Novel Noncontrast-Based Edge Descriptor for Image ... - IEEE Xplore

1086

IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 16, NO. 9, SEPTEMBER 2006

Novel Noncontrast-Based Edge Descriptor for Image Segmentation Byung-Gyu Kim, Member, IEEE, and Dong-Jo Park, Member, IEEE

Abstract—We present an efficient video segmentation strategy based on new edge features to assist object-based video coding, motion estimation, and motion compensation for MPEG-4 and MPEG-7. The proposed algorithm utilizes the human visual perception to provide edge information. Based on the human visual perception, two edge features are introduced and described based on edge features from analysis of a local histogram. An edgeness function is derived to generate the edgeness information map by using the defined features, which can be thought as the gradient image. Then, an improved marker-based region growing and merging techniques are derived to separate the image regions. The proposed algorithm is tested on several standard images and demonstrates high efficiency for object segmentation. Index Terms—Edge information, histogram analysis, human visual perception, image segmentation, object segmentation.

should be applied for best results. Also, edges are often missed where the change of intensity or color is gradual (or weak). We introduce an automatic VO segmentation algorithm based on newly defined edge information. Analysis of local histogram properties is performed to define edge parameters in the analysis window for every pixel. Use of the modality and the defined parameters of the given histogram allows definition of a measure that can provide edge information as a gradient map [14], [17], [18]. This paper is organized as follows. Section II presents the newly defined edge features for image segmentation. Based on the defined features, an improved region-growing method is derived in detail in Section III. The proposed algorithm is tested on several images and gives results of the image segmentation in Section IV. Conclusions are presented in Section V.

I. INTRODUCTION

O

BJECT-BASED video coding is a distinct feature of the MPEG-4 or MPEG-7 standards. It was unavailable in the earlier MPEG-1 and MPEG-2 standards. These newer standards introduce the concept of the video object layer (VOL) to support content-based functionality at the user decoder [1], [2]. Under the concept of VOL, segmentation of VOs in the MPEG-4 and MPEG-7 coding schemes is inevitable in a frame. In most algorithms for VO segmentation, moving objects in the time domain are usually considered to be meaningful objects. Thus, many developed algorithms for VO segmentation are based on the motion field and change detection between consecutive frames [4]–[9]. The motion field can be used but it is noise sensitive and computation is expensive. Various algorithms that are based on edge and color information using color clustering, watershed, or edge masks have been suggested to detect the boundary of the VO [3], [4], [10]–[18]. Approaches based on data clustering are iterative and require exhaustive computation. Edge detection based on edge masks requires spatial convolution for every pixel. In many cases of natural images, there are no perfect closed boundaries by using these edge masks. Thus, some additional postprocessing such as edge linkage, edge clustering, and removal of isolated edges Manuscript received May 23, 2006. This paper was recommended by Associate Editor H. Watanabe. B.-G. Kim is with the Real-time Multimedia Research Team, Embedded Software Technology Center, Electronics and Telecommunications Research Institute (ETRI), Daejeon 305-350, Korea (e-mail: [email protected]; [email protected]). D.-J. Park is with the Department of Electrical Engineering and Computer Science, Korea Advanced Institute of Science and Technology (KAIST), Daejeon 305-701, Korea (e-mail: [email protected]). Color versions of Figs. 3–7 and 9 are available online at http://ieeexplore. ieee.org. Digital Object Identifier 10.1109/TCSVT.2006.879991

II. PROPOSED EDGE FEATURES FOR IMAGE SEGMENTATION An edge is usually defined by the magnitude of a convolution result with spatial masks in the image plane [4]–[7]. Neighboring pixels around the pixel point in question (current pixel) are convoluted with spatial masks to measure how much dose it contain the edge information. According to the magnitude of the current pixel point, a pixel is classified as either an edge or nonedge pixel. This edge definition approach is computationally exhaustive since the convolution should be applied to every pixel in the frame. If the edge is not well defined, postprocessing should be applied, including the edge clustering, removing and linkage techniques. We define the edge features for the edge information based on an analysis of the local histogram rather than using spatial masks as described in [22]. An edge can be considered as adjacent regions having two or more distinct brightness or color values. If the intensity distribution of an image is multimodal, the image is probably edge-contained. Otherwise, the image has no edge information. Figs. 1 and 2 show examples of the intensity distribution of gray levels in a small analysis window. We can see that the distribution of the edge-contained image can be distinguished from the distribution of a nonedge containing image. A. Description of Edge Features These parameters should specify the properties of the local histogram of the window image. For simplicity, we describe the features by employing the case of bimodal histogram. Then, we will extend to the multimodal case. : For the given image with two regions, the • Region ratio distinctiveness between regions also depends on the areas of two regions. This is called as the areal effects [23]. If we assume that two regions (object and background) exist in

1051-8215/$20.00 © 2006 IEEE

KIM AND PARK: NOVEL NONCONTRAST-BASED EDGE DESCRIPTOR

1087

Fig. 2. Example of the intensity distribution for the edge-contained image.

