Incremental Knowledge Representation Model ... - Semantic Scholar

Neural Information Processing – Letters and Reviews

Vol. 10, Nos. 4-6, April-June 2006

LETTER

Incremental Knowledge Representation Model based on Visual Selective Attention Le Dong1, Sang-Woo Ban2, Inwon Lee1and Minho Lee1 1

School of Electrical Engineering and Computer Science, Kyungpook National University 1370 Sankyuk-Dong, Puk-Gu, Taegu 702-701, Korea 2 Department of Information & Communication Engineering, Dongguk University 707 Seokjang-Dong, Gyeongju, Gyeongbuk, 780-714, Korea E-mail: [email protected], [email protected], [email protected], [email protected] (Submitted on December 28, 2005; Accepted on March 20, 2006) Abstract— This Letter proposes a new framework for representing knowledge regarding objects and it infers new knowledge about novel objects based on the preference selection from natural scenes. A method has been developed and implemented to create a knowledge representation hybrid, with ontology maps, for color and form information. Extensive tests with real data have shown the feasibility of this approach. Keywords— Knowledge representation, selective attention model, autonomous mental development, self organizing feature map, adaptive resonance theory network

1. Introduction Knowledge-based clustering and autonomous mental development remains a high priority research topic, among which the learning techniques of neural networks are used to achieve optimal performance. In this paper, we present a new method that can automatically generate a relevance map from sensory data. The proposed model is based on the understating of the visual what pathway in our brain. The selective attention mechanism, with reward and punishment decides on an interesting object by human interaction, and the self organizing feature map (SOFM) makes clusters for the construction of an ontology map in the color and form domains. The clustered information is relevant for describing a specific object, and the proposed model can automatically generate an inference for an unknown object by using learnt information. The previous attention models include Itti, Koch and Ullman’s proposed model based on “feature integration theory”(Treisman) [1]. This is done by using color, intensity, and orientation as a basis, which is the heuristic integration of bases to make a saliency map [2]. Navalparkkam and Itti proposed a goal oriented attention guidance model to estimate the task-relevance of attended locations in a scene [3]. Walther, Itti, Riesenhuber, Poggio, and Koch proposed a combined attentional selection model for spatial attention and object recognition [4]. Tsotsos et al. proposed a biologically motivated attention model for motion [5]. Sun and Fisher proposed hierarchical selectivity for object-based visual attention [6]. Conventional approaches are restricted to sensory data and object-based fields, while our model leads sensory data to knowledge representation. Our motivation is to mimic and understand the mechanism involved in autonomous mental development with human interaction. We consider that topology information of a selected object to represent knowledge. The main contribution of this paper lies in a new framework for the represent knowledge about objects and is inferred new knowledge, about novel objects based on the preference attention of a natural scene. Furthermore, a method to create a knowledge representation model hybrid, with ontology maps for color and form information, is developed. The inference of new objects from previously perceived knowledge, in conjunction with the perception of color and form can autonomously and incrementally be processed. The main structure of this paper is organized as follows: In Section 2, the proposed knowledge representation model will be described, in which the biological background of the proposed model and the previously proposed, bottom-up saliency map model are explained. This section also deals with the intentional selective attention model, according to human interest. In addition, the knowledge representation model, based on selective attention mechanism is shown. Computer simulation results will follow. Concluding remarks and the direction of future research will be presented in the end. 115

Knowledge Representation based on Visual Selective Attention

Le Dong, Sang-Woo Ban, Inwon Lee, and Minho Lee

Secondary Visual cortex

Frontal lobe Prefrontal association area

Natural Image

Retina

1.[Intensity] 2.[Edge] 3.[Color opponency]

LIP

Where path

Hypothalamus Parieto-occipitotemporal association area Limbic association area IT, V4 Photo receptor cell Lateral Geniculate

Ganglion cell

Visual cortex Redundancy Reduction What path

Nucleus(LGN) 4. [Symmetry]

