Pose-invariant Face Recognition through 3D Reconstructions

Thumbnail Image
File version
Primary Supervisor
Gao, Yongsheng
Other Supervisors
Dimitrijev, Sima
File type(s)

Pose invariance is a key ability for face recognition to achieve its advantages of being non-intrusive over other biometric techniques requiring cooperative subjects such as fingerprint recognition and iris recognition. Due to the complex 3D structures and various surface reflectivities of human faces, however, pose variations bring serious challenges to current face recognition systems. The image variations of human faces under 3D transformations are larger than that existing face recognition can tolerate. This research attempts to achieve pose-invariant face recognition through 3D reconstructions, which inversely estimates 3D shape and texture information of human faces from 2D face images. This extracted information is intrinsic features useful for face recognition which is invariable to pose changes. The proposed framework reconstructs personalised 3D face models from images of known people in a database (or gallery views) and generates virtual views in possible poses for face recognition algorithms to match the captured image (or probe view). In particular, three different scenarios of gallery views have been scrutinised: 1) when multiple face images from a fixed viewpoint under different illumination conditions are used as gallery views; 2) when a police mug shot consisting of a frontal view and a side view per person is available as gallery views; and 3) when a single frontal face image per person is used as gallery view. These three scenarios provide the system different amount of information and cover a wide range of situations which a face recognition system will encounter. Three novel 3D reconstruction approaches have then been proposed according to these three scenarios, which are 1) Heterogeneous Specular and Diffuse (HSD) face modelling, 2) Multilevel Quadratic Variation Minimisation (MQVM), and 3) Automatic Facial Texture Synthesis (AFTS), respectively. Experimental results show that these three proposed approaches can effectively improve the performance of face recognition across pose...

Journal Title
Conference Title
Book Title
Thesis Type
Thesis (PhD Doctorate)
Degree Program
Doctor of Philosophy (PhD)
School of Engineering
Publisher link
Patent number
Grant identifier(s)
Rights Statement
The author owns the copyright in this thesis, unless stated otherwise.
Rights Statement
Item Access Status
Access the data
Related item(s)
pose-invariant face recognition
pose invariance
face recognition
face recognition systems
face recognition algorithms
Persistent link to this record