Fig. 1. Example of the intensity distribution for the nonedge image.

the given image, one may be considered as an object and the other as the background. An inverse relationship between the size of an object and its perceptible luminance has been reported [23]. Objects with small retinal images have high luminance thresholds and objects with larger retinal images have lower perceptible thresholds. If the distribution of the image is bimodal or multimodal, the region ratio is described based on the number of pixels in the two distinct regions. In Fig. 3, the distribution of the gray level shows that the image has two distinct regions. Under this situation, the can be written as follows: region ratio

Property 1: If the distribution is unimodal, the region ratio is zero. If the distribution is multimodal, the region ratio is in the . range of From the above property, the region ratio has the largest value is equal to the area of the for when the area of the any . We can also see that the distinctiveness between image regions increases as the region ratio increases. : In an image with several regions, the • Edge potential homogeneity of a region becomes larger as the variation of the intensity decreases. Thus, we introduce another parameter to take into account the variation of the intensity in the region. For any local minimum , the edge potential can be expressed as follows: (4)

for

(1)

where is the gray level of the th local minimum, is the number of the detected local minima. For any , , and are defined as follows: (2) (3) where is the intensity level and level .

is the probability of the

where is the probability of the maximum peak on the is the probability of the maximum left-side for , is the probability at peak on the right-side for and the local minimum . Property 2: If the distribution is unimodal, the edge potential is zero. And if the distribution is multimodal, the edge potential . is in the range of If the distribution of the intensity under the probability of the given local minimum is fixed, decrease of the edge potential causes growth in the number of intensity levels that can exist in the image. This means that the homogeneity of a region of the can become given image decreases. Thus, the edge potential large as the value of the local minimum decreases.

1088


Fig. 5. Variation of the edge potential P with the fixed region ratio R.

Fig. 3. Region ratio (R) and edge potential (P ).

Fig. 4. Variation of the region ratio R with the fixed edge potential P .

B. Analysis of the Designed Edge Features We analyze behaviors of the defined features to obtain edge information in this subsection. To do this, it is assumed that one feature is independent of the other. Also, two Gaussian shapes were used to explain the characteristics of features as the bimodal case. This is easily accomplished by the Gaussian smoothing kernel. under conFig. 4 shows the behavior of the region ratio dition that the edge potential is fixed with a constant value. Let an arbitrary region ratio be and the variation of the number , the changed value of the region of the occupied pixels be ratio by the variation can be written as the following:

relatively , but that of the other re. Thus, gion goes larger simultaneously the visual distinctiveness may maintain to almost constant from viewpoint of the variance of the region. In this case, the visual distinctiveness has been affected by the areal effects [23]. As mentioned in earlier, an inverse relationship between the size of an object and its perceptible luminance has been reported [23]. That is, we can see that the distinctiveness between image regions increases as the region ratio increases. , For analyzing the characteristic of the edge potential we fixed the region ratio to constant value as illustrated in Fig. 5. As we can see from this figure, the edge potential has heavy relation to the variance of a region, too. The edge of one region becomes larger, the variance of the potential distribution of that has smaller under the condition of the constant region ratio. The variance of other region has no change in the intensity. Usually, the homogeneity of a region becomes larger as the variation of the intensity decreases. Therefore, the of one region becomes larger, the distinctiveedge potential ness between regions increases visually from viewpoint of the variance of the region. Based on the analysis of the defined features and , the relationships between features of a region and the visual distinctiveness are as follows: visual distinctiveness

(6)

visual distinctiveness

(7)

Fig. 6 shows the responses of both parameters. To extract the semantic edge information by using the analyzed properties and responses of those, we modeled the desired edgeness function as the Gaussian distributions with several variances as the following:

(5)

(8)

As shown in Fig. 4, the variation of the amount gives arise to the change of the variance of the other region. The region rabecomes larger, the variance of one region has smaller tion

and are the maximum values of feature pawhere rameters which are given as and in Properties, respectively.


