BolortuyaSukh-Erdene
(Sukh-Erdene Bolortuya)
1
조현종
(Hyun-chong Cho)
2†
-
(Interdisciplinary Graduate Program for BIT Medical Convergence, Kangwon National University,
South Korea)
-
(Dept. of Electronic Engineering and Interdisciplinary Graduate Program for BIT Medical
Convergence, Kangwon National University, South Korea)
Copyright © The Korean Institute of Electrical Engineers(KIEE)
Key words
Convolutional neural networks, Face detection, Age estimation, Inception module, Machine learning
1. Introduction
Age estimation is important in social networks to determine whether a person is young
or old. Age is an important attribute in people’s social relations. The ability to
classify age accurately and reliably from facial features remains unsatisfactory in
addressing the needs of social networks and commercial applications. Past approaches
for estimating or classifying age attributes from facial images have relied on differences
in facial feature dimensions or “tailored” face descriptors
(1). Most have employed classification schemes designed especially for age estimation
tasks
(2). Few existing methods were designed to address the challenges of unconstrained imaging
conditions
(3). Moreover, the machine learning methods employed by these methods do not sufficiently
function for the given task. This issue is mainly due to the fact that the facial
appearances of different people of the same age considerably vary.
There are numerous factors affecting one’s facial appearance. First, different people
have different aging rates. Moreover, gender can affect age estimation. As the aging
process is not the same for everyone and depends on several factors such as gender,
race, and living habits, it is very difficult for humans to guess a person’s age by
looking at their pictures
(4). In addition, aging can vary within a person; the facial appearance of a person may
change more slowly in some years but faster in other years. In most cases, the changes
in the appearance of a person within a year are minimal, and it is difficult to tell
if a subject is 40 years old or 41 years old from a single photograph
(5). Moreover, captured facial images are affected by the pose, illumination, and occlusion,
which increase the difficulty of age estimation
(5). In the last few years, deep learning has become the main field for age estimation.
Deep learning has been proven to perform very well for a variety of computer vision
tasks such as human action recognition, handwritten digit recognition, and automatic
face recognition. In relation to the task of soft-biometrics analysis, deep learning
has recently been applied to the task of apparent age estimation. Although a number
of algorithms have been successfully developed for facial age estimation, many challenges
remain. In particular, with the popularization of deep neural networks in computer
vision, a large-scale labeled dataset becomes more important. The commonly used dataset
for age estimation are the FG-NET aging dataset, MORPH dataset
(6), and newly released Adience dataset.
Many studies on age estimation have been conducted. In this paper, we strive to demonstrate
similar advancements using a deep learning architecture that is designed by considering
the limited availability of accurate age labels in existing face dataset. We test
our network on the newly released Adience benchmark age estimation face images
(3). We show that, despite the challenging nature of the images in the Adience dataset,
the simplicity of our network design is evident.
Fig. 1 shows a flowchart of our approach.
Fig. 1. Age estimation flowchart
The problem of automatically extracting age-related attributes from face images has
received increasing consideration in the last few years, and many methods have been
introduced.
Existing methods for age classification are based on calculating the ratios between
different metrics of facial features. Facial features (eyes, mouth, nose, chin, etc.)
are localized, and their sizes and distances are scaled. These methods require accurate
localization of facial features, which alone is a challenging problem. Thus, they
are unsuitable for so-called in-the-wild images, which one may expect to find on social
media platforms. Other methods consider the aging process as a subspace facial aging
pattern, manifold learning, or locally adjusted robust regression. A drawback of these
methods is that they require input images to be close to a full frontal view and well-aligned
(1). These methods thus produce results only for a constrained dataset of near-frontal
images. As a consequence, these methods were shown to be poorly suited for unconstrained
images. Some of the above methods use local features to represent facial images. For
example, Kwon et al
(7) only classified three age groups according to same hand crafted features through
facial skin wrinkle analysis and facial geometry features. Geng et al.
(8) proposed an Aging pattern Subspace (AGES) approach to define an image sequence of
one subject as an aging pattern based on PCA model which obtained the mean absolute
error (MAE) of 6.22 on FG-NET
(9) dataset. Many nonlinear regression approaches, such as quadratic regression
(10), Support Vector Regression (SVR)
(11) and Gaussian Process
(12), have been used for age classification.
