A Logical Framework for Depiction and Image Interpretation

We propose a logtcal framework for deptctton and mterpretatton that formahzes tmage domain knowledge, scene domain knowledge and the deptctton mapping between the tmage and scene domains Thts framework requires three sets of axtoms tmage axioms, scene axtoms and deptctton axtoms An mterpretatton of an tmage ts defined to be a logtcal model of these axtoms The approach Ls dlustrated by a case study, a reconstructton tn first-order logw of a stmphfied map-understanding program, Mapsee. The reconstructton starts with a descrtptton of the map and a spectficatton of general knowledge of maps, geographtc oblects and thetr deptctton relattonshtps For the stmple map world we show how the task level spectficatton may be refined to a provably correct tmplementatton by applying model-preserving transformanons to the mlttal logtcal representatton to produce a set of propostttonal formulas The tmplementatton may use known constraint satt~factton techmques to find the ~et of models of the~e propostttonal formulas In addttton, we sketch prehmmary logtcal treatments for zmage querws, contingent scene knowledge, ambtgutty m tmage descrlptton, occluston, complex oblects, preferred mterpretattons and tmage synthests Thts approach provtdes a formal framework for analyzing and going beyond extsttng systems such as Mapsee, and for understanding the use of constraint sattsfactton techmques It can be used as a foundatton for the spectficatton, destgn and implementation of vtslon and graphtcs systems that are correct with rewect to the task and algortthm levels