Culture-Dependent
Features Computation Model
for Emotion-based
Image Retrieval
Student Name:
Totok Suhardijanto (Doctor
Student 2nd Year)
Supervisor:
Prof. Kiyoki Yasushi
Student ID:
80849341
Abstract
In this
research, we developed a model for computing cultural contents in an image
retrieval system. The key point in the model is the mechanism in providing the
more accurate information about relationship between color and cultural related
emotion aspects in two cultures or more (C1, C2 … Cn). Generally
speaking, in the research area of information retrieval, it is still difficult
to consider cultural related features into an image search system. Current image
retrieval systems are developed based on low-level features such as texture,
shape, color and position. High-level features such as face recognition,
emotions and culture-differences still remain difficult to be used in more
open-ended task of image retrieval systems.
In this
research, we designed an image retrieval system for a specific domain that is
dealing with color and emotion related features. We implemented a method that
allows an information retrieval system to deal with image based on
culture-dependent color-emotion association. A vector space model was
implemented in this system to create a culture oriented computation that is more
feasible to deal with culture related information across cultures. In order to
realizing this system, image color features and a given culture related
color-emotion characteristics converted into vectored numeric data, then stored
as metadata. By using similarity and weighting computation, the system provides
dynamic information of the association between image color and cultural
emotion.
Background
In
recent days, the computer research community began to consider cultural contents
into their research [12]. One of the fields that attract more attention for the
research community is image retrieval research. Nowadays, there are large
quantities of images and visual information available. These images are existing
in structured collections (e.g. museum collections) or independent (e.g. images
found in Web pages in the form of individuals’ photographs, logos, and so on)
[4]. Collection of images became vary not only in number, but also in types
because nowadays people with different cultural backgrounds enable to share
their images over the world. Along with the trend of cross-cultural
communication issues, an image retrieval system that enables to take
culture-dependent issue into consider has more challenges in near future to
provide better ways and approaches in dealing with culture-dependent images.
With
regard to culture-dependent image retrieval, the issue of association between
image color and emotions or impression is essential. This issue has already been
addressed for many years. It has been attracting many scholars from various
areas of studies. The color emotion oriented image retrieval system has been
already proposed in number of researches [5], [6], [10], [11]. Current computer vision techniques allow us to extract
automatically low-level features of images, such as color, texture, shape and
spatial location of image elements, but it is difficult to extract high level
features automatically, such as names of objects, scenes, behaviors and emotions
[7]. Current image retrieval still has semantic gap in dealing with cross
cultural environments [2], [3]. Although several emotion based image retrieval
(EBIR) systems and emotion semantic image research (ESIR) systems have already
been proposed in [7], [10], [11], these systems have not addressed issues yet
for culture-dependent emotion. In our approaches, we consider cultural features
of color and emotion association as one of key point for assessing cultural
differences in image retrieval system. Figures 1 shows the difference between
the previous model and our model.
EBIR/ESIR
(Emotion-Based Image Retrieval/Emotion-Semantic Image Retrieval) System
Overview
Our
Culture-Dependent Emotion-Based Image Retrieval (CD-EBIR) System
Overview
Figure
2: The Architecture Comparison between ESIR/EBIR and our
CD-EBIR
Objective
In
this research we developed a culture-dependent color emotion model for
cross-culture oriented image retrieval system that realizes color emotion spaces
to search images with human emotion aspects. Our system created color impression
spaces based on Plutchick Model of emotion [8], [9].
We created our method based on the theory of color psychology that approaches
color and emotion association according to basic features of emotion and color.
With
regard to the multicultural reality of our world, there are also always slight
differences of color and emotion association across cultures. This approach
makes our method to be more similar to cross-cultural interaction in natural
way. In the natural cross-cultural interaction, there is always a slight
difference in color and emotion association among cultures. A color that means
bravery in one culture could have different impression in another culture.
Figure 2 illustrates the core system in the proposed model.
Figure
2: Core System (Culture-Dependent Emotion-Based Image Retrieval (CD-EBIR)
Model)
We
generated culture-dependent emotion based image metadata through these following
steps. First, color features are extracted using 3-D Color Vector Quantization
of RGB color space [1]. In this step, color features are converted into RGB
color space. Separately, culture-dependent emotion features are acquired from
cultural knowledge through a survey that is addressed to respondents from three
different cultures including Vietnamese, Indonesian and Japanese. Respondents
were asked to indicate their emotion or impression based on Plutchik model of emotion [8], [9] that associates to
certain colors. This survey result is used for creating culture-dependent color
semantic metadata.
After
the culture-dependent emotion information is obtained, this emotion feature is
then represented as a vector to make it possible to map onto the RGB color space
to correlate with nearest colors. Because a particular color can associate with
different emotions and vice versa, in order to generate dynamically
representative colors of emotion, we use automatic clustering method. After
representative colors are chosen and irrelevant colors are excluded, we used
them in creating culture-dependent color emotion metadata. In our
culture-dependent emotion-based image retrieval, the query emotion is processed
and mapped onto the color-emotion space. We implemented feature weighting and
merging processing in order to measure the similarity for selecting the result.