Facial age estimation methods based on label-sensitive learning and age-oriented regression
were proposed as improved versions of the relevant component analysis and locally
preserving projections
(2). These methods are used for distance learning and dimensionality reduction, respectively,
with the application of active appearance models as image features. All of these methods
were proven to be effective for small and constrained benchmarks for age estimation.
To the best of our knowledge, the most effective methods were demonstrated with the
“Group Photos” benchmark for understanding the images of groups of people. For the
age estimation of unfiltered faces, state-of-the-art performance on this benchmark
was achieved by employing local binary pattern (LBP) descriptor variations
(1) and a dropout support vector machine (SVM) classifier.
This paper provides a methodology to estimate the real age groups of human by analyzing
frontal face images. This process involves four stages: face detection, pre-processing,
classification, and age estimation. The system uses two types of face images: training
face images where age group is known and test face images where the age group is unknown.
We show that our proposed method, applied to the more challenging Adience benchmark
designed for the same task, outperforms other recent methods.
2. Materials and Methods
2.1 Database
In this study, we used the Adience dataset for training and testing. Adience images
attempt to resize all of the in-visage features, the lighting quality, the head pose,
noise, and other aspects. In addition, the images of the dataset were obtained without
careful preparation or posing. We used 4,000 facial images from the Adience dataset
(1). It includes 1,723 male, 1,777 female, and 500 baby facial images
(1). Each image is added with the person’s age range (eight possible ranges)
(1).
Table 1 summarizes the details of the dataset used. The dataset was acquired from the Computer
Vision Lab at the Open University of Israel(OUI)
(1). We selected images that were generally front-facing, i.e., 4,000 facially detected
images.
Table 1. The dataset
|
Group label
|
Age group in years
|
Male
|
Female
|
Number
|
|
1
|
0~2
|
-
|
-
|
500
|
|
2
|
4~6
|
259
|
241
|
500
|
|
3
|
8~13
|
236
|
264
|
500
|
|
4
|
15~20
|
256
|
244
|
500
|
|
5
|
25~32
|
238
|
262
|
500
|
|
6
|
38~43
|
259
|
241
|
500
|
|
7
|
48~53
|
257
|
243
|
500
|
|
8
|
60+
|
218
|
282
|
500
|
|
|
Total
|
1,723
|
1,777
|
4,000
|
2.2 Face Detection and Preprocessing
Face detection is the basis of the method described by Viola and Jones
(13) using Haar-like features. The Viola and Jones face detection method has four steps
(14): selection of Haar-like features; creation of an integral image; AdaBoost training;
and cascade classification. Most human faces have some similar features (nose, eyes,
mouth, etc.) and Haar-like features that match them. Each Haar-like feature composed
of two, three, or four rotatable black and white rectangles
(14). Integral image is an intermediate image representation. Haar-like features can be
computed rapidly using integral image. AdaBoost is a weight-updating algorithm that
selects a few number of strong features from a large set of possible features
(15).
Cascade classifier has a several stages. Each stage makes a newly weak learner. The
cascade classifiers pass over the background regions of the image to be quickly discarded
while spending more calculation time on regions that are more promising as being facial
ones. The cascade classifiers having more features will achieve higher detection rates
and lower false positives.
After face detection, our system crops the detected facial area because the cropped
facial area is more effective forage cap age estimation. Facial area cropping reduces
the image background and other noise. The cropped facial image then proceeds to histogram
equalization to balance the image contrast because every image may not have good contrast
between bright and dark. Next, an image sharpening method is used to distinguish between
different colors. A rapid transition from black to white appears sharp. A series conversion
from black to gray to white appears blurry. Sharpening images increases the contrast
along the edges where different colors meet
(16). Thus, facial wrinkles and age differences are more effectively shown. The basic
sharpening method uses the two-dimensional (2D) finite impulse response (FIR) filter
and Lab color space transformations. First, an RGB image is changed to the Lab color
space. Then, only the L channel passes through the 2D FIR filter, which is a high-pass
filter.
Fig. 2 shows face detection and preprocessing.
