Zengyin Zhang

I'm now a Visiting Scholar of Robotics Institute in Carnegie Mellon University, working with Fernando de la Torre in Component Analysis Lab. Current project aims at Image Alignment.
My personal page in RI can be found here.
Here is my email address: zhangzyin#gmail.com or zhangzy#cmu.edu
My office is
EDSH 110, Robotics Institute,
Carnegie Mellon University.
5000 Forbes Ave, Pittsburgh, PA 15213, USA.

Education

Sept. 2005 - July 2008 Peking University
Master of Engineering - Electronics Engineering and Computer Science
Sept. 2001 - July 2005 Peking University
Bachelor of Science - School of Mathematical Sciences

Awards

2007 Excellence of Academic Performance Award (Awarded to the top 5%)
2006 Outstanding Student in Academy, Moral and Health (Awarded to the top 5%)
2001 Third prize in ACM Contest of Peking University
2000 Bronze Medal in National Olympiad in Informatics of China

Research Experience

Sept. 2008 - Now Accurate Localization of Eyes and Mouth
Aimed at precisely localizing the corners of eyes and mouth under various environment. Recently, Principal Component Analysis of appearance is used to model the interest areas. Local Binary Pattern feature and Histogram of Orientated Gradient feature are accepted in the system now.
Dec. 2007 - July. 2008 Research on Camera-based Chinese Document Image Processing
The most challenge problem in Camera-based document images is the warped and non- orthogonal paper surface. How to model these surfaces was the core of thesis. With the hints of visualized 3D mesh, major and minor texture flows were defined to abstract the mesh of surface. Finally a carefully designed generative developable surface model was proposed to depict the document surface which also gave reasonable constraints so that the warped surface could be rectified. This is a part of project on Multimedia Information Retrieval sponsored by National High Technology Research and Development Program of China.

Here is some results of capturing my master thesis:
Camera captured image: Processed image:
src image src image result result
Sept. 2007 - Jan. 2008 Research on High-level Feature Extraction from Video
This is a hot spot research on Document Analysis and Recognition now. The most challenging problem is how to rectify the wrapped document surface under arbitrary perspective projection so that it can be easily analyzed by current plane-based technology. Analogous to the visualization of 3D object by meshes, I adopted major and minor texture flow of documents to describe the 3D information of a paper surface, while the major and minor texture flow corresponded to the direction of line text and its orthogonal direction, respectively. Histogram distribution analysis of different orientated projections was applied to estimate the local major texture flow. A structural wavelet whitespace detector (specific to the Chinese characters) was designed to estimate the local minor texture flow. I proposed a global continuity constraint to suppress the disturbances of noise to the texture flows estimation. Then I represent the wrapped paper in reality in a generative developable surface model. It is hard to rectify the developable surface model directly since the number of estimate parameters becomes infinite. Inspired by the differential thought, a group of ruling lines were estimated to segment the surface into a batch of conjoint planes. With the estimated texture flows of document, ruling lines were solved in a mathematical form under the constraint of continuity of slope. Finally, the rectification of paper surface can be simply divided into rectifying a batch of perspective planes.