Abstract
In order to support human users with daily tasks, computers need to
be aware of human presence and identity. Today, it is very easy to
obtain visual information using, e.g. smart phones. Therefore, AI
based algorithms for face detection and recognition are becoming
increasingly relevant. Face recognition is a very easy and convenient
technique compared to other biometrics methods like finger print or
iris recognition. It does, however, still not perform very robust
when compared to humans and is less reliable than alternative
biometrics methods.
This is due to many variations existing in photographs and videos
(e.g. illumination, pose), which create challenging problems for a face
recognition algorithm. We focused mainly on three challenges in this
thesis. The first two are localization and rotational alignment of the
faces, which are preprocessing steps before the recognition. The third
step is face identification itself, based on very little training data.
For localization, we developed an eye detector localizing eye centers
after face detection. Rotational alignment is done using the angles
from these eye centers. For the single and small sample problems, we
proposed two novel methods to handle the insufficient amount of data.
The experiments lead to two important conclusions: First, a big generic
dataset can help to improve recognition performance significantly for
the new target faces. Second, if the number of photographs is limited,
then smart sampling of many patches helps improving identification
accuracy.
Summarizing, although our research contributes to solving face
identification in small sample conditions, further research is
necessary to obtain more robust results.
be aware of human presence and identity. Today, it is very easy to
obtain visual information using, e.g. smart phones. Therefore, AI
based algorithms for face detection and recognition are becoming
increasingly relevant. Face recognition is a very easy and convenient
technique compared to other biometrics methods like finger print or
iris recognition. It does, however, still not perform very robust
when compared to humans and is less reliable than alternative
biometrics methods.
This is due to many variations existing in photographs and videos
(e.g. illumination, pose), which create challenging problems for a face
recognition algorithm. We focused mainly on three challenges in this
thesis. The first two are localization and rotational alignment of the
faces, which are preprocessing steps before the recognition. The third
step is face identification itself, based on very little training data.
For localization, we developed an eye detector localizing eye centers
after face detection. Rotational alignment is done using the angles
from these eye centers. For the single and small sample problems, we
proposed two novel methods to handle the insufficient amount of data.
The experiments lead to two important conclusions: First, a big generic
dataset can help to improve recognition performance significantly for
the new target faces. Second, if the number of photographs is limited,
then smart sampling of many patches helps improving identification
accuracy.
Summarizing, although our research contributes to solving face
identification in small sample conditions, further research is
necessary to obtain more robust results.
Original language | English |
---|---|
Qualification | Doctor of Philosophy |
Awarding Institution |
|
Supervisors/Advisors |
|
Award date | 30-Sep-2016 |
Place of Publication | [Groningen] |
Publisher | |
Print ISBNs | 978-90-367-9158-8 |
Electronic ISBNs | 978-90-367-9159-5 |
Publication status | Published - 2016 |