Fig. 2. Preprocessing: (a) input image, (b) detected and cropped face, (c) after applying
histogram equalization, and (d) after applying the sharpening filter
2.3 Convolutional Deep Leaning Based on an Inception Module
We used a deep convolutional neural network architecture code named Inception that
achieves the new state of the art for classification, GoogLetNet
(17). It has 22 layers with nine inception modules. The inception module is shown in
Fig. 3. The idea of the inception layer is to cover a larger area while also retaining a
fine resolution for minor information in the images. Thus, it convolve the different
sizes in parallel from the most accurate details (1×1) to a larger one (5×5)
(17). The basic principle is that a Gabor filter with different sizes will better handle
numerous object scales with the benefit of all filters assuming that the inception
layer is learnable on the inception layer being learnable
(17). The most thorough approach for increasing the quality for deep learning is to use
larger layers and a greater amount of data. However, the existence of numerous parameters
also means that the model is more susceptible to over-fitting. Thus, to avoid an increase
in the number of parameter in the inception layers, all bottleneck procedures are
exploited
(17). The bottleneck procedure consists of 1×1, 3×3, and 5×5 convolutions, which appear
as a bottleneck. Using the bottleneck processes, we can remake the inception module
with more non-linearities and fewer parameters
(17). Additionally, a max pooling layer is added to the sum of the content of the forward
layer. All of the results are associated one after the other and transferred to the
next layer. The input size of the perceptivity field in the convolutional deep-learning-based
inception module is 224×224 in the RGB color map. It has 22 layers and almost twelve-fold
fewer parameters (it is faster than AlexNet, which is the first convolutional neural
net architecture and much more accurate).
Fig. 3. Inception module with dimension reductions
2.4 Training and Testing
The Adience benchmark is one of the most recent dataset designed for age estimation
from face images. In this study, the Adience dataset was used to evaluate the efficiency
of the proposed method. It consists of the unconstrained face images of 2,284 subjects
and has eight age groups. We used 4,000 face images, and each age group comprised
500 images. We cleaned the dataset to remove images in which faces were not fully
viewable. The training dataset used 3,200 images, with 400 in each of the eight age
groups. The testing set comprised 800 images, and each age group had 100 images. The
initial learning rate was 0.0001 with mini-batches of 10 images and 10,000 iterations.
The training dataset used k-fold cross validation. In k-fold cross validation, the
dataset is first divided into five equally sized segments. Accordingly, in each iteration,
another fold of the dataset was held out for validation, while the remaining k−1 folds
were used for training
(18).
The dataset was generally stratified prior to being divided into k folds. Stratification
ordered the data to ensure that each fold was representative of the whole. In our
experiment, each test class comprised 20% of the data. Our method was implemented
using the Matlab (2017a) open- source framework. Training was performed using a GeForce
GTX 750Ti with 640 CUDA cores and 8GB of video memory. Training each network required
approximately 2 h, and predicting the age for a single image using our network required
approximately 450ms.
We compared four methods: “LBP + FPLBP + Dropout 0.5,” “LBP + FPLBP + Dropout 0.8,”
“AlexNet
(19),” and “Deep-CNN
(1).” The face representation for the first two methods uses the LBP and four-patch LBP
(FPLBP)
(3). The standard linear dropout SVM was used for classification. The LBP description
of a pixel image was produced by thresholding the 3×3 neighborhood with the central
pixel and devolving the result as a binary code
(20). The LBP operator has been extended to use different neighborhoods sizes by allowing
different sampling points and a larger neighborhood radius.
The FPLBP is another variation of the LBP
(20). In the FPLBP, two rings with different radii are used and centered at the pixel.
The 3×3 patches are distributed around these two rings, and two center symmetric patches
in the inner ring are compared with two center symmetric patches in the outer ring
positioned patches away along the circle.
An SVM is a discriminative classifier formally defined by a separating hyperplane.
Given labeled training data in supervised learning, the SVM outputs an optimal hyperplane
that categorizes new examples. “Dropout 0.5” denotes a dropout SVM with a 50% probability
of dropping features, and “Dropout 0.8” means that 80% of the input features are dropped,
randomly and independently, from the input feature vector.
AlexNet is the name of a convolutional neural network that competed in the ImageNet
Large Scale Visual Recognition Challenge. AlexNet has eight layers. The first five
are convolution layers, and last three are fully connected layers. The first convolutional
layer performs convolution. Its size is 11×11, and the max pooling size is 3×3 with
local response normalization (LRN). The same operations are performed in the second
convolution layer with a size of 5×5. The size of the third, fourth, and fifth convolutional
layers is 3×3. Two fully connected layers are used with dropout followed by softmax
layers at the end
(19).
