Optimal Score Fusion via a Shallow Neural Network to Improve the
Performance of Classical Open Source Face Detectors
Moumen T. El-Melegy, Hesham A. M. Haridi, Samia A. Ali and Mostafa A. Abdelrahman
Department of Electrical Engineering, Assiut University, Assiut, Egypt
Keywords: OpenCV, Dlib, Shallow Neural Network, Skin Detector, HOG-based Face Detector, Classical OpenCV Face
Detector, CNN-based Face Detectors.
Abstract: Face detection exemplifies an essential stage in most of the applications that are interested in visual
understanding of human faces. Recently, face detection witnesses a huge improvement in performance as a
result of dependence on convolution neural networks. On the other hand, classical face detectors in many
renowned open source libraries for computer vision like OpenCV and Dlib may suffer in performance, yet
they are still used in many industrial applications. In this paper, we try to boost the performance of these
classical detectors and suggest a fusion method to combine the face detectors in OpenCV and Dlib libraries.
The OpenCV face detector using the frontal and profile models as well as the Dlib HOG-based face detector
are run in parallel on the image of interest, followed by a skin detector that is used to detect skin regions on
the detected faces. To figure out the aggregation method for these detectors in an optimal way, we employ a
shallow neural network. Our approach is implemented and tested on the popular FDDB and WIDER face
datasets, and it shows an improvement in the performance compared to the classical open source face detectors.
1 INTRODUCTION
Face detection is one of the most broadly explored
topics in computer vision and pattern recognition,
which represents the initial and vital stage of many
application pipelines, such as: face verification (Tu et
al., 2017), face tracking (Kim et al., 2008), face
clustering (Cao et al., 2015), and face identification
(Parkhi et al., 2015). From many literature surveys
like (Yang et al., 2002; Zafeiriou et al., 2015), we
observe that face detection has sighted considerable
breakthroughs since the revival of deep learning once
again in 2006 (Wang and Raj, 2017). Since that time,
many well-established face detectors depending on
that technique are provided in literature like CNN-
based face detectors (Li et al., 2016; Hu and
Ramanan, 2017; Tang et al., 2018). However, there
are many industrial applications (Shaikh et al., 2016;
Frejlichowski et al., 2016; Zheng et al., 2016;
Puttemans et al., 2016a; Puttemans et al., 2016b) still
utilize the classical detectors existing in OpenCV
(Bradski, 2000) and Dlib (King, 2018) libraries.
OpenCV library (Bradski, 2000) has a face detector
that relies on the seminal work of (Viola and Jones,
2001) and depends on a cascade of classifiers using
Haar-like features. As another example, Dlib library
(King, 2018) also includes a face detector that counts
on SVM as a classifier using HOG (Histogram
Oriented of Gradient) features (Dalal and Triggs,
2005). These open source face detectors are unable to
give a higher performance on the well-known public
datasets like FDDB (Jain and Learned-Miller, 2010)
and WIDER FACE (Yang et al., 2016) compared to
CNN-based detectors. Several reasons created this
situation (Yang et al., 2002; Zafeiriou et al., 2015),
such as: These classical detectors work effectively in
detecting frontal faces and fail at extreme in-plane
and out-plane rotations. In addition, they lack of
robustness in detecting faces under extreme lighting
conditions. Moreover, these detectors tend to fail in
discovering tiny and occluded faces. Thus, any
endeavors to enhance their performance will have an
effective impact on the applications that count on
them.
There are many methods, such as RSFFD1
(Robust Score Fusion Face Detection) (Rara et al.,
2010), RSFFD2 (El-Barkouky et al., 2012), and
IterativeHardPositives+ (Puttemans et al., 2017), that
try to boost the performance of classical OpenCV
face detector. We follow the same direction trying to
improve the performance of classical face detectors in
660
El-Melegy, M., Haridi, H., Ali, S. and Abdelrahman, M.