On-Center, Off-Surround at All area

:Bottom-up :Top-down

Figure 1. Understanding the Biology Processing Mechanism

2. Visual Selective Attention and Knowledge Representation Model In the vertebrate retina, three types of cells are important processing elements that perform edge extraction. They are photoreceptors, horizontal and bipolar cells, respectively [7, 8]. According to well-known facts, edge information is obtained by the role of cells in a visual receptor, and the data are delivered to the visual cortex through the LGN and the ganglion cells. By the output signal of the bipolar cell, the edge signal is detected. On the other hand, a neural circuit in the retina creates opponent cells from the signals generated by the three types of cone receptors [9]. R+G- cell receives inhibitory input from the M cone and excitatory input from the L cone. The opponent response of the R+G- cell occurs because of the opposing inputs from the M and L cones. The B+Y- cell receives inhibitory input by adding the inputs from the M and L cones and an excitatory input from the S cone. These preprocessed signals are transmitted to the LGN through the ganglion cell, and the on-set and off-surround mechanism of the LGN and the visual cortex intensify the phenomena of opponency [9]. Moreover, the LGN and the primary visual cortex detect the shape and pattern of an object. In general, the shape or pattern of an object has symmetrical information, and this is one of the most important features in the construction of a saliency map. Even though the role of the visual cortex, in finding a salient region is important, it is very difficult to model the details of the visual cortex. Due to Barlow’s hypothesis, we simply consider the role of the visual cortex as the redundancy reduction [10]. On the other hand, the lateral intraparietal cortex(LIP) spacing provide a retinotopic spatio-featural map that is used to control the spatial focus of attention and fixation, which is able to integrate feature information in its spatial map [11]. As an integrator of information, the LIP provides the inhibition of return (IOR) mechanism required here to prevent the scan path returning to previously inspected sites [11]. Also, it is clear that several limbic structures, including the hypothalamus, are particularly involved in the affective nature of sensory sensation [11]. These affective qualities are also called reward or punishment, or satisfaction or aversion [11]. The reward and punishment centers undoubtedly constitute one of the most important controllers of our bodily activities, our drives, our aversion and our motivations. According to Guyton’s studies, the limbic system, for reward and punish, has a lot to do with selecting the information that we learn [11]. The frontal lobe integrates information from what pathway and where pathway, and it generates not only an inference from constructing knowledge, but it provides feedback signals to the parietal cortex areas including the LIP areas [11]. Figure 1 shows the basic for the biology processing mechanism.

2.1 Bottom-up saliency map model In order to model the human-like, bottom-up visual attention mechanism [7-9, 11], we used four bases of edge (E), intensity (I), color (RG and BY) and symmetry information (Sym), as shown in Figure 2. The roles of the retina cells and the lateral geniculate nucleus (LGN) are reflected in the previously proposed attention model [12]. The feature maps ( I , E , S , and C ) are constructed by the center surround difference and normalization (CSD & N) of the four bases, which mimic the on-center and off-surround mechanism in our brain. Then, they 116



Figure 2. The architecture of the saliency map model (r: red, g: green, b: blue, I: intensity feature, E: edge feature, S: symmetry feature, RG: red-green opponent coding feature, BY: blue-yellow opponent coding feature, CSD & N: center-surround difference and normalization, I : intensity feature map, E : edge feature map, S : symmetry feature map, C : color feature map, ICA: independent component analysis, SM: saliency map, SP: saliency point, IOR: inhibition of return)

Figure 3. The architecture of the preference and refusal selective attention model using Fuzzy ART network ( I : intensity feature map, E : edge feature map, S : symmetry feature map, C : color feature map, SFA: Fuzzy ART network for symmetry feature map, CFA: Fuzzy ART network for color feature map, EFA: Fuzzy ART network for edge feature map, IFA: Fuzzy ART network for intensity feature map, Ti: intensity feature weight, Te: edge feature weight, Ts: symmetry feature weight, Tc: color feature weight) are integrated by the independent component analysis (ICA) algorithm [12]. In order to consider the shape of an object, we consider symmetry as an additional basis. Symmetry information is obtained by the noise tolerant general symmetry transform (NTGST) method [12]. The ICA can be used to modeling the role of the primary visual cortex for the redundancy reduction, according to Barlow’s hypothesis and Sejnowski’s results [13]. Barlow’s hypothesis is that human visual cortical feature detectors might be the end result of a redundancy reduction process [10], and Sejnowski’s result is that the ICA is the best way to reduce redundancy [13]. After the convolution between the channel of feature maps and filters obtained by ICA learning, the saliency map is computed by the summation of all feature maps for every location [12, 14].