Deep-CNN
(1) has five layers. The first three are convolutional layers, and the last two are fully
connected layers. The first convolutional layer is an input of 7×7 pixels, ReLU (Rectified
Linear Unit), a max pooling layer taking the maximal value of 3×3 regions with two-pixel
strides, and a local response normalization layer. The size of the second convolutional
layer is 5×5 pixels, and it consists of ReLU, a max pooling layer, and local response
normalization. The size of the last convolutional layer 3×3 pixels, and it consists
of ReLU and a max pooling layer. The first fully connected layer receives the output
of the last convolutional layer and contains 512 neurons, ReLU, and a dropout layer.
The second fully connected layer contains 512 neurons, followed by ReLU and a dropout.
The last fully connected layer maps the final classes for the age. The output of the
last fully connected layer is fed into a softmax layer that assigns a probability
for each class. Deep-CNN
(1) is a smaller network design motivated to reduce the risk of over-fitting.
3. Results
Experiments were conducted to evaluate the proposed method. Our proposed method includes
“face detection,” “histogram equalization,” an “image sharpening filter,” and a “convolutional
deep-learning-based inception module.”
Table 2 presents a matrix of the multi-class age estimation, and a comparison of the performance
of the proposed method and recent related methods is presented in
Table 3. The
Table 3 is shows the latter shows the exact accuracy and so-called one-off accuracy, which
is where the result is off by one adjacent age label. The exact accuracy means that
our method classifies the age group correctly. The one-off accuracy includes one adjacent
age group, above or below. On the basis of our results, the proposed method significantly
outperforms in terms of the exact accuracy and one-off accuracy. These results confirm
the efficiency of the proposed work.
Table 2 indicates that the “60+ years” age label is classified with the highest accuracy
of 76%. The labels of “8-13,” “38-43,” and “48-53” years of age are classified with
the lowest accuracies of 32%, 38%, and 29%, respectively. The label of “48-53” years
of age shows the lowest accuracy. The “48-53“ age group is misclassified into ”60+”
group as 43%. One of possible reasons is that the facial shape of these two age groups
are very similar. And “48-53“ age group does not have their own facial features (e.g.
wrinkles) which can distinguish them from ”60+” group.
Table 2. Confusion matrix of Facial age estimation
|
Group label
|
1
|
2
|
3
|
4
|
5
|
6
|
7
|
8
|
|
Actual (age group)
Estimated (age group)
|
0~2
|
4~6
|
8~13
|
15~20
|
25~32
|
38~43
|
48~53
|
60+
|
|
0~2
|
0.69
|
0.5
|
0.03
|
0.02
|
0.02
|
0
|
0
|
0
|
|
4~6
|
0.28
|
0.46
|
0.12
|
0.05
|
0
|
0.01
|
0
|
0
|
|
8~13
|
0.01
|
0.03
|
0.32
|
0.11
|
0.02
|
0.03
|
0.01
|
0.03
|
|
15~20
|
0.02
|
0
|
0.46
|
0.65
|
0.11
|
0.07
|
0.03
|
0.02
|
|
25~32
|
0
|
0
|
0.03
|
0.12
|
0.54
|
0.21
|
0.07
|
0.05
|
|
38~43
|
0
|
0
|
0.03
|
0
|
0.29
|
0.38
|
0.17
|
0.03
|
|
48~53
|
0
|
0.01
|
0.01
|
0.05
|
0.01
|
0.12
|
0.29
|
0.11
|
|
60+
|
0
|
0
|
0
|
0
|
0.01
|
0.18
|
0.43
|
0.76
|
Table 3. Overall age estimation results with different approaches
|
Method
|
Exact accuracy
|
One-off accuracy
|
|
LBP + FPLBP + Dropout 0.5(3)
|
44.5%
|
80.6%
|
|
LBP + FPLBP + Dropout 0.8(3)
|
45.1%
|
79.5%
|
|
Deep-CNN(1)
|
50.7%
|
84.7%
|
|
AlexNet(19)
|
31.75%
|
66.12%
|
|
Our Proposed method
|
51.12%
|
89.37%
|
Correctly estimated images are shown
Fig. 4, and
Fig. 5 shows incorrectly estimated images. In
Table 4, we tested 4000 images with histogram equalization, sharpened and original.
Table 4 shows preprocessing filter can increase the age estimation accuracy.
Fig. 4. Correctly estimated images for age ranges of (a) 0-2, (b) 4-6, (c) 8-13, (d)
15-20, (e) 25-32, (f) 38-43, (g) 48-53, and (h) 60+.
Fig. 5. Incorrectly estimated images. Actual age group (estimated age group): (a)
0-2 (4-6), (b) 4-6 (0-2), (c) 8-13 (48-53), (d) 15-20 (48-53), (e) 25-32 (8-13), (f)
38-43 (15-20), (g) 48-53 (25-32), and (h) 60+(8-13).
