Information Technology Reference
In-Depth Information
Ta b l e 1 .
Comparison with different approaches
Method
PN
SN
SN
Decision Rule
Measure
Recognition Rate(
%
)
[14]
97
-
6
-
Five-fold
90.9
[29]
96
320 7(6)
-
Ten-fold
88.4
[30]
90
284
6
-
-
93.66
[31]
90
313
7
-
Leave-One-Subject-Out
93.8
[32]
97
374
6
-
Ten-fold
91.44
CSF
97
374
6
-
Ten-fold
94.92
CSFMC
97
374
6
Median rule
Ten-fold
95.19
BCSFMC
97
374
6
Median rule
Ten-fold
96.32
measures. It should be noted that the results are not directly comparable due to differ-
ent experimental setups, processing methods, the number of sequences used etc., but
they still give an indication of the discriminative power of each approach. From this ta-
ble, we can see that CSF obtained better result than block-based LBP-TOP that divided
face image into 8
. Additionally,
CSFMC and BCSFMC are slightly better compared to CSF. BCSFMC outperformed
all the other methods.
×
8 overlapping blocks [32], with an increase of
3
.
48%
5Con lu ion
In order to boost facial expression recognition, we propose a component-based spa-
tiotemporal feature (CSF) to describe facial expressions from video sequences. In our
approach, facial interest points in an initial frame are detected by ASM that is robust
to errors in fiducial point localization. According to those interest points, facial com-
ponents are computed on areas centered at those points in a sequence, providing less
redundant information than block-based methods. Comparing with appearance and ge-
ometric approaches, our component-based spatiotemporal approach belongs to hybrid
methods with advantages from both. However, our method describes the dynamic fea-
tures from video sequences. Furthermore, for boosting CSF and reducing the compu-
tational cost of each classifier, AdaBoost is utilized to select the most discriminative
spatiotemporal slices from the facial components. Finally, we also present an approach
for fusing several individual classifiers based on mean, median or product rule.
In experiments on the Cohn-Kanade database we have demonstrated that the CSF de-
scriptors with multi-classifier fusion and AdaBoost feature selection lead to a promising
improvement in facial expression classification. In future work we plan to explore how
our approach could be adopted to very challenging problems including more severe
head pose variations and occlusion. Spontaneous facial expressions common in many
practical applications of facial expression recognition will also be studied.
Acknowledgements
The financial support provided by the Academy of Finland is gratefully acknowledged.
The first author is funded by China Scholarship Council of Chinese government. This
Search WWH ::
Custom Search