A weighted dominant color descriptor for content-based image retrieval Ahmed Talib a,b,⇑ , Massudi Mahmuddin a , Husniza Husni a , Loay E. George c a Computer Science Dept., School of Computing, University Utara Malaysia, 06010 Sintok, Kedah, Malaysia b IT Dept., Technical College of Management, Foundation of Technical Education, 10047 Bab Al-Muadham, Baghdad, Iraq c Computer Science Dept., College of Science, Baghdad University, 10071 Al-Jadriya, Baghdad, Iraq article info Article history: Received 20 August 2012 Accepted 2 January 2013 Available online 18 January 2013 Keywords: Dominant color descriptor MPEG-7 Object- and content-based image retrieval Semantic feature Similarity measures Salient object detection Background dominance problem Linear Block Algorithm abstract Color has been extensively used in the process of image retrieval. The dominant color descriptor (DCD) that was proposed by MPEG-7 is a famous case in point. It is based on compactly describing the promi- nent colors of an image or a region. However, this technique suffers from some shortcomings; especially with respect to object-based image retrieval. In this paper, a new semantic feature extracted from dom- inant colors (weight for each DC) is proposed. The newly proposed technique helps reduce the effect of image background on image matching decision where an object’s colors receive much more focus. In addition, a modification to DC-based similarity measure is also proposed. Experimental results demon- strate that the proposed descriptor with the similarity measure modification performs better than the existing descriptor in content-based image retrieval application. The proposed descriptor considers as step forward to the object-based image retrieval. Crown Copyright Ó 2013 Published by Elsevier Inc. All rights reserved. 1. Introduction Image retrieval become one of the most famous research direc- tions nowadays because it uses to search an image in archive, do- main-specific, personal and web image databases. For retrieving color images from multimedia database, low level features and especially the color feature, have been widely used in this regard. This is because color represents the most distinguishable feature compared with other visual features, such as texture and shape. From perspective of feature extraction, color-based image descrip- tors can be divided into two categories: (i) Global descriptors that consider the whole image to obtain their features, there is no par- titioning or pre-processing stage during feature extraction process. The resulted descriptors from this approach are simple and fast but it lacks to spatial color information and high discriminating power. The most famous example about this representation is global color histogram [1]. (ii) Local descriptors that obtain their features from local regions or partitions of image. This can be done by dividing the image into either fixed size or different size regions. The former type is called fixed partitioning-based approaches and they have more spatial information about colors in the image; an example of this approach is cell color histogram [2]. The latter type is called segmentation-based approaches where the regions of image can be extracted by either segmentation or clustering methods. These descriptors usually have better accuracy than others but introduce more complexity of feature extraction process; examples of this approach are color-based clustering [3] and dominant colors [4,5]. In addition to the global and local approaches to represent the im- age, there are other interesting methods that get more attention re- cently which are local invariant feature-based approaches. These approaches introduced features that are invariant to different image transformation such as translation, scaling, rotation and affine trans- formation. Salient edges and regions (saliency map) detection are part of these approaches and they are widely used in many computer vision applications including object recognition [6] and image retrieval [7]. Combination of different representation methods (at the feature or rank level) leads to improvement of image retrieval accuracy [8]. Therefore, in this work, combination of global and local features (at feature level) is proposed to enhance color-based image retrieval. In this respect, MPEG-7 Committee proposed many color, tex- ture and shape descriptors to be used in image and video retrieval [9,10]. Authors in [11–13] maintain that human visual system first helps identify prominent colors in the image and second it pro- cesses any other details. The whole process resembles the way hu- mans recognize image from its dominant colors without paying any attention to their distribution. MPEG-7’s DCD (MP7DCD) pro- vides compact and effective representations for colors in an image or region of interest [9]. Recently, compactness property of domi- nant colors representation becomes more attractive for many researchers to reduce size of color descriptors from several hun- dred bins (histogram-based methods) into few colors (8 colors in MP7DCD) such as the works that have been done in [12,14–16]. 1047-3203/$ - see front matter Crown Copyright Ó 2013 Published by Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.jvcir.2013.01.007 ⇑ Corresponding author at: Computer Science Dept., School of Computing, University Utara Malaysia, 06010 Sintok, Kedah, Malaysia. E-mail address: s91707@student.uum.edu.my (A. Talib). J. Vis. Commun. Image R. 24 (2013) 345–360 Contents lists available at SciVerse ScienceDirect J. Vis. Commun. Image R. journal homepage: www.elsevier.com/locate/jvci