Table 4. Overall age estimation results with different pre- processing
|
Input image type
|
Exact accuracy
|
One-off accuracy
|
|
Original images
|
42%
|
78.81%
|
|
Histogram images
|
46.62%
|
84.25%
|
|
Sharpen images
|
48.25%
|
84%
|
|
Our proposed method
|
51.12%
|
89.37%
|
From
Table 2, the baby faces of the first two labels (“0-2” and “4-6”) look very similar and are
difficult to estimate. The groups associated with Labels 6 (38-43) and 7 (48-53) are
classified with accuracies of 38% and 29%, respectively. It is difficult to estimate
the age of people in their 30s and 40s. However, the one-off accuracies of Labels
6 and 7 are 71% and 89%, respectively. The oldest group, label 8 (60+) was classified
to 76%, indicating that older people’s facial features are distinguishable. For this
range of ages, people’s facial wrinkles are more distinctive than those in other age
groups.
We also tested underage people under 20 years of age and gender-based age estimation.
From
Table 5, the exact accuracy and one-off accuracy of males are higher than those of females.
As a consequence of the influence of makeup, it can be more difficult to estimate
the real age of women. In k-fold cross-validation, 377 female images and 323 male
images were used for the test. Further, Label 1 (0-2) was excluded in this experiment
because it is very difficult to estimate the gender for this age group.
Table 5. Gender-based age estimation
|
|
Male
|
Female
|
|
Exact accuracy
|
52.95%
|
49.14%
|
|
One-off accuracy
|
89.42%
|
86.69%
|
The results for underage estimation for the underage group are listed in
Table 6.
Fig. 6 shows some correctly estimated face images. Labels 1-4 are in the underage group.
This experiment might help to control illegal underage activities such as entering
bars or wine shops and purchasing tobacco products.
Fig. 7 shows incorrectly estimated face images as adult’s age for the underage group. Adult
age estimation for people over 20 years of age is presented in
Table 7. Labels 5-8 are in the adult age group. The results of adult age estimation are slightly
lower than those for underage estimation.
Table 6. Underage estimation for the underage group
|
Misclassified
|
Correctly classified
|
|
6.25%
|
93.75%
|
Fig. 6. Correctly estimated as underage. The age groups are (a) 15-20, (b) 15-20,
(c) 15-20, and (d) 15-20
Fig. 7. Incorrectly estimated as adult age. The actual (estimated) age groups are
(a) 15-20 (48-53), (b) 15-20 (25-32), (c) 15-20 (38-43), and (d) 15-20 (38-43)
Table 7. Adult age estimation for the adult age group
|
Misclassified
|
Correctly classified
|
|
8.75%
|
91.25%
|
Table 8 summarizes the results for Labels 4 (15-20) and 5 (25-32). These two age groups are
at the borderline between adult and underage. “Correctly classified” in
Table 8 means that the age group was correctly classified as adult or underage, even though
the age label cannot be precisely estimated.
Table 8. Age estimation of borderline groups
|
Age group in years
|
15-20
|
25-32
|
|
Exact accuracy
|
65%
|
54%
|
|
Correctly classified
|
83%
|
85%
|
4. Conclusions
In this paper, we proposed a method to perform age estimation based on facial images
by using a convolutional neural network based on an inception module. It was modified
and fine-tuned to perform age estimation, and it was trained for face detection for
a large dataset. The proposed method outperformed the previous approach for the Adience
dataset by 4%, which is the newest challenging age estimation benchmark consisting
of unconstrained facial images. In addition to facial age estimation, we tested underage
and gender-based age estimation. Underage estimation helps to prevent illegal underage
activities such as entering bars and purchasing alcohol or tobacco products.
Our method significantly outperforms recent related approaches. These results provide
a remarkable baseline for deep-learning-based approaches. The proposed method can
be improved with more data sets. Future work will focus on obtaining a higher accuracy
for age estimation with large data sets and the development of a gender estimation
architecture for the proposed method.
Acknowledgements
This work was supported by the National Research Foundation of Korea(NRF) grant funded
by the Korea government(MSIT) (No. 2017R1E1A1A03070297). This research was supported
by the MSIT(Ministry of Science and ICT), Korea, under the ITRC(Information Technology
Research Center) support program(IITP-2018-0-01433) supervised by the IITP(Institute
for Information & communications Technology Promotion).