2.2 Intentional selective attention model according to human interest The previous training selective attention model is based on salient areas which are generated by the bottom-up process; therefore areas that are interesting to the human supervisor may not be included in the 117



plausible scan path. It is necessary to seek approaches that generate salient areas, which focus on an intentionally selected interesting area in a static natural scene, as well as incremental learning with human interaction. Figure 3 shows the architecture of the proposed selective attention model that reflects human interest. In the training mode, the inputs of the Fuzzy ART network consist of feature information such as intensity, color, edge and shape characteristics, which are related with the attention area selected by the supervisor. Then, the supervisor decides whether it is a preference area or a refusal area. If the selected area is refusal area, the refusal part of the Fuzzy ART model trains and memorizes the characteristics of that area which are to be ignored in later processing. If some areas are decided to be preference by the supervisor, that area is trained by the preference part of the Fuzzy ART model. After the training process is successfully completed, they memorize the characteristics of the feature information of the refusal areas and preference areas. If a test image contains an area where by feature characteristics occur in the resonance in the Fuzzy ART memory for human preference, the selected area is intensified in the saliency map by endowing a positive weight value to the output of the Fuzzy ART network. This is done for preference which is to be selected as the most salient areas. On the contrary, if the Fuzzy ART memory for human refusal has a resonance for an area in a test image, the related area deteriorates in the saliency map by endowing a negative weight value to the output of the Fuzzy ART network. This is done so as not to be selected as a salient area. The resonance for both Fuzzy ART networks, for preference and refusal, is controlled by setting the threshold for the vigilance value. Also, the weight values for intensifying and deteriorating the saliency map can be trained by a learning algorithm, which will be considered later. In this paper, the weight values are fixed to a value of which the magnitude is obtained by the scaling of vigilance values in each feature map, so that each feature characteristic can be considered respective to reflect the relative importance of feature information in order to construct the intentional selective attention model.

2.3 Knowledge representation model based on selective attention Figure 4 shows a part of what pathway in our visual brain and its functions. The saliency map (SM) selects a candidate of interesting regions using bottom-up features such as intensity, edge, color and symmetry [8]. The hypothalamus model, which is implemented by the Fuzzy ART, generates a modified scan path by reflecting reward and punishment. The selected object information is transmitted to the prefrontal cortex through the V4 and IT areas. The SOFM, based on feature information and a distance matrix method, is used to produce two-

Prefrontal cortex

From sensory data to knowledge representation

Inference module

FEF, Pulvinar

Object representation IT

Form perception

V4

Working Memory

Knowledge representation

Color perception

Knowledge Representation Model based on incremental SOFM

LIP

V1

Preference & Refusal Model combined with Bottom-up Saliency Map LGN ← Retina I (BW)

I

SM

Entropy Maximization ICA

E

CSD

Sym

& N

E

image

RG BY

C

Limbic system Hypothalamus Preference & Refusal representation

Figure 4. Visual what pathway and functions in the brain organs 118

Input

Sym

12



Color abstract Object Labeling/Naming Representation

Color perception

Form abstract Representation

Fact

Form perception

Fuzzy ART + Topology preserving

Primitive features for color perception

Intensity Feature

Inference

Primitive features for form perception

Edge Feature

Color Feature

Symmetry Feature

Saliency Point Preference & Refusal Selective Attention Model Combined with Bottom-up Saliency Map

