Image segmentation is a widely-researched topic with many algorithms available. Our goal is to segment an image, in an unsupervised way, into several coherent parts with the help of superpixels. To achieve that, we propose an iterative segmentation algorithm. The algorithm models the image by a Markov random field, whose nodes are the superpixels, and each node has both color and texture features. The superpixels are assigned labels according to their features with the help of support vector machines and the aforementioned MRF and the number of segments is iteratively reduced. The result is a segmentation of an image into several regions with requiring any user input. The segmentation algorithm was tested on a standard evaluation database, and performs on par with state-of-the-art segmentation algorithms in F-measures. In terms of oversegmentation, our approach significantly outperforms the state of the art by greatly reducing the oversegmentation of the object of interest.
|
