Graphics Reference
In-Depth Information
Although the recognition accuracy of this method is considerably high for neutral and near
neutral face recognition, its ability to handle recognition under facial expressions is limited.
2.4.2 Regional 3D Facial Features
The rigid matching of sub-facial surfaces (regions of the 3D face) or the extraction of (rigid)
features from specific regions of the face has proven to mitigate the effects of facial expressions
on the recognition accuracy. Chances are facial expressions leave some regions undeformed,
depending on the expression at hand. Even, deformed regions could appear locally minimal.
ICP-based approaches:
Indeed, some of the most successful approaches to 3D face recogni-
tion are those based on region-matching using ICP. The reason behind their success is that ICP
is very accurate when matching rigid surfaces but is, on the other hand, sensitive to deforma-
tions. As discussed earlier, region-based matching mitigates the affects of facial expressions
and consequently the accuracy of ICP is retained, to an extent. Chang et al. (2006) performed
3D face recognition using ICP on multiple overlapping nose regions. For matching a probe to
a gallery facial scans, the landmarks (e.g., the nose tip, eye pits, and the ridge of the nose) are
first detected using curvatures. On the basis of the locations of the landmarks, the overlapping
regions in the probe scan are extracted. ICP is then used to individually register each extracted
region to the gallery scan. From the registration errors of the overlapping regions, a total
dissimilarity measure between the probe and the gallery scans is computed on the basis of
the product, sum, and minimum rules. The product and the sum rules have demonstrated a
better recognition accuracy than the minimum rule, but they (the product and sum rules) were
comparable to each other. By matching facial scans using ICP on the forehead and the nose
as two separate regions, a higher recognition accuracy has been achieved (Mian et al., 2005,
2007). In order to reduce the computational complexity of ICP, a rejection classifier was used.
The rejection classifier is a histogram of the number of the 3D points of the face falling in
between concentric spheres centred at the tip of the nose with increasing diameters called
spherical face representation (SFR).
Miscellaneous region-based approaches:
The 3D face recognition approach proposed by
Gunlu and Bilge (2010) divides a pose-corrected range image of the 3D face into a number
of squares. The discrete cosine transform (DCT) is then applied to each region, producing
a collection of DCT coefficients. The most discriminative and expression invariant DCT
coefficients are selected as features for recognition, by means of a feature selection approach.
Intuitively, as DCT is a template matching approach, one would expect that this approach to
give a higher recognition accuracy than a holistic PCA due to the use of feature selection.
However, the selected features may still be affected by pose variations as facial expressions
also introduce pose correction errors.
The curvature-based approach proposed by Moreno et al. (2003), one of the early ones in
this category, segments the 3D face into regions on the basis of the sign of the mean
H
and
Gaussian
K
curvatures. A number of attributes of the segmented regions are used as features
for recognition. These attributes include their areas, distances, and angles between their centers
gravity, average
H
and
K
. It is noted that unlike most other approaches in this category, the
division into regions is not meant to achieve invariance to facial expressions but is rather a