Natural image

22

Figure 5. Knowledge representation model based on visual information processing dimensional maps with a distance representing relief structure for selected regions. The clustered information in the color and the form domains is used to make an ontology map, which can be integrated with the incremental learning mechanism. This kind of hybrid structure can play a key role in the development of an autonomous mental development system. Exploring the results related with functions in the brain showed that the color and form perception can be used for object representation. Together, they further constitute the knowledge representation module which always interacts with the inference module. All of these modules are related with the functions of the prefrontal cortex and they have a close relationship with the working memory. On the other hand, the preference and refusal representation are associated with the role of the limbic system and especially, with the function of the hypothalamus. The bridge from the sensory data to the knowledge representation lies in the motivation to mimic and understand autonomous mental development through human interaction. Figure 4 ingeniously integrates the bottom-up saliency map model with the preference and refusal representation. Then, the knowledge representation model can reasonably and subtly be yielded. The knowledge representation model based on visual information processing, is described in Figure 5. After natural images are input in to the system, the saliency point was selected through the preference and refusal selective attention model, which is combined with the bottom-up saliency map model. Then, the four feature maps constitute primitive features for the color perception and the form perception respectively. The intensity and color features are employed as the basis for the color perception, while the edge and symmetry features are used as the foundation for the form perception. The intensity and color features are the most basic features to reflect color perception; on the other hand, the edge and symmetry features can be regarded as basis for the form perception in the primitive perspective. The color and form perception were implemented by the Fuzzy ART integrated with the topology preserving mechanism, which was realized by the SOFM unit, as shown in Figure 6. This is where the unit in layer F2 of the Fuzzy ART structure was replaced with the SOFM unit. Topographic maps are used to reflect the results of the SOFM unit. Topographic maps are related with the idea of competitive learning which incorporates the winner node and the neighborhood around the winner node. The transformation of the input pattern space into the output feature space preserves the topology. Neurons in the neighbourhood of the winner node respond to similar inputs. Neurons are tuned to particular input patterns in such a way that they become arranged with respect to each other. A meaningful coordinate system for the different input features is created and spatial locations signify intrinsic statistical features of the input patterns. 119



This approach hopefully enhances the dilemma of stability and plasticity to some extent. Topology can be retained completely for each cluster and at the same time, the total tendency of the input patterns can remain due to the inherence of the Fuzzy ART mechanism. This method can be easily extended to task specific application which can represent and reflect global topology through hierarchical structures and topology relevance. From this, novel objects can be inferred based on the topology from a comprehensive side. In some cases, it is not necessary to explore the topology in one cluster or between different clusters whose topologies belong to various categories. In such situations, usually prior conditions or information related with applications should be informed in advance. For example, regarding round-shape objects such as balls, faces and some single dots, there is no need to set up a specified topology between them even though they share a similar shape. It is wise to set different classification rules rather than to take time by developing procedures to find the relevance between them. Therefore, the advantage of this integrated mechanism is that the stability in the convention Fuzzy ART is enhanced by the introduction of the topology concept using the SOFM, while plasticity is maintained by the Fuzzy ART architecture.

F2 Y1

Y2

……

Yj

……

STM

j

SOFM unit

Top-down

LTM

F1 X1

Add Class

Bottom-Up

wji

X2

……

Xi

wji

Weight update

STM …… X2m

A = ( ai, aiC )

Reset(j)

Y

N

Feature information by Preference selection 18

Figure 6. Fuzzy ART with preserving topology Finally, inferences can be induced from the learnt information facts which consist of object labeling/naming, the color abstract representation and the form abstract representation. We can use the extracted information from the perceptions of both the color and form domains to represent knowledge of objects. Color abstract representation and form abstract information can be regarded as the basis for inferring new objects, while labeling/naming can be designated by the human supervisor or by some given reasoning rules which will not be considered here. Furthermore, it can be concluded that ontology maps can be generated through the conceptual scheme by the given relevance between the represented object and its components, as well as by the inferred novel object and its constitutive elements. The determinate rules should refer to the category of reasoning which will be considered later. There are still some analyses that can be conducted for further relevance maps as to different category fields from the semantic sense or other sources, which also can be hybrid with the structure of Fuzzy ART. The incremental mechanism can be enhanced by full integrations with various hierarchical frameworks [15].

3. Experimental Results Figures 7 and 8 show the simulation results of the proposed selective attention model that can reflect human preference and refusal. In Figure 7, the orange color is decided as the preference feature by the human supervisor. The characteristic of the orange color is trained in the Fuzzy ART network to memorize the color feature of orange as preference one. As shown in Figure 7 (b), after training the preference toward orange color, it becomes the most salient area and several areas with the similar orange color feature are selected as salient 120



(a)

(b)

Figure 7. Simulation results of the preference and refusal attention model: (a) the scan path and the saliency map generated by the bottom-up saliency map model; (b) the scan path and the saliency map generated by the preference & refusal intentional selective attention model after training for the orange color preference.

(a)

(b)

Figure 8. Simulation results of the preference and refusal attention model: (a) the scan path and the saliency map generated by the bottom-up saliency map model; (b) the scan path and the saliency map generated by the preference & refusal intentional selective attention model with the skin color preference mechanism.

