Semantic image segmentation is a process of dividing a picture into a finite amount of non-intersecting and substantive regions. A contextual framework referred to as Contextual Hierarchical Model (CHM) was adapted to learn the suitable data throughout hierarchical framework semantic image segmentation. CHM consists of multiple hierarchy levels and effectively model key information within a fiction occurs. Due to lack of global constraints, the CHM has short comings in terms of class average accuracy. So, in the research work; a classical Conditional Random Field (CRF) is introduced to solve the above problem. In CRF, the global constraints are defined through an energy function on a discrete random field. However, this model does not incorporate the allocation of multiple classes to the unique region of CRF graph model under the global node. So, a hierarchical CRF (HCRF) model is proposed to overcome the issues in CRF and to define the global constraints. In HCRF, the global constraints are defined by describing the energy function on unary, pairwise and higher order potentials. In higher order HCRF (HHCRF), an economical technique for a HCRF is intended by using the structures that depends upon long continuous label sequences, as long because the range of fundamentally different label sequences utilized in the structures is tiny. This leads conditional random fields to attain the efficient learning algorithm. Hence the proposed CRF, HCRF and HHCRF are trained at multiple resolutions and the output of these models is used to train Logistic Disjunctive Normal Networks (LDNN) for semantic image segmentation.
Volume 11 | 01-Special Issue
Pages: 48-57