1089

TABLE I PERFORMANCE COMPARISON WITH THE WATERSHED-BASED METHOD WITH THE CANNY OPERATOR FOR SYNTHETIC IMAGES

Fig. 6. Actual responses of (a) the region ratio R and (b) the edge potential P .

Here, we should select the variance of the desired Gaussian model. As we see in Fig. 6, the variance of the Gaussian model depends on the size of the analysis window through the actual response in our study. Especially, the variance of each model is , where determined to fit the actual response as denotes the size of the window for the histogram analysis. The size of the analysis window determines the size of the image region that can be detected. Thus, the size of the window should be chosen based on features of the image. The presented edgeness function may be very useful to localize the edge information because of being without a term of the contrast. This means that the proposed edge features, and , have ability to extract the edge information although the difference of brightness between adjacent two regions is very weak in many cases. For case of multiple valleys in the histogram, two feature parameters can be determined as for each local minimum point . If there are two local minimum points and , two

sets of parameters and are computed for and . each minimum After obtaining these feature parameters, the edgeness funccan be computed by the (8) for each local minimum tion . Thus, the edgeness function can exist as many as the number of local minimum dose. By the definition of the edgeness function, this measure becomes larger as parameters go to their maximum values. So, we take the maximum value among ’s as the edge response the computed the edgeness function of the given window image at the current pixel. , the amount of the edWith the edgeness function geness with the windowed region is computed for the entire image. Since this image can be thought as a gradient image (edgeness map) of the original image, this edgeness map is a good indicator of whether an area is in the interior of the region or near the boundary. For color image segmentation in RGB space, is generated by the edge information image for a pixel taking the maximum value among the analyzed edge measures for , , and components. By the proposed edge features and edgeness function, examples for extracting the edge information are displayed in Fig. 7. In this result, the Canny gradient operator is adopted for comparison with our method. The size of the analysis window was set 4 4 in our approach and 5 5 for the Canny operator . In the result, some edge regions where have a weak change of the intensity are detected well in comparison with those of the Canny gradient operator. Also, many edge-like responses can be effectively removed because of being based on the local homogeneity. These are good and valuable properties for the conventional marker-based region segmentation in computer vision and image processing. The characteristics of the generated edgeness map allow us to use a region growing method to segment the image. Region growing consists of determining marker points and expanding from marker locations. Region growing is followed by region merging and filtering to yield the final segmentation result. III. SPATIAL SEGMENTATION A. Marker Extraction A set of initial marker areas are selected as the basis for region growing. These marker areas are localized as minima of the local edge information map. In general, extracting good

1090


Fig. 7. Extracted edgeness information maps. First column: original images. Second column: edge images by Canny operator. Third column: edge images by the proposed edge features.

marker areas is a nontrivial problem and an important procedure for better segmentation. Many algorithms that use region growing select marker areas by heuristic thresholding. In our study, the connected regions that have zero values in the edgeness map are chosen as effective markers for region-growing and flooding. Thus, any thresholding method is not necessary unlike other works which utilize the spatial gradient operators. Remaining pixels are the uncertainty region for the growing procedure. B. Marker Growing For marker extraction, the edgeness map is divided into the two categories of marker areas and uncertainty regions. A marker area is assigned a label, then the uncertainty regions are grouped into the marker area. A modified approach is used for implementation of this procedure, as follows.

1) These uncertainty pixels are then grown with the following color similarity:

(9) are color vectors of a uncertainty pixel and where and are the th marker area, respectively. Also, , , and Euclidean distances of the color components between the uncertainty pixel and the marker area. If the above color distance between an uncertainty pixel and a labeled area is smaller than the predefined color distance, which is set at 15, this pixel is assigned to that label. 2) The average values of color bands , , and for each labeled region are updated. For an unclassified pixel with only one labeled region in its neighborhood, the pixel is


1091

Fig. 8. Segmentation results for visual comparison in synthetic images. First column: noisy images. Second column: watershed-based segmentation with the Canny operator. Third column: proposed algorithm.

assigned to the region. If the pixel has more than one labeled region in its neighborhood, it is assigned to the label of several regions if the shortest color distance to this pixel for each assigned region is less than the proper threshold value. The proper threshold value is determined as the minimum color distance between neighboring regions. 3) The remaining uncertainty pixels are assigned to the label regions with the shortest color distance to each pixel. This

process is repeated until there are no remaining uncertainty pixels. C. Region Merging and Filtering Region growing is followed by region merging and filtering to yield the final segmentation result. We use a minimum size of 50 pixels for a region. Small areas that are less than the defined

