Supervised image segmentation using learning and merging

The segmentation problem can be viewed as a learning and merging problem based on superpixels (image segments), which can incorporate a group of cues to guide the segmentation. So the proposed multi-label segmentation algorithm mainly consists of two stages: the learning stage and the merging stage. In the learning stage, Gaussian Mixture Models (GMMs) firstly learn color models for different components of objects. Based on the likelihood, we execute the alpha-expansion algorithm only once in order to alleviate the shrinking bias. The initial labels help determine whether a superpixel is too noisy, and the contour responses between superpixels can distinguish spurious boundaries. Those superpixels containing too much noisy pixels and spurious boundaries will be unlabeled. In the merging stage, unlabeled superpixels may have similar color information while differing in texture information. Therefore, they can be correctly classified by a novel region merging algorithm based on maximal similarity. In this way the advantages of features in different levels are enhanced by uniting them in different stages. Finally, the proposed method is evaluated on the Berkeley segmentation benchmark, the Graz benchmark and the Grabcut benchmark. Experimental results show that our method obtains the highest accuracy on the Graz benchmark, and the performance on other benchmarks can also be comparable or better than current leading algorithms.