Hierarchical Image Segmentation Using Semantic Edge Constraint

Segmentation is one of the fundamental problems in the field of computer vision. Although there is abundant research work published in this topic, it is still very challenging, because detecting the continuous edge contour itself is a laborious task. In this paper, we explore an approach on segmentation by using semantic edge constraint. Firstly, the local features are extracted and then trained by using the discriminative classifier within the hierarchical regions. Next, the spatially consistency constraint is introduced to strength the connection between the neighboring regions, and consequently the scores for each group of pixel is calculated. Secondly, the semantic edge constraint, an oriented gradient signal which is computed on semantic channel, is combined with the normalized edge signal to detect the effective edges on the objects contour. Finally, a greedily merging method is adopted to obtain final hierarchical segmentation regions. Experiments achieves the state-of-the-art performance on the challenging MSRC 21 datasets.