Contextual Object Recognition with Conditional Random Fields

For contextually consistent target recognition in satellite imagery, a contextual conditional random field (CRF) model is proposed. First of all, context invariance for the target is determined by sparse auto-encoders. The area represented by the most repetitive invariance is used as central node in CRF model. Marking such an area directly as the target or applying a rule-based methodology concludes in false alarms or missing results. Therefore, a star-shaped CRF, which models contextual relationships, is used. Other nodes of the CRF are chosen as land-use land-cover classes in the surroundings of the candidate target area. These classes are obtained by merging segments with the same label after feeding best known discriminative features to support vector machines. The same features are extracted from the merged class areas to be used as node features in CRF. Edge features in CRF are essential for representing contextual relations and they are chosen as class co-occurrence frequencies. For each target candidate, a dynamic CRF model is generated and in those models, each node can have two states (true or false). The proposed method is robust in terms of being threshold-free and selecting contextual invariance via sparse auto-encoders. Performance of the method is competitive to rule-based methods and segmentation-based classification methods.