Iterative Semi-Supervised Learning Approach for Color Image Segmentation

Image segmentation is an important step in many image processing techniques. In this paper, a new semi-supervised approach for color image segmentation is proposed. This method takes advantage of a limited human assistant. After an unsupervised segmentation stage, classes of some regions are questioned from the user. These user hints are used as an initial sample data and will be iteratively expanded based on the existing relevancy between adjacent pixels. This relevancy is measured by probabilities calculated by a classifier which has learned the existing samples prior to that iteration. The learner is a multinomial logistic regression (MLR) classifier. The extended seed is used for training of a support vector machine (SVM) classifier in order to perform the final segmentation. The result of this segmentation fulfills the intention of the user and extracts the targeted classes. Experimental results show that our proposed method makes a noticeable improvement in the accuracy with respect to comparable algorithms.