Optimal Score Fusion via a Shallow Neural Network to Improve the Performance of Classical Open Source Face Detectors.
DOI: 10.5220/0007691206600667
In Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2019), pages 660-667
ISBN: 978-989-758-354-4
Copyright
c
2019 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
Figure 1: Our fusion method: In the matching block, the
output consists of three scores for each detected face. If
any detected face obtained by one detector has no matches
with the other detected faces of the other detectors, the other
scores of that face will be zeros.
OpenCV and Dlib libraries by:
Running OpenCV Haar-based face detector on
the image of interest using frontal and profile
models.
Running Dlib HOG-based face detector in
parallel with the OpenCV detectors.
Using a skin detector to detect skin regions in
each face rectangle obtained from the above two
points targeting to reduce the number of false
positive faces (the obtained faces that are not
truly faces) by taking into account skin color as
an important feature for human faces.
Employing a shallow neural network to learn
the best method for aggregating the confidence
scores of each face rectangle obtained from
frontal, profile, HOG-based, and skin detectors,
for more details see Figure 1.
The rest of this paper is organized as follows.
Section 2 provides the related work achieved in the
same direction of our target, while section 3 describes
our approach for boosting the performance of
classical open source face detectors. The
experimental results are provided in section 4
followed by a conclusion and potential future work in
section 5.
2 RELATED WORK
OpenCV (Bradski, 2000) and Dlib (King, 2018)
libraries are the most renowned libraries employed in
developing computer vision applications. They are
updated from time to time with new algorithms from
the community of academic researchers and industrial
partners to help industrial users to accurately build
their own working applications. (Viola and Jones,
2001) is one of the algorithms that is included in
OpenCV and used extensively for object detection,
especially face detection. They used AdaBoost
algorithm to learn a cascade of classifiers to
distinguish between faces and non-faces using Haar-
Like features. The authors in (Dalal and Triggs, 2005)
counted on HOG (Histogram of Oriented Gradients)
features and SVM as a classifier. The accuracy of all
these detectors still suffers when applied to public
datasets, such as FDDB (Jain and Learned-Miller,
2010) and WIDER FACE (Yang et al., 2016). These
recent public datasets have many challenges such as
occlusion, illumination, and very tiny faces, however
the open source face detectors are originally designed
to detect frontal faces only. This situation motivates
many researchers to improve the performance of
these detectors.
The designers of Dlib library pursued the same
direction, and they enhanced the face detector
inspired from (Dalal and Triggs, 2005) by creating
five HOG filters for the sliding window that is used
to search about frontal and semi-frontal faces in an
image, and they incorporate the updated detector in
the library, but its accuracy still needs to go up. In (Li
and Zhang, 2013), the authors adopted SURF
(Speeded Up Robust Features) features (Bay et al.,
2008) instead of using Haar-Like features of the
original OpenCV face detector, and they used logistic
regression to learn the best features that differentiate
between faces and non-faces instead of using
AdaBoost algorithm, aiming to raise the accuracy of
the detector. In IterativeHardPositives+ face detector
(Puttemans et al., 2017), the authors improved the
negative training sample collection method, and they
used an active learning scheme to iteratively append
hard positive (positive rectangles categorized as
negatives in the preceding iteration) and hard
negative (negative rectangles labelled as positives in
the former iteration) samples to the training process
of the OpenCV detector. Also, they made a new
annotation file for FDDB dataset, but despite their
efforts, the accuracy of their detector became worse
than the original OpenCV face detector but faster than
it. As we can see, despite the attempts to boost the
performance of open source face detectors, the
Optimal Score Fusion via a Shallow Neural Network to Improve the Performance of Classical Open Source Face Detectors
661
accuracy of the detectors in (Bay et al., 2008; Li and
Zhang, 2013; Puttemans et al., 2017; King, 2018) still
suffers. In addition, some of them added more
complexity and computational cost on the original
detector such as the number of filters in (King, 2018).