Figure 3 shows the system architecture of the proposed
system.
Figure
3 The Architecture of CD-EBIR
System
In
order to obtain culture information of color and emotion, we designed and
distributed survey to subjects in three different cultures including Japanese,
Vietnamese, and Indonesian. According to this information, we created metadata
repository of culture-oriented emotion features. Separately we developed a
culture-oriented image feature metadata by extracting color features from a
cultural image such as painting. Based on culture-oriented emotion and color
feature metadata, we created a color-emotion space. We map color and emotion
features onto space to make it possible to measure closest distance between the
features based on the similarity and weight computation. A culture-dependent
emotion-color subspace is implemented to reduce the result lists and to select a
particular culture-oriented features that are closely related to the query. An
automatic merging system is applied to integrate the result lists of collection.
This step also involves performance analysis of the experimental results to
measure performance and quality of the system. We implemented all the components
of our technique and evaluated the prototype in a small scale dataset. The
experimental dataset consists of 1000 images.
Result
Our
culture-dependent emotion-based image retrieval system shows an ability to deal
with image search in three different cultures. This system includes three core
components, that is, color and emotion-space, culture-specific subspace, and
automatic weighting function. In the color-emotion space, the query is processed
by analyzing its emotion-related features and computing it through automatic
feature weighting to correlate them with the correspondent color features. The
information of correlation between emotion and color features are useful to
retrieve a set of images that match to the query. A culture specific subspace
which is implemented is also useful to select a culture that is targeted by the
query. The output of this system is provided into lists of images that are
related to a specific culture.
For
the experimental studies, we implemented our system to image datasets containing
Japanese and Indonesian painting images. We designed a database of Japanese
Ukiyo-e based on Tokyo National Gallery Database and
Indonesian Balinese paintings. The experimental results show that our system can
select and find out a set of culture-oriented emotion-based images by submitting
query with emotion and impression keywords.
Figure
3 Our Web-based Cultural-Dependent Image Search System
In
comparison to human impression based selecting result, the precision rate in our
system is still under 70%. It is still difficult to distinguish colors based on
a set of closest emotion and impression such as joy, contentment, and
satisfaction. However, our system can indicate a set of emotion related colors
in specific culture. For instance, Japanese and Vietnamese happiness relate to a
set of colors between red and blue, whereas Indonesian happiness refers to a set
of colors between yellow and green.
References
[1]
Barakbah, A.R. and Y. Kiyoki.
3D-Color Vector Quantization for Image Retrieval Systems, International Database Forum (iDB),
Iizaka, Japan, 2008.
[2] Semeulders, A.W.M. Content-Based image
retrieval at the end of early years. IEEE Trans. On Pattern Analysis and Machine
Intelligence 22, 12, 2000, pp.
1349-1380
[3]
Zhao, R., William, I. and Grosky.
Bridging the semantic gap in image retrieval. Distributed multimedia databases:
techniques & applications. Idea Group
Publishing Hershey PA USA, 2002, 14-36.
[4]
Kherfi, M.L., Ziou, D. and
Bernardi, A. Image Retrieval From the World Wide Web: Issues, Techniques, and
Systems. ACM Computing Surveys, Vol.
36, No. 1, March 2004, pp. 35–67.
[5]
Sasaki, S., Itabashi, Y., Kiyoki,
Y. and Chen, X. An Image-Query Creation Method for Representing Impression by
Color-based Combination of Multiple Images. Frontiers in Artificial Intelligence and
Applications; Vol. 190, Proceeding of the 2009 Conference on Information
Modelling and Knowledge Bases XX, 2009.
[6] Wang, S. and Wang, X. Emotion
Semantics Image Retrieval: A Brief Overview. ACII 2005, LCNS 3784, 2005, pp.
490-497.
[7] Wang, W.N. and Yu, Y.L. Image
emotional semantic query based on color semantic description. Proceedings of ICMLC 2005, 2005, pp.
4571-4576.
[8] Plutchik, R. Emotion: Psychoevolutionary Synthesis.
New York: Harper, 1980.
[9] Plutchik, R. Emotions and Life: Perspective from
Psychology, Biology, and Evolution. Washington D.C.: American Psychology
Associations, 2003.
[10] Wang, W.N and He, Q.H. A Survey on Emotional
Image Retrieval. Proceeding of ICIP
2008, 2008, pp. 117-120.
[11] Solli, M. and Lenz, R. Color Based
Bags-of-Emotions. CAIP 2009, LCNS
5702, 2009, pp. 573-580.
[12] Tosa, N., Matsuoka, S., Ellis, B., Ueda, H. and
Nakatsu, R. ‘Cultural computing with context aware application: ZENetic
computer’, Lecture Notes in Computer
Science, Vol. 3711, pp.13–23, 2010.