Research

Iris biometrics uses the texture pattern of the iris to create a unique identifier for a person. The iris image to the near right shows moderate pupil dilation and only minor occlusion of the iris region by the lower eyelid. The image to the far right shows a cosmetic contact lens, which obscures nearly all of the natural iris texture. We have published results dealing with "fragile bits" in the iris code; the effects of varying pupil dilation, contact lenses, and template aging; the similarity of iris texture in genetically identical irises; and prediction of "soft biometric" attributes from iris texture.

To the left is an aerial image of a Florida community before Hurricane Dennis in 2005. To the right is an image taken after the storm. Detecting changes based on before- and after-storm images can enable automated and consistent damage assessment. My work in this area is a collaboration with Professor Ahsan Kareem in Civil Engineering. This problem is difficult because the images are not necessarily registered, structures that appear in one image may not appear in the other image, and images taken at different times or under different lighting conditions may have different color properties. A narrated ppt by PhD student Jim Thomas gives an overview of correcting for color change that is due to images being taken under different conditions.

We collected a large biometric dataset from volunteer "identical twins" subjects at the 2009 and 2010 Twins Days Festival in Twinsburg, Ohio. Current commercial face recognition technology has difficulty telling twins apart. However, it appears that human observers can do better by focusing on fine details of the face image. Also, while iris biometrics does not 'see' the similarity in the iris textures of twins, our experiments show that human observers can tell if a pair of iris images come from identical twins or from unrelated persons. We have also explored the use of 3D shape of the face as a means to distinguish twins from each other. We hope that the study of human ability to distinguish twins can lead to improved biometric algorithms.

Current Research - Biometrics.
My research group has done extensive work in iris biometrics, 2D, 3D and infra-red face recognition, face recognition in video, 2D and 3D ear biometrics, multi-modal biometrics and related topics. We have been involved in support of the Human ID Gait Challenge, the Face Recognition Grand Challenge, the Face Recognition Vendor Test 2006, the Iris Challenge Evaluation, and the Multiple Biometric Grand Challenge.

Our lab's early intellectual property in iris biometrics is no longer available for license, but more recent IP in the iris biometrics area is still available. For example, see our method of masking inconsistent bits in the iris code and our method of reducing the effects of dilation on iris matching.

Current Research - Data Mining.
My research in data mining is focused on creating ensembles of classifiers for problems that exhibit "extreme" characteristics such as a high imbalance between classes in the training data, unusually large size of training data, and noise in the class labels of the training data.

My most highly-cited paper in this area introduced the synthetic minority over-sampling technique, or "SMOTE". SMOTE is an approach to dealing with class imbalance in the training data by creating additional synthetic training samples.

Medical Image Analysis.
I previously worked in several areas of medical image analysis, including detection of signs of cancer in mammograms, algorithms for estimating shunted blood flow in the heart, and CT stereotaxis calculations used by neurosurgeons operating on the brain.

Object Recognition.
I made a series of contributions in two areas of object recognition: object recognition through reasoning about functionality and aspect graphs for object representation.

Publications.
A list of recent and selected publications is available, with links to pdf versions of many. My most-cited papers are largely the same in Google Scholar and ISI Web of Science, except for a CVPR 2005 conference paper that has 1000+ citations on Google Scholar that is not tracked by Web of Science. Highly-cited papers on both sources include the data mining / classifier ensemble papers on "SMOTE" in JAIR and "local accuracy estimates" in Trans PAMI, an evaluation of range image segmentation algorithms in Trans PAMI, papers on 3D + 2D face recognition in Trans PAMI and CVIU, papers on evaluation of edge detection algorithms in Trans PAMI and CVIU, a paper on the Human ID Gait Challenge in Trans PAMI and a book chapter on the Digital Database for Screening Mammography. My "h index" for citations currently runs over 50 based on my Google Scholar profile.

Grants.
My research in the area of biometrics has been supported by several federal agencies, including IARPA, FBI, CIA, TISWG, DARPA and NSF. My research in the area of data mining and classifier ensembles has been supported primarily by Sandia National Labs. A list of grants active since 2000 totals over $11M.

PhD graduates.

A list of the wonderful colleagues that I have had the pleasure of working with as PhD students is available, with links to some of the dissertations.

Deborah Thomas and Karen Hollingsworth are two recent PhD graduates. The picture of us to the right was taken at the May 2011 graduate school ceremony. Deborah is currently teaching at Bethel University in Minnesota, and Karen is current a Research Assistant Professor at Notre Dame, supported by an IC Community Postdoc grant.

Of my nineteen PhD graduates to date, seven are in academia and twelve are in industry, and, totally by coincidence, seven are women and twelve are men.

Graduate students currently working with me include Jim Thomas, Estefan Ortiz, Matthew Pruitt, Jeremiah Barr, Amanda Sgroi, Jay Doyle, Vipin Vijayan and Cory Hayes.