RIASSUNTO
The geospatial land recognition is often cast as a local-region based classification problem. We show in this work, that prior knowledge, in terms of global semantic relationships among detected regions, allows us to leverage semantics and visual features to enhance land use classification in aerial imagery. To this end, we first estimate the top-k labels for each region using an ensemble of CNNs called Hydra. Twelve different models based on two state-of-the-art CNN architectures, ResNet and DenseNet, compose this ensemble. Then, we use Grenander’s canonical pattern theory formalism coupled with the common-sense knowledge base, ConceptNet, to impose context constraints on the labels obtained by deep learning algorithms. These constraints are captured in a multi-graph representation involving generators and bonds with a flexible topology, unlike an MRF or Bayesian networks, which have fixed structures. Minimizing the energy of this graph representation results in a graphical representation of the semantics in the given image. We show our results on the recent fMoW challenge dataset. It consists of 1,047,691 images with 62 different classes of land use, plus a false detection category. The biggest improvement in performance with the use of semantics was for false detections. Other categories with significantly improved performance were: zoo, nuclear power plant, park, police station, and space facility. For the subset of fMow images with multiple bounding boxes the accuracy is 72.79% without semantics and 74.06% with semantics. Overall, without semantic context, the classification performance was 77.04%. With semantics, it reached 77.98%. Considering that less than 20% of the dataset contained more than one ROI for context, this is a significant improvement that shows the promise of the proposed approach.