A CNN-BASED SEGMENTATION MODEL FOR SEGMENTING FOREGROUND BY A PROBABILITY MAP

This paper proposes a CNN-based segmentation model to segment foreground from an image and a prior probability map. Our model is constructed based on the FCN model that we simply replace the original RGB-based three channel input layer by a four channel, i.e., RGB and prior probability map. We then train the model by constructing various image, prior probability maps and the groundtruths from the PASCAL VOC dataset, and finally obtain a CNN-based foreground segmentation model that is suitable for general images. Our proposed method is motivated by the observation that the classical graphcut algorithm using GMM for modeling the priors can not capture the semantic segmentation from the prior probability, and thus leads to low segmentation performance. Furthermore, the efficient FCN segmentation model is for specific objects rather than general objects. We therefore improve the graph-cut like foreground segmentation by extending FCN segmentation model. We verify the proposed model by various prior probability maps such as artifical maps, saliency maps, and discriminative maps. The ICoseg dataset that is different from the PASCAL Voc dataset is used for the verification. Experimental results demonstrates the fact that our method obviously outperforms the graphcut algorithms and FCN models.