Furthermore, the detectors that depends on SURF or
SIFT (Li and Zhang, 2013) features confront another
problem because these feature descriptors are patent
protected; they cannot be used for commercial
purpose except with a permission from the original
inventors.
There are other methods that follow the same
direction, but they use the original simple building
blocks that already exist in OpenCV library such as
skin detection, Viola-Jones face detector, and Viola-
Jones facial part detector, see for example RSFFD1
(Rara et al., 2010) and RSFFD2 (El-Barkouky et al.,
2012). RSFFD1 was one of the best performers in the
competition done by (Parris et al., 2011). RSFFD2 is
a modified version of it, where saliency and skin
information are added, and it consists of four-step
pipeline. The first step is used to generate three scales
for the image under interest. Then, OpenCV face
detector is run on each scale, and each detected
rectangle is assigned a score from 1 to 3 depending
on how many times this rectangle appears in the three
scales. In the second step, the same detector is applied
on each detected rectangle from the first step to detect
facial parts (two eyes and mouth), and each rectangle
is given a score from 0 to 3 relying on how many parts
found. The third step is used to run a skin detection
algorithm on each rectangle given by the first step,
and depending on how many skin pixels found, each
rectangle is allocated a discrete score from 0 to 3. The
fourth or the last step is employed to calculate the
saliency map for each rectangle detected by the first
step and depending on the saliency pixels acquired,
each rectangle is assigned a discrete score from 0 to
3. At the end, each candidate face has 4 different
scores one for each step. After that these scores are
added up to give a value from 1 to 12 to each
rectangle, where the higher score rectangles are more
likely to be true faces counting on the threshold value
that will be used.
3 OUR APPROACH
The techniques used in (Viola and Jones, 2001;
ElBarkouky et al., 2012; Puttemans et al., 2017; King,
2018) suffer from some drawbacks, such as:
They have a problem in detecting tiny, non-
frontal, and occluded faces.
The different scores of the information sources
(RSFFD2 steps) in (El-Barkouky et al., 2012)
are summed directly to give a final discrete
score from 1 to 12, but other aggregation
methods can provide better results.
The running time of saliency map algorithm
used in the fourth step in (El-Barkouky et al.,
2012) is very long making the time of the entire
pipeline exceeds 2 seconds on any device with
limited hardware.
The confidence score of OpenCV face detector
in each scale in (El-Barkouky et al., 2012) is not
taken into account. This information is very
important which can be used to improve the
confidence and the accuracy of the results.
The new annotation file for FDDB dataset
(Puttemans and Goedeme, 2017) used in
(Puttemans et al., 2017) is not accurate, and we
did not get a good reason for that from the
authors when we emailed them.
Due to all the above limitations, each one from
these techniques has an accuracy less than 80% (the
maximum detected true faces to the total ground-truth
faces). The basic idea of our approach is to deal with
these limitations to increase the performance of
detection. Our model consists of four steps, for more
details see Figure 1. In the first step, we run in parallel
on the image the three detectors (OpenCV frontal
Haar-based, OpenCV profile Haar-based, and Dlib
HOG-based face detectors). The output from this step
is the detected faces with their scores. It is well-
known that OpenCV has two types of face detector,
one counts on Haar-Like features and the other relies
on LBP (Local Binary Pattern) features, and the
accuracy of the former outperforms the accuracy of
the latter. This is the reason why we depend on Haar-
based face detectors. The second step is to match the
obtained faces of each detector with the others. The
output from this step is the detected faces with three
scores for each one without relying on the number of
detectors that discover it. That is to say, if any
detector fails to discover a specific detected face of
the other detectors, its score will be zero for that face.