1092


Fig. 9. Segmentation results for comparison. First column: original. Second column: conventional watershed. Third column: probabilistic relaxation based on MRA. Fourth column: proposed algorithm.

size are merged into the neighborhood region with the shortest color distance to the small area. IV. EXPERIMENTAL RESULTS A variety of images were tested to validate the performance of the proposed object segmentation algorithm. These images

were selected frames of standard MPEG image sequences with the size of QCIF. Some synthetic images of size 128 128 were tested to show the robustness of the algorithm regarding noise. The conventional marker-based watershed algorithm and probabilistic relaxation method based on multiresolution analysis (MRA) were used to compare results with the proposed


1093

TABLE II PERFORMANCE COMPARISON WITH TWO EXISTING METHODS [3], [17]

algorithm [3], [17], [18]. In these methods, the gradient image is need to extract the marker areas. Most of them utilize Canny operator or morphological gradient. In this paper, Canny operis used for the gradient ator with the size of 5 5 and image of the watershed algorithm. In our study, we developed a segmentation algorithm based on new edge descriptor to give fewer segments and better image quality. This kind of algorithm can yield reliable quality performance with less manipulations for segments in the next processing stages. Therefore, we employ three measurements, that can consider the described facts, to evaluate the segmentation performance. These are the number of segments, goodness, and the signal-to-noise ratio (SNR) or peak SNR (PSNR). : Generally, there exists a tradeoff • Number of segments between the number of segments and the details of segmentation. Too many segments can give satisfactory results based upon visual observation while causing an oversegmentation problem in many marker-based segmentation algorithms. Too few segments can result in loss of information from many detailed regions of the original image. Therefore, it is necessary to determine the proper number of segments to produce reliable details in the result. • Goodness: The Goodness function was defined by Liu [19] as follows:

the original image and the segmented image in the th region. Equation (10) is composed of two terms. The first , penalizes segmentation that forms too many term, , is a local measure that regions. The second term, penalizes small regions or regions with a large color error since indicates whether or not a region is assigned an appropriate color. We scale down the variable F by a factor 1/1000. Smaller values of result in better segmentation. • PSNR: This measurement is widely used in image enhancement, image restoration and image compression [24]. Similar to the Goodness function , the segmentation result is compared with the original image. Usually, the mean square PSNR of the segmented image to the original image is written as:

(11)

is the segmented image and where original image. For applying the image segmentation method with ments, the average PSNR can be written as

is the seg-

(12) (10) The above equation is often rewritten as where is the number of regions in the segmented image, is the number of pixels in the th region, and is the sum of the Euclidean distance of the color vectors between

(13)

1094


As the value of the increases, the segmentation algorithm yields better results for any given number of image regions.

TABLE III CPU-TIME OF THE CONVENTIONAL CONVOLUTION-BASED METHOD (UNIT: ms)

A. Test for Synthetic Images In synthetic image testing the ground-truth is known. Synthetic images with two Gaussian distributions and , were generated. The strength of the given noise is and . To compare the performance of affected by terms segmentation, one parameter is devised as the noise strength. The noise strength of an image is defined as

TABLE IV CPU-TIME OF THE PROPOSED METHOD (UNIT: ms)

(14) From (14), the noise strength becomes larger as the denominator decreases or the numerator increases. This parameter has larger, the given noise becomes heavier. Fig. 8 shows results of the proposed segmentation algorithm and watershed-based algorithm with the Canny operator in noisy images. We used the size of 128 128 for an original image and 4 4 analysis window for extracting the property of the local histogram. In this figure, results were obtained after region filtering based on the same condition of color and size of the segmented region. In all cases, the proposed algorithm achieves better results than the watershed-based algorithm, which shows many false regions detected. The proposed algorithm based on new edge features effectively suppresses noisy terms although extensive noise exists. From results, it was shown that the proposed segmentation algorithm was more robust to a heavy noise than the employed method for comparison. The quantitative qualities of the segmentation results are illustrated in Table I. The proposed algorithm yields better results than the watershed-based method with the Canny operator due to an ability to suppress noisy clutters. From these tables, the segmentation results of the proposed algorithm are not only closer to the original image, but also contain fewer segments than watershed based segmentation with the Canny gradient operator. The more intensive noise exists, the performance of the proposed algorithm is better. B. Test for Natural Images For test of the spatial segmentation, several images from standard MPEG sequences were selected to validate the performance of the proposed algorithm. The images selected were from the Claire, Akiyo, Flower garden, Lena, and Foreman sequences. The size of the analysis window was set as 4 4 in the proposed algorithm. Figs. 9 shows the final results of segmentation. Visually, the proposed algorithm gives a reliable and detailed segmentation result with a smaller number of segments (or classes) for all tested images. Segmentation results are shown in Table II. The initial number of markers was chosen as the number of segments in Table II for comparison with the watershed method based on the Canny operator.

