An Iterative Image Segmentation Algorithm Utilizing Spatial Information

otherwise it is segmented into another class. Multi-class thresholding is defined in a similar manner. There exist various algorithms for the determination of the thresholding value [2-4]. ABSTRACT: An iterative image segmentation algorithm that segments an image on a pixel-by-pixel basis is described. The observation information to be utilized is the joint gray level values of the pixel to be segmented and those of its neighborhood pixels. The iterative process is initialized by thresholding the image with Otsu' s method. Each pixel is segmented into a class when the a posteriori probability, conditioned on the observation information, that it belongs to this class is maximum. The newly segmented image is employed to re-estimate the a posteriori probabilities and the segmentation process is repeated until there is no further pixel classification change in a particular run. Among those segmented images generated in the iterative process, the best segmented inzage is chosen according to a maximum entropy criterion. Simulation studies demonstrate that the proposed algorithm can achieve very significant improvement in segmentation performance as compared to the more popular thresholding approach. Furthermore, the performance is neither sensitive to the initial thres.hold value nor the form of the probability densif)' function of the image. Segmentation of practical images also demonstrates that the proposed algorithm is capable of good segmentation results for real-life images. An important advantage of thresholding is that the threshold value can be derived in high speed if Qnly simple statistical properties, such as pixels' probability density function (pdt), of the image are considered [5]. But since an image pdf does not contain sufficient amount of information to characterize the image faithfully, a pdf-based thresholding algorithm may not always arrive at a good threshold value. A possible improvement is to make use of more information, such as spatial information, to determine the threshold value. In recent years, several new thresholding methods that utilize spatial information in determining the threshold value have been reported [6-8]. Another limitation of thresholding lies in its fundamental characteristics and hence is more difficult to overcome. After a threshold value has been determined, all pixels having the same gray level value will invariably be segmented into the same class. The segmentation result is limited by the degree of overlap of the pdf's of the constituent sub-images irrespective of how the threshold value is derived. When the degree of overlap is small, good thresholding results are still possible; if the degree of overlap is substantial, it is impossible to obtain good thresholding results. INTRODUCTION 1.