The third step is to use the skin detector in (Brancati
et al., 2016) to detect skin color in each obtained
rectangle from the previous step. The score of this
detector is represented by the ratio of the total skin
pixels in each rectangle to its total area. In the fourth
step, to calculate the final score of our model, Instead
of summing the scores from various sources of
information as in (El-Barkouky et al., 2012), our
approach employs a shallow neural network to learn
the optimal aggregation method for these scores. The
network consists of three layers: input, hidden, and
VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications
662
Figure 2: The proposed shallow neural network used to
represent the function
.
The inputs are related to a candidate detected face and each
one represents a score with a value in the range [0; 1].
is the confidence score of a detected rectangle
obtained by OpenCV face detector using Haar-based frontal
model.
is the score given by OpenCV face detector
using Haar-based profile model.
is the score that
represents the ratio of the total skin pixels in the detected
rectangle to its total area, and
is the score of the same
rectangle detected by Dlib HOG-Based face detector.
output layer, as shown in Figure 2. The input layer
has four neurons: one for the confidence score
coming from the frontal model (
), the second
for the confidence score of the profile model
(
), the third for the score of skin detector
(
, the ratio of the total skin pixels found in a
detected rectangle to its total area), and the fourth for
the confidence score coming from Dlib detector
(
). The output layer consists of one neuron for
the final score of our model,
. The best number
of hidden neurons is determined experimentally using
cross-validation on a subset of WIDER FACE
dataset, and it is found to be 10 neurons. In addition,
the activation function of each neuron in the hidden
and output layer is function; its values range
from 0 to 1. Furthermore, our network is implemented
using neural network toolbox of Matlab 2017a and
trained using gradient descent with momentum. To
prepare the training data for our network, we use
WIDER FACE dataset. We use 12800 images from
this dataset grouped as follows: 8000 for training,
2400 for validation and 2400 for testing. Each image
may contain from a single face to more than 500 face.
For training and validations we need to prepare good
representative examples for faces and non-faces. For
that sake, we rely on the concept of IOU (Intersection
Over Union) (Jain and Learned-Miller, 2010) that is
given by the equation:
(1)
Where
is the detected face rectangle and
is the
ground-truth rectangle. This concept measures the
overlapping between detected rectangles and ground-
truth. Its values range from 0 to 1. In our model, any
candidate face with IOU larger than 0.6 with any
ground-truth face, its four scores will be considered a
good positive example (a true face) with label 1, and
anyone with IOU less than 0.4, its four scores are
treated as a good negative example (a non-face) with
label 0; anyone having IOU in the interval [0.4, 0.6]
is ignored completely in training. So, by using these
positive and negative examples with their labels, a
neural network can learn the best aggregation method
for these four scores to give one score for each
detected rectangle. At the test stage, all we need to do
is to apply the detectors of our approach on the image,
then we feed the four scores of each detected
rectangle to the trained shallow neural network to
give one confidence score for it.
4 EXPERIMENTAL RESULTS
In this section, we evaluate the proposed method on
WIDER FACE and FDDB datasets using several
criteria in comparison with other methods.
4.1 Choosing the Best Detectors
To adopt the best detectors in our approach, the
detectors that give a better accuracy (a larger number
of true faces), a less number of false faces, and a less
detection time, we study the effect of changing
different detectors’ parameters on the accuracy and
detection time, for more details see Figure 3. For Dlib
HOG-based face detector, it is favourable to run it on
an image scale of 1.5 and a threshold value for its
confidence score equals -1. That is to say, any
detected rectangle has a score larger than or equals -
1, the detector will consider it as a face. For the
OpenCV frontal and profile detector, it is preferable
to run them on an image of scale 1 and scale factor
equals 1.1. The scale factor judges how the detector
changes its sliding window size when searching an
image for faces.
