model to track high-level representations instead of individual subjects. One example of important applications is the recognition of the behavioral patterns of cell swarms when individual cells are not all trackable. Finally, arising from a real problem in colonoscopy imaging, it is my interest to generalize the EMD matching model from matching shapes to matching structural features, such as networks.