DOMINANT COLOR EXTRACTION

Gozde Bozdagi, Bülent Tavli
 

Current research on image indexing focuses on the choice of the suitable features for image representation. Among these featues, texture, shape, color, location are the most frequently used ones. The comparison of these features used in the literature is based on the complexity of search and the feature vectors. In this work, we propose a simple method for color based indexing that provides efficient indexing and fast search. Since color is a vital part of our everyday experience, it is useful for powerful visual processing in the world around us. Other than exact color distribution within a digitized image, people are usually interested in the most dominant color components. Due to this fact, we try to extract the dominant colors together with their locations and the area from the images to be used in indexing and retrieval. The proposed method is based on coarse segmentation of color images using a scale-space filter. Application of scale space filters for determining salient features is first proposed by Witkin. Later on, Lim and Lee combined it with a Fuzzy clustering method to have a more accurate color segmentation. In this work we use a similar idea however reduce the complexity by modifying the clustering method.

The proposed algorithm is composed of the floowing steps:

• Obtain the histograms of R, G, B components of the image seperately.
• Obtain the scale-space representation of each histogram. In order to do this, convolve the histograms by a Gaussian filter given as


h(x, s) = exp[- x ² / (2 s ² ) ] / ( s (2 p) ^½ ),




where s changes at each scale and take the second order derivative. A sample image is given in Fig. 1, its R-histogram and scale-space representation obtained by this method is shown in Fig. 2.
• Determine the peak locations from the scale-space representation. The number of locations depends on the number of dominant regions that is wanted to be extracted from the image. The location of peaks are shown in Fig. 2 based on the fact that the 3 dominant regions are going to be extracted.
• Trace the peak locations to the lower scale in order to find thresholds for each region.
• Repeat same process (3-4) for each color component.
• Take the intersection of the extracted regions and find the overall thresholds.
• Segment the image based on these thresholds.
• Filter the image to get rid of small regions. The three dominant colors extracted by the proposed method is shown in Fig 3.

Fig.1

Fig.2

Fig.3