MAS622J: FACE RECOGNITION PROJECT
Niloy Mukherjee, Nikolaos Mavridis, MIT Media
Lab, Fall’ 02 |
|
Introduction
The
traditional problem
Our
problem setting
Overview
Ôhe mean face image of the training set:
as expected, highly symmetric, and often in symmetry lies beauty!
|
Introduction
Automated
face recognition is a challenging pattern recognition problem, with an
already vast existing literature covering various aspects of it, and a number
of commercially available systems. Its applications are numerous, and range
from biometric identification to generalised surveillance to advanced
human-computer interaction and beyond. A
somewhat outdated but nevertheless concise review can be found in [1]. Unfortunately,
as justified by the latest escalation of world terrorism, face recognition
together with other biometric identification techniques will certainly become
much more widespread in the future. There are even statistics that public
opinion (a figure of 86% is quoted!) has become more tolerant and sometimes
even demands this [2]. Thus, entry/exit logging in many public places may not
be very far away, and successful deployment of smaller-scaler non-govermental
applications such as cardholder verification in ATM’s will offer huge
immediate savings to banks. The
cognitive science / neuroscience viewpoint of face recognition is also very
interesting. For example, there is strong evidence for the conjecture that
the human face recognition system is distinct from general object
recognition. The rare disorder of “prosopagnosia” [3] (prosopo (ðñüóùðï) = face, agnosia (á-ãíùóßá) = not knowing) provides the main clues for
the above conjecture, as in that case humans with normal general object
recognition performance have great difficulty recognising faces. The traditional problem
The
traditional setting is as an identity recognition or authentication problem:
in the first case, the system is given samples of the faces of a set of
people, and is asked to identify an unlabelled picture with one of the people
it has been trained for (sometimes also with a “reject” option). In the
second case, the system is again trained on a set of people, and is given a
novel picture together with the supposed identity of the person, and should
decide whether the novel picture really matches the identity. Typical
applications of the first include matching the photo of an unknown suspect
with somebody in a criminal database, and of the second cardholder
verification in ATM’s. A main
justification of the difficulty of the problem lies in the wide variation of
appearances of faces. Humans can successfully recognise faces (also see
[4])under short-term variations caused by the environment or the viewpoint
such as pose, added artifacts such as glasses, other occlusions, different
illumination conditions, expressions, but also longer-term variations such as
hairstyle, beard-growing, and the effect of ageing. Of course, humans often
exploit multimodal (whole-body, gait movement, speaker id) and larger
contextual information towards their final decision; but even after
constraining the information to a simple 2D picture, humans do an impressive
job. However, modern face recognition systems have been shown to outperform
humans in some cases: for example, Baback reports superiority in gender
identification with images not containing hair – see [5]. Our problem setting
There are
many other useful related problems, apart from the traditional face
ID/verification. For example, other forms of categorisation: based on gender,
age, colour, expression, and recognition of face artifacts such as a
moustache or a hat. This is exactly what we will deal with in this project:
the categorisation of faces in a discrete and finite nine-dimensional space,
along suitably quantized dimensions of gender, age, colour, expression,
moustache, beard, glasses, bandana and hat. Apart
from providing a description of the face (useful in metadata /
description-based searching etc.), and expression estimation (useful for
emotional state estimation, and affective HCI), this information might also
be advantageous towards boosting performance in the traditional problem, for
example by excluding the area of moustache after its identification and
localisation as “missing info” and then proceeding with the search, in case
no matching pictures with moustaches where found. One can think of many other
uses; their value of course remains to be proved in practice. Overview
A
discussion of the problem and a description of the data set serves as a
natural first step. Preprocessing and some possible representations of face
images follow. An extensive discussion of feature selection, aiming not only
towards recognition directly but also towards localisation of the relevant
information in a subarea of the facial image, is then presented, together
with some first results. Further classification methods and results obtained
so far, as well as a future directions section conclude. References: [1]
Chellappa, R.; Wilson, C.L.; Sirohey, S. , “Human and machine recognition of
faces: a survey” , Proceedings of the IEEE , Volume: 83 Issue: 5 , May 1995
,Page(s): 705 -741 [2]
Sullivan, B , “Warming to big brother”, MSNBC Tech-Science News, http://www.msnbc.com/news/654959.asp?0si=-&cp1=1 [3] Burman,
C ,“Prosopagnosia pages”, [4] Bruce,
V.; Hancock, P.J.B.; Burton, A.M. “Comparisons between human and
computer recognition of faces”, Automatic
Face and Gesture Recognition, 1998. Proceedings. Third IEEE International
Conference on , 1998 Page(s): 408 –413 [5] Moghaddam, B.; Yang, M-H., "Gender
Classification with Support Vector Machines", IEEE International
Conference on Automatic Face and Gesture Recognition (FG), March 2000 Next page: The problem & datasets |