CDSEG: Community Detection for Extracting Dominant Segments in Color Images S. Hamid Amiri, A. A. Abin, and Mansour Jamzad Department of Computer Engineering, Sharif University of Technology, Tehran, Iran s_amiri@ce.sharif.edu, abin@ce.sharif.edu, jamzad@sharif.edu Abstract— Segmentation plays an important role in the machine vision field. Extraction of dominant segments with large number of pixels is essential for some applications such as object detection. In this paper, a new approach is proposed for color image segmentation which uses ideas behind the social science and complex networks to find dominant segments. At first, we extract the color and texture information for each pixel of input image. A network that consists of some nodes and edges is constructed based on the extracted information. The idea of community detection in social networks is used to partition a color image into disjoint segments. Community detection means partitioning vertices of a network into different non-overlapped groups (communities) such that the density of intra–group edges is much higher than the density of intergroup edges. There is a very close relation between communities in the social network and segments in an image. Our results show that community detection approaches will improve the segmentation output compare to other methods available in the literature. Keywords- Image segmentation; Social networks; Community detection; Modularity measure. I. INTRODUCTION Image segmentation is an essential and critical component of any image analysis or pattern recognition system. It is considered as a preprocess for a number of applications in the image processing such as image and video retrieval, extraction of region of interest (ROI) in an image or scene, and medical image processing. Furthermore, it is one of the most difficult tasks in the image processing because its output has a great influence on the performance of many image analysis systems. A precise definition of segmentation is given in [4,10] as follow: “Image segmentation is a process of dividing an image into different regions such that each region is homogeneous, but not the union of any two adjacent regions” [10]. Nevertheless, according to [4], the image segmentation problem is basically one of psychophysical perception, and therefore not susceptible to a purely analytical solution. According to [16], Image segmentation algorithms are classified into three categories. First, feature-space-based clustering approaches [2,3,5,6] that utilizes image features such as color and texture to group the feature samples into compact, but well-separated clusters. These approaches ignore the spatial structure and edge information. Also, the pixels from disconnected regions of the image may be grouped together if their feature spaces overlap. The second, spatial segmentation or region-based methods [11] perform the segmentation based on some homogeneity criteria. Since these methods may produce a large number of homogenous regions, some merging algorithm must be applied to the regions. The third category is graph-based approaches [12,15,16] that combine the image features with spatial information and use some factors such as similarity and continuation to extract the segments. The common idea among these approaches is the formation of a weighted graph based on the pixels or regions. The constructed graph is partitioned into multiple components to minimize some cost function. In this paper, we propose a new approach called CDSEG to find dominant segments of color images. Dominant segments mean those regions that correspond to the main object in the image. Thus, we are not interested in the regions with small number of pixels. CDSEG uses the community detection algorithms whose aim is to analyze a social network. These algorithms find the groups in the network such that individuals could be divided into them. There is a very close relation between community detection and image segmentation because the ultimate goal of image segmentation (community detection) is to partition pixels of the input image (nodes of the network) into disjoint segments (communities). Since each network is represented by a graph, the main step of CDSEG is the graph construction that maps input image onto a network. CDSEG consists of the following steps: Preprocessing the input image to merge pixels with high similarity, Constructing a weighted network using color and texture information, Extracting the communities by applying the Newman-Fast algorithm [8] to the network, Postprocessing the detected communities to form the dominant segments. The quality of proposed method is sensitive to the network modeling step. The remainder of paper is organized as follows. Section 2 presents a review on the related concepts of community detection. Details of CDSEG steps are presented in section 3. Experimental results are discussed in Section 4 and concluding remarks are given in Section 5. II. COMMUNITY DETECTION Many systems studied in different fields of science could be represented by a network. Examples contain Internet, neural networks, communication and social networks etc. To represent the network, one graph is constructed such that each entity of the network is mapped onto one node (vertex) and some links are used to connect pair of nodes. Community structure, which is a property of complex networks, is described as the gathering of vertices into groups 7th International Symposium on Image and Signal Processing and Analysis (ISPA 2011) September 4-6, 2011, Dubrovnik, Croatia Image Processing Image Segmentation 177