The proposed algorithm based on local histogram analysis can reduce the number of noisy segments significantly compared with the conventional watershed algorithm. Thus, the proposed algorithm is able to remove small noise clusters during marker extraction. All images have lower Goodness values than watershed algorithm images using fewer segments. We investigate the consumed time as the computational cost for extracting the edge information in the proposed algorithm. As described in advance, most algorithms for vision applications employed the spatial mask such as the Sobel, Prewitt, and Canny operators to examine the edge possibility of a pixel. Thus, we compare the part of the local histogram analysis in our algorithm with the other’s spatial mask for extraction of the edge information. Tables III and IV show results of the consumed CPU-time for generating the edge information map of QCIF images on an HW platform of a Pentium-4 PC with a 2.6GHz CPU and 512 MBytes of RAM. From these tables, it can be known that the proposed local histogram analysis-based approach is faster than the convolution-based method for extracting the edge information map. In detail, the consumed time of 4 4 window in the proposed approach is faster than 5 5 mask of the convolution-based method by a factor of approximately 3 times. In our work, we emphasized the image or region segmentation as the preprocessing part for object-based video coding. To define an object (or VO), the region partitioning must be executed in advance. Based on the segmented regions, shapes and motions of the semantic objects may be computed for removing data redundancy. In this process, there exist advantages because of fewer segments based on the proposed algorithm. For example, motion estimation is faster than others due to less number of the segmented regions with credible image quality. V. CONCLUSION We have developed a novel video segmentation algorithm based on the newly defined concept of edge information. The


introduced edge information is determined from analysis of the local histogram. A discriminant function based on local histogram analysis is introduced to construct an edge information map that describes the existence of two or more distinct brightness regions in a window image. According to the characteristics of the gradient image, a modified region-growing method is applied for region partitioning. Unlike other methods that use spatial gradient operators, the proposed edge feature can extract the edge region where a change of intensity or color is weak. This can give better results in the stage of region growing. Experimental results are provided to demonstrate the feasibility of the proposed algorithm. ACKNOWLEDGMENT The authors would like to thank the reviewers for their valuable comments and Dr. Cassell for helping to check the English. REFERENCES [1] R. Koenen, “MPEG-4 multimedia for our time,” IEEE Spectrum, vol. 36, no. 2, pp. 26–33, Feb. 1999. [2] MPEG Requirements Group, “MPEG-7 context and objectives,” presented at the MPEG Atlantic City Mgt., 1998, Doc. ISO/MPEC 2460. [3] L. Salgado, N. Garcia, J. M. Menéndez, and E. Rendón, “Efficient image segmentation for region-based motion estimation and compensation,” IEEE Trans. Circuit Syst. Video Technol., vol. 10, no. 7, pp. 1029–1039, Oct. 2000. [4] T. Meier and K. N. Ngan, “Automatic segmentation of moving objects for video object plane generation,” IEEE Trans. Circuit Syst. Video Technol., vol. 8, no. 5, pp. 525–538, Sep. 1998. [5] ——, “Video segmentation for content-based coding,” IEEE Trans. Circuit Syst. Video Technol., vol. 9, no. 8, pp. 1190–1203, Dec. 1999. [6] C. G. Kim and J. N. Hwang, “A fast and robust moving object segmentation in video sequences,” in Proc. Int. Conf. Image Process., 1999, vol. 2, pp. 131–134. [7] E. A. Edirisinghe and J. Jiang, “A contour analysis based technique to extract objects for MPEG-4,” in Proc. Int. Conf. Image Process., 1999, vol. 1, pp. 369–374. [8] I. Kompatsiaris and M. G. Strintzis, “Spatiotemporal segmentation and tracking of objects for visualization of videoconference image sequences,” IEEE Trans. Circuit Syst. Video Technol., vol. 10, no. 8, pp. 1388–1402, Dec. 2000. [9] L. Shi, Z. Zhang, and P. An, “Automatic segmentation of video object plane based in object tracking and matching,” in Proc. IEEE Int. Symp. Intell. Multimedia, Video Speech Process., 2001, pp. 510–513. [10] P. Salembier, A. Oliveras, and L. Garrido, “Antiextensive connected operators for image and sequence processing,” IEEE Trans. Image Process., vol. 7, no. 4, pp. 555–570, Apr. 1998. [11] S. Herrmann, H. Mooshofer, H. Dietrich, and W. Stechele, “A video segmentation algorithm for hierarchical object representations and its implementation,” IEEE Trans. Circuit Syst. Video Technol., vol. 9, no. 8, pp. 1204–1215, Dec. 1999. [12] C. Gu and M. C. Lee, “Semantic video object segmentation and tracking using mathematical morphology and perspective motion model,” in Proc. Int. Conf. Image Process., 1997, vol. 2, pp. 514–517. [13] J. P. Gambotto, “A region-based spatio-temporal segmentation algorithm,” in Proc. Int. Conf. Pattern Recognit., 1992, vol. 3, no. 3, pp. 189–192.

