Face Image Understanding

Introduction
In recent years there has been significant progress in the automatic interpretation of images of human faces. The applications of such techniques include: Access Control, Behaviour Monitoring, Expression Recognition, Image Enhancement, Synthesis, Database Indexing, and many others. Existing techniques tend to address each of these areas individually, producing highly specific solutions to well-constrained problems. The aim of the Faces Project in the Wolfson Unit is to develop a generic approach to all these applications, providing a single, unified system capable of performing any face interpretation task in a wide range of circumstances, from high-resolution stills to low-resolution and noisy video images.

Approach
The difficulty in understanding faces comes from the large degree of variability possible in images of any face. Our system addresses the interpretation problem by `learning' about this variability from large sets of training data. In particular, the system must learn about variation in:

1.: Identity
2.: Expression
3.: Viewpoint
4.: Lighting

Statistical techniques have allowed us to build a photo-realistic model of faces, incorporating all this variation. (See the middle image at the top of this page.) This model can not only deal with all types of variation, but separate the variation into particular types, revealing the Identity, Expression, Viewpoint and Lighting for any face image.

Current Research
We have recently develop a novel method of locating deformable objects (such as faces) in images. These are known as Active Appearance Models and are the subject of ongoing research. We have applied this approach to face images and shown that, using the model parameters for classification we can obtain good results for person identification and expression recognition using a very difficult training and test set of still images.

We have also demonstrated how this method can be used in the interpretation of video sequences. The aim is to improve recognition performance by integrating evidence over many frames. A face appearance model can be partitioned to give sets of parameters that independently vary identity, expression, pose and lighting. We exploit this idea to obtain an estimate of identity which is independent of other sources of variability and can be straightforwardly filtered to produce an optimal estimate of identity. This leads to a stable estimate of ID, even in the presence of considerable noise. This approach can be used to produce high-resolution visualisation of poor quality sequences.

Research Directions
The `Faces Group' aim to bring together several of our recent advances in a real-time demonstrator. This state-of-the-art system will interpret video sequences of faces and even recognise faces where the picture quality is so low it is impossible for a human observer to do so. There remain many interesting areas of investigation for new researchers, including:

Models of Dynamic Behaviour
Biometric Data Fusing (audio and video)
Automatic Model Building
Fast/Reliable Initial Detection
Non-Linear/Multi-Part Models
Synthetic Face Generation

Acknowledgements
Funding for this project is provided by the EPSRC and British Telecom PLC.

Enquiries
The Faces Group welcomes inquiries from any interested students. To speak to people working on Faces projects, contact:

Tim Cootes: tim.cootes@man.ac.uk
Louise Butcher: Louise.Butcher@man.ac.uk

Research Courses Clinical Radiology Industrial liaison Contact Us Join Us Search Home

Top Page