Narrative image is one type of visual literature resources that intends to tell stories. In face of its obscure and abstract nature,existing narrative-relevant semantic models appear inadequate for its semantic annotation and description. To address the issue,this paper designs an ontology model for deep semantic annotation applicable for narrative images.
The model resorts to the design thought as in modular ontology,integrates with Open Annotation Collaboration(OAC)framework,and takes plot,entity,dynamics and context as its core. The model addresses issues such as the organization and representation of spatio-temporal information in narrative images,and the mapping between narrative model and annotation information.
Plot lies at the core of the framework. Each plot takes place in a specific context and involves some entities,while different plots are semantically related in terms of sequence or overlap. In a specific plot annotation area within an image,a dynamics is normally used to present dynamic features of the plot. The descriptive framework,as a whole,is an abstract semantic annotation model and accomplishes the semantic description and annotation by mapping annotated objects onto the abstract model. Plot constitutes the core of the annotation framework and corresponds to the Body of OAC in image annotation.
In detail,temporal semantic relations between plots contain nextplot,lastplot,overlaps,meets,starts,during,finishes,equal. In addition to the above relations,other obscure and abstract spatio-temporal semantic relations can also be identified in plots. This paper places TemporalSpatialEntity under the category of Entity. To meet the need for annotation of spatio-temporal information in narrative images,this paper divides TemporalSpatialEntity into Place and TemporalThing,which are respectively further divided into two types-Concrete(ConcretePlace,ConcreteTemporalThing)and Abstract(AbstractPlace,AbstractTemporalThing). Dynamics is another indispensable element in the model. As a core sub-category in the process of annotation,Action has two object properties,namely hasActor and hasRecipient；the former being the agent of the action while the latter the patient. At the same time,Dynamics can also have a type(DyanmicsType). Context is a special item in the plot annotation of narrative images,and it provides the annotation model with additional extension ability. Instead of being directly accessible from image,information involved in context is hidden behind the image,yet helpful in understanding the image.
To represent the complex relations of overlap between different annotated areas within an image,this paper puts forward OAC extended vocabularies,including no-overlap,complete-overlap,inclusiveness,and partial-overlap. In the end,this paper performs a semantic annotation experiment on Dunhuang narrative murals,using an image semantic annotation tool embedded with the model,and the annotation data is converted and stored in the form of RDF triples. The experiment verifies the model's reliability and applicability. 8 figs. 1 tab. 22 refs.