Carnegie Mellon University
March 21, 2016

Bhagavatula and former students receive Best Paper Award

ECE Professor Vijayakumar Bhagavatula and his co-authors, former Ph.D. students Jonathon Smereka and Andres Rodriguez, has received the Best Paper award at the IEEE International Conference on Identity, Security, and Behavior Analysis held in Sendai, Japan. Their paper, “Selecting Discriminative Regions for Periocular Verification,” introduces a novel unsupervised approach to automatically select regions in the periocular image for improved match performance.

“The work in this paper aims to leverage how discrimination ability changes across a biometric so regions that are highly separable between non-similar users will have more influence on the final match score,” said Jonathon Smereka. “Previously, this was done by employing some method of traditional image segmentation, which would find components separated by prominent edges, that represent 'things' or 'parts of things'. We found that by redefining what constitutes as a 'good' region between a match pair (by using a metric of similarity, rather than dissimilarity between neighboring regions), we can achieve state-of-the-art recognition rates with simple feature based matching.”

A fundamental step in biometric recognition is to identify discriminative features in order to maximize user separation. Matching systems will often require manually choosing these discriminative regions of interest for feature extraction and/or score fusion. Specifically within periocular recognition scenarios, previous works segment the eyebrow and/or eye. While such efforts demonstrate the discriminative power of these regions, in this paper we show that there are various scenarios where blindly employing this type of segmentation is not consistently effective. Thus, we introduce a novel unsupervised approach to automatically select regions in the periocular image for improved match performance. A periocular image is segmented into rectangular regions (this process is referred to as patch segmentation) which improve the overall discrimination ability of the biometric samples being matched. We demonstrate the efficacy of this approach via extensive numerical results on multiple periocular biometric databases exhibiting challenges commonly found in uncontrolled acquisition environments. As the proposed approach is shown to be equivalent to or better than state-of-the-art on each dataset, our results indicate that our patch segmentation is an important step which can greatly influence system performance.