Color Image Segmentation based on Automatic Derivation of Local Thresholds

In this paper a new method for color image segmentation is presented. The proposed algorithm divides the image into homogeneous regions by derivation of local thresholds via local information. The algorithm contains two main steps. First, the watershed algorithm is applied on the image gradient magnitude. Its results are used as an initial segmentation for the next step, which is region merging process. During that process regions are merged and local thresholds are derived one-by-one at different times by analyzing local characteristics of the regions. Every threshold refers to specific region and defines it as ífinal regioní (non-mergeable). Thus, regions are handled separately; some regions grow while others were already defined as ífinal regionsí. The significant use of local information improves the quality of the segmentation result. Experimental results have demonstrated the efficiency of the proposed method. The algorithm is found to be reliable and robust for different images. 1 Introduction Image segmentation is a process of dividing an image into (non-overlapping) homogeneous regions with respect to a chosen property. It is an important task for variety of application such as object extraction, image compression, objects tracking and objects recognition. The existing segmentation methods 1, 2 can be roughly divided into four different groups: Histogram-based methods, boundaryñbased methods, region-based methods and hybrid-based methods. Most of the Histogram-Based methods deal with gray level images, which are represented by one-dimensional histogram. The range of intensities is assumed to be constant and the histogram is considered as being a probability density function of a Gaussian. Then, the segmentation problem is reformulated as parameter estimation followed by pixel classification. For 3-D color space (RGB, HSV, etcí) some techniques were developed (for example,3). The idea of Boundary-based methods is to search for pixels that lie on a region boundary. These pixels are called edges 4. An edge is characterized by a significant local change in image intensities. Edges are detected by looking at neighboring pixels. The basic assumption is that the change in pixels values between neighboring pixels inside a region is not as significant as the change in pixels values on the regions