Event Location: NSH 1507
Bio: Shengyang Dai is a Ph.D. candidate in the Electrical Engineering and
Computer Science department of Northwestern University, under the
supervision of Professor Ying Wu. He was a research intern with NEC
Laboratories America, Cupertino, CA in 2006, a research intern with
the Interactive Visual Media Group, Microsoft Research, Redmond, WA in
2007, and a software engineer intern with the Vision Research Group,
Google Research, Mountain View, CA in 2008. He obtained his B.S. and
M.S. degrees from the Electronic Engineering department of Tsinghua
University, Beijing, China in 2001 and 2004 respectively. He received
the Everly Fellowship at Northwestern University in 2008. He is a
Student member of the IEEE. He holds 1 US patent and has 2 more
patents pending. His research interests include image/video
enhancement, vision based detection, tracking, image search, and
machine learning.

Abstract: Alpha matting has been extensively studied in computer graphics. We
explore the special properties of the image alpha channel, and
introduce the alpha channel image modeling method to tackle some very
challenging image processing tasks. We first show how to apply this
technique on estimating space-variant motion blur. Motion blur
estimation has been intensively studied for decades. While for the
space-variant motion blur, previous solutions can only address some
special cases with the help of additional information. A major
contribution of our work is a new finding of an elegant motion blur
constraint. Exhibiting a very similar mathematical form as the optical
flow constraint, this linear constraint applies locally to pixels in
the image. Therefore, a number of challenging motion blur estimation
problems can be addressed in a unified way. The alpha channel modeling
technique is also successfully applied on color image super
resolution. A Softcuts method is proposed to analytically measure the
smoothness of image edges with gradual transitions. By combining the
Softcuts measure and the alpha channel image modeling, different image
edges can be processed simultaneously in a unified way to achieve
significant enhancement of the image quality.