4.2 WIDER FACE Dataset
In this experiment, we use the test set of WIDER
FACE (2400 images) that is formed as described
before, and we adopt AUC (Area Under Curve)
Optimal Score Fusion via a Shallow Neural Network to Improve the Performance of Classical Open Source Face Detectors
663
Figure 3: The effect of changing image scale, scale factors (FSF and PSF of OpenCV frontal and profile face detectors,
respectively), and the threshold value (DThr) of the confidence score of HOG-based face detector on the performance and
time of detection: Image scale 1 on the left and image scale 2 on the right.
concept for ROC curve (Sensitivity vs. 1-Specificity)
as an assessment criterion. As shown in Figure 4, the
AUC of our approach is indeed better than those of
the RSFFD2 and OpenCV face detector.
4.3 FDDB Dataset
In this experiment, we evaluate our approach
compared to OpenCV Haar-based face detectors with
their default parameter values (image scale equals 1
and scale factor is 1.1), Dlib HOG-based face detector
with its default parameter values (image scale equals
1 and the threshold value for its confidence score is
0), RSFFD2, and IterativeHardPositives+ on the
complete FDDB dataset. Note, our shallow neural
network is trained only on the WIDER FACE dataset.
Figure 5 demonstrates an enhancement in
performance of our model compared to the other
detectors. Our model has a maximum true positive
rate of 84.5% which exceeds the rates of the other
detectors in the comparison. In addition, the number
of true faces detected by our approach at any
operating point selected on ROC curves exceeds its
counterpart for the other detectors. All the algorithms
included in the evaluation are implemented and tested
using C++ running on Intel Core I7-6700k CPU, 4
GHz. One key aspect of the proposed approach is that
the additional computational overhead is limited as
demonstrated in Table 1. The table gives the detection
time of five methods in comparison. Furthermore, we
use the same evaluation criterion used by FDDB to
examine the impact of our approach on the gap
between OpenCV and CCN-based face detectors. As
a result of their notability, we use Conv3D (Li et al.,
2016) and HR-ER (Hu and Ramanan, 2017) as
examples of CNN-based face detector. Figure 6
obviously shows that our technique indeed takes a
notable step towards narrowing the gap between the
two types of detectors.
Figure 4: Comparison between our approach, RSFFD2, and
OpenCV frontal face detector on a test set from WIDER
FACE dataset using the AUC (area under the curve) with
the highest AUC (0.97) for our approach.
4.4 Visual Comparison
Figure 7 offers some results for our model and
OpenCV frontal face detector on some images from
FDDB dataset. In all images, our model outperforms
this detector via detecting the same candidate faces
(green rectangles) in addition to new candidate faces
(blue rectangles). Figure 8 presents some results for
our approach and IterativeHardPositives+ face
detector on some images from FDDB dataset.
Although, the two detectors suffer from a few false
VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications
664
Figure 5: Comparison between our approach, RSFFD2,
OpenCV face detector using frontal and profile models, and
Dlib HOG-based face detector running on FDDB dataset
using discrete score.
candidates (red rectangles), our approach beats this
detector via ascertaining the same candidate faces
(green rectangles) in addition to new candidate faces
(blue rectangles).
Table 1: Timing results (in secs) of OpenCV face detectors,
Dlib HOG-based detector, our model, and RSFFD2 running
on FDDB dataset.
Method
Whole FDDB
Per Image
OpenCV Frontal
113.8
0.04
OpenCV Profile
125.2
0.04
HOG-Based
142.25
0.05
Our Approach
256.05
0.09
RSFFD2
6771.1
2.38
5 CONCLUSIONS
In this paper, we have proposed a fusion method to
combine OpenCV and Dlib face detectors in one
detector with the target of enhancing their
performance. It is constructed from simple models
already existing in OpenCV and Dlib libraries, such
as: OpenCV frontal and profile face detector, skin
detector, and Dlib HOG-based face detector.
