CUE-LEVEL CHANGE DETECTION

This proposal describes an innovative coarse-to-fine approach to the detection of militarily relevant changes in electrooptical remotely sensed imagery. Our approach is based on the use of a generalized cross-image modeling technique to act as a focus-of-attention mechanism, and graph monomorphism finding algorithms for object-level reasoning. The focus-of-attention algorithms are robust with respect to misregistration effects and differential illumination and determine the likely areas of possible change. Knowledge of the approximate location and size of change regions reduces tremendously the computational requirements of object-level analysis and provides good initial estimates for processing parameters. Detailed analysis and matching of segmented regions is used to determine locations of actual physical change. Prioritization strategies are developed to permit an image analyst to rank the exploitation mission relevance of each change cue.