S19

S19

S17

S16

S15 S13

S13 S10 S7

0.3-0.6 0-0.3

S7 S5

Red color in red ball

19

16

13

7

S3 10

Winner node

4

S1

S9

1

19

16

13

7

10

4

1

S4

S11

0.3-0.6 0-0.3

S1

Blue color in blue cap

Figure 9. Simulation results of the SOFM for color perception

S19 S16 S13 0.3-0.6 0-0.3

S10 S7

Ball form in red ball

19

16

13

7

10

Winner Node

4

S1

1

19

16

13

10

7

4

1

S4

S19 S17 S15 S13 S11 S9 S7 S5 S3 S1

0.3-0.6 0-0.3

Cap form in blue cap

Figure 10. Simulation results of the SOFM for form perception

121



Inference New Object Red Cap

Red Color

Cap Form

S19 S16 S13 0.3-0.6 0-0.3

S10 S7

Red color in red ball

19

16

13

7

10

4

S1

1

19

16

13

7

10

4

1

S4

Inference New Object Blue Ball

Ball Form

S19

S19

S17

S16

S15 S13 S9

S13 0.3-0.6 0-0.3

S10

S5

S4

Blue color in blue cap

19

16

13

10

7

4

1

19

16

13

10

7

4

S3 S1

0.3-0.6 0-0.3

S7

S7

1

0.3-0.6 0-0.3

Cap form in blue cap

Blue Color

S11

S19 S17 S15 S13 S11 S9 S7 S5 S3 S1

S1

Ball form in red ball

Figure 11. Simulation results of the knowledge representation model regarding the inference of new objects areas, which show that the proposed model can appropriately reflect human preferences. Furthermore, Figure 8 shows the simulation results for skin color preference. As shown in the saliency map in the right-side image in Figure 8 (b), the areas in the skin color becomes much more salient than any other areas due to the preference and refusal mechanism. Moreover, Figure 8 shows that the proposed model can be utilized to localize skin color candidate areas. According to the simulation results, after the training process of the Fuzzy ART network is successfully finished, the proposed model can generate a more plausible scan path as in the human visual attention system, in consideration of human supervisor preference and refusal. We demonstrate the benefit of this novel model by using a real world example from the graphical cluster domain. Clustering and labeling are important in the construction of primitive knowledge. New object clusters can be generated automatically with our proposed knowledge representation model. As shown in Figures 9~11, Euclidean distance is employed to reflect the topology preservation from the perspective of lattice structure and the output dimension is defined as 20x20, as shown in the simulation results. A blue cap and a red ball are employed as training data. Our knowledge representation model can successfully generate new object clusters by simple inference when a red cap and blue ball appear as test objects. This is done by using a combination of learnt factors of both color and form information. In Figure 9 and 10, the winner node signifies the minimum distance between the input patterns and weight vectors for each output unit. As the scaling shows, the distance increases as the location departs from the winner node and the neighbourhood of the winner node is generated by blue and red regions, which can be regarded as 122



the topology preserving knowledge representation and inference. For training data, we detect the topology of the color and form perception of the red ball and the blue cap respectively, which indicated that the red, blue colors, cap and ball forms were represented by topology preserving. This information was detected and expressed from learnt objects and we use this information to infer new knowledge about novel objects. In this case, the blue ball and the red cap were employed as test data. The topology for the test data, from the color and form perception can be combined to comprise inferences of new objects through the comparison of topology representation. The identical color/form should have the similar topology representation and the location of the winner node of the same color/form should be relatively close. Therefore, the blue color and the ball form can detect and infer the blue ball, while the red color and the cap form should be labelled as the red cap. The training data and test data were located at rather complex environments, which did not affect the knowledge representation results of the objects in consideration of the selective attention model. A slight variation in the luminance or the visual angle could affect the knowledge representation results, which also indicated the robustness of our approach.

4. Conclusion and Future Work The proposed knowledge representation model, from sensory data, can be regarded as a framework for autonomous mental development through human interaction, while the proposed attention model can reflect human interests such as preference and refusal through human interaction. The approach can automatically generate relevance maps from sensory data and inferences for unknown objects, by using learnt information based on the visual what pathway in our brain, when the preference & refusal model combined with the bottomup saliency map. The hybrid selective attention mechanism, with the preference and refusal model, can decide on an interesting object by human interaction. The Fuzzy ART structure, with topology preservation, allows clusters to construct an ontology map in the color and the form domains. The next challenge is to consider an incremental representation framework taking into account the ontology, which can be developed with a growing hierarchical SOFM. This dynamically growing architecture can evolve into a hierarchical structure of self-organizing maps, according to the characteristics of the input data. It, therefore, can provide a new perception of the space we are navigating. Also, we are considering a Bayesian approach in order to construct a relevance map in ontology.