Furthermore, it employs a shallow neural network to
optimally learn the best aggregation method to
combine all these information sources. We have
examined our approach on the FDDB and WIDER
FACE datasets, and the results have shown that our
adaptations have produced a reasonable increase in
performance. We believe that our approach has taken
a notable step towards narrowing the gap between
classical open source and CNN-based face detectors,
but we are not there yet. As a future task, it would be
Figure 6: Comparison between our approach, OpenCV face
detector, and deep learning-based approaches (Conv3D (Li
et al., 2016) and HR-ER (Hu and Ramanan, 2017)) running
on FDDB using discrete score.
interesting to tighten this gap even further. We think
that there is still some room to increase the number of
detected faces. Our model is only evaluated on frontal
and profile faces, but it could be modified to detect
faces with severe pose variations. Also, our model
could be tested on the other state-of-the-art face
detectors.
ACKNOWLEDGEMENTS
This research is supported by Information
Technology Industry Development Agency of Egypt
(grant # CFP130).
REFERENCES
Bay, H., Ess, A., Tuytelaars, T., and Gool, L. V. (2008).
Speeded-up robust features (surf). In Journal Computer
Vision and Image Understanding, pages 346 – 359.
Science Direct.
Bradski, G. (2000). The opencv library. In Doctor Dobbs
Journal, pages 120 – 126.
Brancati, N., Pietro, G. D., Frucci, M., and Gallo, L. (2016).
Human skin detection through correlation rules between
the ycb and ycr subspaces based on dynamic
color clustering. In Journal of Computer Vision and
Image Understanding. Science Direct.
Cao, X., Wei, X., Han, Y., and Lin, D. (2015). Robust face
clustering via tensor decomposition. In IEEE
Transactions on Cybernetics, pages 2546 – 2557. IEEE.
Dalal, N. and Triggs, B. (2005). Histograms of oriented
gradients for human detection. In International
Conference on Computer Vision and Pattern Recognition
Workshops (CVPR), pages 886 – 893. IEEE.
Optimal Score Fusion via a Shallow Neural Network to Improve the Performance of Classical Open Source Face Detectors
665
a) b)
Figure 7: Detection results on FDDB: a) The output of OpenCV face detector, b) The output of our approach. Green rectangles
exemplify true faces detected by the two algorithms. Blue rectangles symbolize the new true faces detected by our approach.
Red rectangles represent false positive faces. In the images that contain no detected rectangles, the algorithm fails to detect
any face at all.
a) b)
Figure 8: Detection results on FDDB: a) The output of IterativeHardPositives+ (Puttemans et al., 2017), b) The output of our
approach. Green rectangles exemplify true faces detected by the two algorithms. Blue rectangles symbolize the new true faces
VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications
666
detected by our approach. Red rectangles represent false positive faces. In the images that contain no detected rectangles, the
algorithm fails to detect any face at all.
El-Barkouky, A., Rara, H., Farag, A., , and Womble, P.
(2012). Face detection at a distance using saliency
maps. In IEEE Computer Society Conference on
Computer Vision and Pattern Recognition Workshops
(CVPRW), pages 31 – 36. IEEE.
Frejlichowski, D., Gosciewska, K., Forczma, P.,
Nowosielski, A., and Hofman, R. (2016). Applying
image features and adaboost classification for vehicle
detection in the sm4public system. In Image Processing
and Communications Challenges. Springer.
Hu, P. and Ramanan, D. (2017). Finding tiny faces. In
Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition. IEEE.
Jain, V. and Learned-Miller, E. (2010). Fddb: a benchmark
for face detection in unconstrained settings. In technical
report, University of Massachusetts Amherst.
Kim, M., Kumar, S., Pavlovic, V., and Rowley, H. (2008).
Face tracking and recognition with visual constraints
in real-world videos. In IEEE Conference on
Computer Vision and Pattern Recognition, pages 1 –
8. IEEE.
King, D. (2018). In http://dlib.net.
Li, J. and Zhang, Y. (2013). Learning surf cascade for fast
and accurate object detection. In International
Conference on Computer Vision and Pattern
Recognition Workshops (CVPR), pages 3468 – 3475.
IEEE.