1095

[14] C. Gu and M. C. Lee, “Semantic video object tracking using regionbased classification,” in Proc. Int. Conf. Image Process., 1998, vol. 3, pp. 643–647. [15] J. Y. Zhou, E. P. Ong, and C. C. Ko, “Video object segmentation and tracking for content-based video coding,” in Proc. Int. Conf. Multimedia Expo., 2000, vol. 3, pp. 1555–1558. [16] Y. Deng and B. S. Manjunath, “Unsupervised segmentation of colortexture regions in images and video,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 23, pp. 800–810, 2001. [17] H. Gao, W.-H. Siu, and C.-H. Hou, “Improved techniques for automatic image segmentation,” IEEE Trans. Circuit Syst. Video Technol., vol. 11, no. 12, pp. 1273–1280, Dec. 2001. [18] J.-B. Kim and H.-J. Kim, “Multiresolution-based watersheds for efficient image segmentation,” Pattern Recognit. Lett., vol. 24, pp. 473–488, 2003. [19] J. Liu and Y.-H. Yang, “Multiresolution color image segmentation,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 16, no. 7, pp. 689–700, Jul. 1998. [20] K.-M. Chen and S.-Y. Chen, “Color texture segmentation using feature distributions,” Pattern Recognit. Lett., vol. 23, pp. 755–771, 2002. [21] T. Pun, “Entropic thresholding, a new approach,” Computer Graph. Image Process., vol. 16, pp. 210–236, 1981. [22] B.-G. Kim and P.-S. Mah, “Non-contrast based edge descriptor for image segmentation,” in Proc. 17th IEEE Int. Conf. Pattern Recognit., 2004, vol. 1, pp. 572–575. [23] Graham and H. Clarence, Vision and Visual Perception. New York: Wiley, 1965. [24] A. K. Jain, Fundamentals of Digital Image Processing. Englewood Cliffs, NJ: Prentice Hall, 1997. Byung-Gyu Kim (M’04) received the B.S. degree from Pusan National University, Pusan, Korea, in 1996 and the M.S. and Ph.D. degrees from the Department of Electrical Engineering and Computer Science, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea, in 1998 and 2004, respectively. In March 2004, he joined the Real-Time Multimedia Research Team, Electronics and Telecommunications Research Institute (ETRI), Daejeon, Korea, where he is currently a Senior Researcher. His research interests image segmentation for the content-based image coding, realtime multimedia communication and intelligent information system for image signal processing. Dr. Kim is a member of the IEEE Computer and Communication Societies, IEICE, and the Korea Multimedia Society (KMMS).

Dong-Jo Park (M’86) received the B.S. degree in electronics engineering from Seoul National University, Seoul, Korea, in 1976, and the M.S. and Ph.D. degrees in electrical engineering from the University of California, Los Angeles, in 1981 and 1984, respectively. From 1984 to 1985, he was a Member of the Technical Staff, Electronics and Telecommunications Research Institute (ETRI), Daejeon, Korea. In 1985, he joined the Department of Electrical Engineering and Computer Science, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea, where he is currently a Professor. His research interests in the area of communication and image signal processing. He is a member SICE, KIEE, KICS, and KISS.