Acknowledgment This research was funded by the Brain Neuroinformatics Research Program of the Korean Ministry of Commerce, Industry and Energy (2005).

References [1] A.M. Treisman, G. Gelde, “A Feature-Integrations Theory of Attention,” Cognitive Psychology, Vol. 12, No. 1, pp. 97-136, 1980. [2] L. Itti, C. Koch, and E. Niebur, “A model of saliency-based visual attention for rapid scene analysis,” IEEE Trans. Patt. Anal. Mach. Intell., Vol.20, No. 11, pp.1254-1259, 1998. [3] V. Navalpakkam and L. Itti, “ A goal oriented attention guidance model,” Biologically Motivated Computer Vision, Springer, H. H. Bulthoff, pp. 453–461, 2002. [4] D. Walther, L.Itti, M. Riesenhuber, T. Poggio, and C. Koch, “Attentional selection for object recognition – a gentle way,” BMCV 2002, Tubingen, Germany, Nov. 22-24, pp. 472-479, 2002. [5] J. K. Tsotsos, et al., “Modelling visual attention via selective tuning,” Artificial Intelligence, Vol. 78, pp. 507-545, 1995. [6] Y. Sun, R. Fisher, “ Hierarchical Selectivity for Object-Based Visual Attention,” Lecture Notes in Computer Science, Vol. 2525, Springer-Verlag, Heidelberg, pp. 427-438, 2002. [7] E. Majani, R. Erlanson, and Y. Abu-Mostafa, The eye, Academic, New York, 1984. [8] S.W. Kuffler, J.G. Nicholls, and J.G. Martin, From neuron to brain, Sinauer Associates, Sunderland, U.K, 1984. [9] E. Bruce Goldstein, Sensation and perception, 4th ed., An international Thomson publishing company, USA, 1996.

123



[10] H.B. Barlow and D.J. Tolhust, “Why do you have edge detectors?,” Optical society of America Technical Digest, Vol. 23, No. 172, 1992. [11] A.C. Guyton, Textbook of medical physiology 8th ed., W.B. Saunders Company, USA, 1991. [12] S.J. Park, K.H. An and M. Lee, “Saliency map model with adaptive masking based on independent component analysis,” Neurocomputing, Vol. 49, pp. 417-422, 2002. [13] A. Bell, T.J. Sejnowski, “The independent components of natural scenes are edge filters,” Vision Research, Vol. 37, pp. 3327-3338, 1997. [14] A. Ratnaparkhi, “Maximum entropy models for natural language ambiguity resolution,” Ph.D. Dissertation, Computer and Information Science, Univ. of Pennsylvania, USA, 1998.

Le Dong is currently a Ph. D. candidate, School of Electrical Engineering & Computer Science, Kyungpook National University, Taegu, Korea. Her research interests include brain science and engineering, knowledge representation, neural networks, pattern recognition techniques, and biologically motivated active vision systems.

Sang-Woo Ban received the Ph.D. degree in electrical engineering from the Kyungpook National University, Taegu, Korea in 2006. He is currently an Assistant Professor with the Department of Information & Communication Engineering, Dongguk University, Gyeongju, Gyeongbuk, Korea. His research interest includes brain science and engineering, intelligent sensor systems, neural networks, pattern recognition techniques, and biologically motivated active vision systems.

Inwon Lee is currently a master candidate, School of Electrical Engineering & Computer Science, Kyungpook National University, Taegu, Korea. His research interests include brain science and engineering, motion analysis, neural networks, pattern recognition techniques, and biologically motivated active vision systems.

Minho Lee graduated from the Korea Advanced Institute of Science and Technology in 1995, and is currently a Professor of School in the Electrical Engineering & Computer Science, Kyungpook National University, Taegu, Korea. His research interests include active vision systems based on human eye movements, selective attention, independent component analysis, active noise control, and intelligent sensor systems. (Home page: http://abr.knu.ac.kr)

124