Li, Y., Sun, B., Wu, T., and Wang, Y. (2016). Face
detection with end-to-end integration of a convnet and
a 3d model. In European Conference on Computer
Vision (ECCV). Springer.
Parkhi, O., Vedaldi, A., and Zisserman, A. (2015). Deep
face recognition. In Proceedings of the British Machine
Vision Conference. BMVA Press.
Parris, O., Wilber, M., Heflin, B., Rara, H., El-barkouky,
A., Farag, A., Movellan, J., Castriln-Santana, M.,
Lorenzo-Navarro, J., Teli, M., Marcel, S., Atanasoaei,
C., and Boult, T. (2011). Face and eye detection on
hard datasets. In IEEE IAPR International Joint
Conference on Biometrics (IJCB), pages 1 – 10. IEEE.
Puttemans, S., Ergun, C., and Goedeme, T. (2017).
Improving open source face detection by combining
an adapted cascade classification pipeline and active
learning. In TEMPLATE’06, 1st International
Conference on Template Production. VISAPP.
Puttemans, S. and Goedeme, T. (2017). https://iiw.
kuleuven.be/onderzoek/eavise/facedetectiondataset/
home/.
Puttemans, S., Van, W., Ranst, and Goedeme, T. (2016a).
Detection of photovoltaic installations in rgb aerial
imaging: a comparative study. In GEOBIA. GEOBIA.
Puttemans, S., Vanbrabant, Y., Tits, L., and Goedeme, T.
(2016b). Automated visual fruit detection for harvest
estimation and robotic harvesting. In International
Conference on Image Processing Theory, Tools and
Applications (IPTA). IEEE.
Rara, H., Farag, A., Elhabian, S., Ali, A., Miller, W.,
Starr, T., and Davis, T. (2010). Face recognition ata-
distance using texture and sparse-stereo reconstruction.
In Fourth IEEE International Conference on
Biometrics: Theory Applications, pages 1 – 7. IEEE.
Shaikh, F., Sharma, A., P.Gupta, and Khan, D. (2016). A
driver drowsiness detection system using cascaded
adaboost. In Imperial Journal of Interdisciplinary
Research. IJIR.
Tang, X., Du, D., He, Z., and Liu, J. (2018). Pyramidbox:a
context-assisted single shot face detector. In European
Conference on Computer Vision. IEEE.
Tu, C., Lin, M., and Hsiao, S. (2017). Subspace learning for
face verification. In ICASI, International Conference
on Applied System Innovation, pages 582 – 585. IEEE.
Viola, P. and Jones, M. (2001). Rapid object detection using
a boosted cascade of simple features. In Proceedings
of the IEEE Conference on Computer Vision and
Pattern Recognition, pages I–511 – I–518. IEEE.
Wang, H. and Raj, B. (2017). On the origin of deep
learning. In arXiv: 1702.07800v4. arXiv.
Yang, M., Kriegman, D., and Ahuja, N. (2002). Detecting
faces in images: A survey. In IEEE Transactions on
Pattern Analysis and Machine Intelligences, pages 34
– 58. IEEE.
Yang, S., Luo, P., Loy, C., and Tang, X. (2016). Wider face:
a face detection benchmark. In Proceedings of the
IEEE International Conference on Computer Vision
and Pattern Recognition, pages 5525 – 5533. IEEE.
Zafeiriou, S., Zhang, C., and Zhang, Z. (2015). A survey on
face detection in the wild: past, present and future. In
Journal of Computer Vision and Image Understanding,
pages 1 – 24. Science Direct.
Zheng, Y., Yang, C., Merkulov, A., and Bandari, M.
(2016). Early breast cancer detection with digital
mammograms using haar-like features and adaboost
algorithm. In Sensing and Analysis Technologies for
Biomedical and Cognitive Applications. SPIE.
Optimal Score Fusion via a Shallow Neural Network to Improve the Performance of Classical Open Source Face Detectors
667
