Research Highlights

Muscular Movement Model

This paper presents a novel and effective approach to automatic 3D facial expression recognition based on the Muscular Movement Model (MMM). It first automatically segments the input face by localizing the corresponding points around each muscular region of the reference face using Iterative Closest Normal Pattern (ICNP). A set of features of multiple differential quantities, including coordinates, normals and shape index values, are then extracted to describe the geometry deformation of each segmented region. We analyze the importance of these muscular areas, and a score level fusion strategy which optimizes the weights of the muscular areas by using a genetic algorithm is proposed in the learning step.

Muscular Movement Model based Automatic 3D Facial Expression Recognition, Qingkai Zhen, Di Huang*, Yunhong Wang, Liming Chen, MultiMedia Modeling, 8935 , 522-533,2015.

[PDF] [Code(coming soon)] [Project page]

Detection Results on Video Sequences

Based on skin color segmentation combined with the saliencymodel, a novel method is proposed to detect human faces invideos. Firstly, a skin color model in the YCbCr chrominancespace is built to segment skin-color pixels from background inframes. Then, the candidate regions of human face can beextracted using mathematical morphological operators. Finally, asaliency model is proposed to detect eyes and mouths in faceregions.

Trackingand Recognition Face in Videos with Incremental LocalSparse Representation Model, Chao Wang, Yunhong Wang*,Zhaoxiang Zhang, Optical Engineering, 52(10), 1-13, 2013.

[PDF] [Code(coming soon)] [Project page]

Face Aging Effect Simulation

A novel approach using hidden factor analysis joint sparserepresentation is presented to address the issue of face agingsynthesis. In contrast to the majority of the published tasksthat handle the facial texture integrally, the proposed agingapproach separately models the person-specific facialproperties that tend to be stable in a relatively long periodand the age-specific clues that change gradually over time;and then merely transforms the age component to a target agegroup via sparse reconstruction, thus yielding aging effectson faces. Furthermore, the proposed aging synthesis modelpresents a good universality by accounting for the uncertainexternal facial variations. A series of evaluations prove itsvalidity with respect to the ability of identity preservationand age accuracy of synthesis.

Face Aging Effect Simulation using Hidden Factor AnalysisJoint Sparse Representation, Hongyu Yang, Di Huang,Yunhong Wang, Heng Wang, Yuanyan Tang, 2015.

[PDF] [Code(coming soon)] [Project page] [Data]

BUAA-VisNir Face Database

Face recognition is one of the most important directions inour lab. Our work includes 2D face recognition with visiblelight/near infrared images, 3D face dataset collection &recognition, synthesis of aging face, etc.


Gait Recognition

Gait is one of the most important features in Biometrics. It is user friendly and widely utilized in video surveillance. We build a large multi-view gait dataset for our research. Based on that, we propose several hierarchical model of gait features, which are based on HMM model or CRF.

DERF: Distinctive Efficient RobustFeatures From the Biological Modeling of the P Ganglion Cells

A new local image descriptor, termed distinctive efficientrobust features (DERF), is derived by modeling the responseand distribution properties of the parvocellular-projectingganglion cells in the primate retina. DERF featuresexponential scale distribution, exponential grid structure,and circularly symmetric function difference of Gaussian (DoG)used as a convolution kernel, all of which are consistent withthe characteristics of the ganglion cell array found inneurophysiology, anatomy, and biophysics. In addition, a newexplanation for local descriptor design is presented from theperspective of wavelet tight frames.

DERF: Distinctive Efficient Robust Features From the Biological Modeling of the P Ganglion Cells,Weng Dawei, Wang Yunhong, Gong Mingming, Tao Dacheng, Wei Hui,Huang Di, Image Processing, IEEE Transactions on, 2015, 24(8):2287-2302.

[PDF] [Code(coming soon)] [Project page]

Person Re-identification

A novel Relevance Metric Learning method with ListwiseConstraints (RMLLC) by adopting listwise similarities, whichconsist of the similarity list of each image with respect toall remaining images, is proposed. By virtue of listwisesimilarities, RMLLC could capture all pairwise similarities,and consequently learn a more discriminative metric byenforcing the metric to conserve predefined similarity listsin a low dimensional projection subspace. Despite theperformance enhancement, RMLLC using predefined similaritylists fails to capture the relative relevance information,which is often unavailable in practice. To address thisproblem, a rectification term is further introduced toautomatically exploit the relative similarities. And anefficient alternating iterative algorithm to jointly learn theoptimal metric and the rectification term is developed.

Relevance Metric learning for Person Re-identification by Exploiting Global Similarities,Jiaxin Chen, Zhaoxiang Zhang, Yunhong Wang, in ICPR,pp.1657-1662, 2014.

[PDF] [Code(coming soon)] [Project page]

Object Detection and Tracking using Static Information

We operate object detection and tracking using staticinformation, e.g. gradient and texture descriptors, in eachframe of one video, and recognize their category with dynamicinformation such as speed and directions. Based on adeformable vehicle model, we can classify them in more detailcategories, which includes cars, sedans or pickups.

Our research interest also includes signature and speech features. They can be used for recognition and validation.

Digital watermarking can be employed for tamper detection and copyright protection. Several methods have been propesed by us for automatic tamper region detection. And biometric watermarking is one excellent application for information hiding.

Object-based Feature Extraction and Semi-supervised Classification for Urban Change

This paper presents a novel approach for urban changedetection of high resolution (HR) remote sensing images. Toovercome deficiency of traditional pixel-based methods andbetter annotate HR images, object-based strategies areadopted. Firstly change vector analysis (CVA) and local binarypatterns (LBP) are utilized to extract the object-specificfeatures based on the image-objects acquired by multitemporalsegmentation. Then sparse representation is further exploitedto characterize highly effective sparse features. Finally, thefinal change map is obtained by support vector machine (SVM)with the pseudotraining set acquired by expectationmaximization (EM).

Object-based Feature Extraction and Semi-supervisedClassification for Urban Change Detection UsingHigh-resolution Remote Sensing Images, Bin Hou, QingjieLiu, Yunhong Wang, IEEE International Geoscience and RemoteSensing Symposium, 2015.

[PDF] [Code(coming soon)] [Project page]

Object Detection based on Salient Region Detection

For the part of remote sensing image processing, our workfocus on registration and fusion with images from varioussensors. Furthermore, we propose a novel approach for objectdetection based on salient region detection.