References
Levi G., Hassner T., 2015, Age and gender classification using convolutional neural
networks, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition
Workshops, pp. 34-42
Chao W. L., Liu J. Z., Ding J. J., 2013, Facial age estimation based on label-sensitive
learning and age- oriented regression, Pattern Recognition, Vol. 46, pp. 628-641
Eidinger E., Enbar R., Hassner T., 2014, Age and gender estimation of unfiltered faces,
IEEE Transactions on Information Forensics and Security, Vol. 9, pp. 2170-2179
Hosseini S., Lee S. H., Kwon H. J., Koo H. I., Cho N. I., 2018, Age and gender classification
using wide convolutional neural network and Gabor filter, in International Workshop
on Advanced Image Technology 2018 (IWAIT 2018)
He Y., Huang M., Miao Q., Guo H., Wang J., 2017, Deep embedding network for robust
age estimation, in 2017 IEEE International Conference on Image Processing (ICIP),
pp. 1092-1096
Hu Z., Wen Y., Wang J., Wang M., Hong R., Yan S., 2017, Facial age estimation with
age difference, IEEE Transactions on Image Processing, Vol. 26, pp. 3087-3097
Kwon Y. H., 1994, Age classification from facial images, in Computer Vision and Pattern
Recognition, 1994. Proceedings CVPR'94, 1994 IEEE Computer Society Conference on,
Vol. 권, No. 호, pp. 762-767
Geng X., Zhou Z. H., Zhang Y., Li G., Dai H., 2006, Learning from facial aging patterns
for automatic age estimation, in Proceedings of the 14th ACM inter- national conference
on Multimedia, pp. 307-316
Lanitis A., Cootes T., 2002, Fg-net aging data base, Cyprus College, Vol. 2, No. 5
Guo G., Fu Y., Dyer C. R., Huang T. S., 2008, Image- based human age estimation by
manifold learning and locally adjusted robust regression, IEEE Transactions on Image
Processing, Vol. 17, pp. 1178-1188
Guo G., Mu G., Fu Y., Huang T. S., 2009, Human age estimation using bio-inspired features,
in Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pp.
112-119
Xiao B., Yang X., Zha H., Xu Y., Huang T. S., 2009, Metric learning for regression
problems and human age estimation, in Pacific-Rim Conference on Multimedia, pp. 88-99
Viola P., Jones M. J., 2004, Robust real-time face detection, International Journal
of Computer Vision, Vol. 57, pp. 137-154
Bolortuya S. E., Kim M. J., Cho H. C., 2016, A study of automatic face detection system
for side-view face images, in Information and Control Symposium, pp. 117-118
Viola P., Jones M., 2001, Rapid object detection using a boosted cascade of simple
features, in Proceedings of the 2001 IEEE Computer Society Conference on Computer
Vision and Pattern Recognition, pp. I-I
Ng Y. S., Tai H. T., 2006, Edge enhancement of gray level images, US Patent No. US7079287B1
Szegedy C., Liu W., Jia Y., Sermanet P., Reed S., Anguelov D., Erhan D., Vanhoucke
V., Rabinovich A., 2015, Going deeper with convolutions, in 2015 IEEE Conference on
Computer Vision and Pattern Recognition (CVPR)
Refaeilzadeh P., Tang L., Liu H., 2009, Cross-validation, in Encyclopedia of Database
Systems, Springer, pp. 532-538
Krizhevsky A., Sutskever I., Hinton G. E., 2012, Imagenet classification with deep
convolutional neural networks, in Advances in Neural Information Processing Systems,
pp. 1097-1105
Wolf L., Hassner T., Taigman Y., 2008, Descriptor based methods in the wild, in Workshop
on Faces in 'Real- Life' Images: Detection, Alignment, and Recognition
저자소개
Bolortuya Sukh-Erdene (Bolortuya Sukh-Erdene)
Received the M.S. degree in the Department of Electronic Engineering & Interdisciplinary Graduate Program for BIT Medical Convergence, Kangwon National
University, South Korea.
His research interests include image processing and deep learning.
Received the M.S. and Ph.D. degrees in Electrical and Computer Engineering from the
University of Florida, USA in 2009.
During 2010~2011, he was a Research Fellow at the University of Michigan at Ann Arbor,
USA.
From 2012 to 2013, he was a Chief Research Engineer in LG Electronics, South Korea.
He is currently an assistant professor at Kangwon National University, South Korea.