Image Segmentation Based on Semantic Knowledge and Hierarchical Conditional Random Fields

Semantic segmentation is a fundamental and challenging task for semantic mapping. Most of the existing approaches focus on taking advantage of deep learning and conditional random fields (CRFs) based techniques to acquire pixel-level labeling. One major issue among these methods is the limited capacity of deep learning techniques on utilizing the obvious relationships among different objects which are specified as semantic knowledge. For CRFs, their basic low-order forms cannot bring substantial enhancement for labeling performance. To this end, we propose a novel approach that employs semantic knowledge to intensify the image segmentation capability. The semantic constraints are established by constructing an ontology-based knowledge network. In particular, hierarchical conditional random fields fused with semantic knowledge are used to infer and optimize the final segmentation. Experimental comparison with the state-of-the-art semantic segmentation methods has been carried out. Results reveal that our method improves the performance in terms of pixel and object-level.