Scene Classification Algorithm Based on Covariance Descriptor

Scene classification has been a hot topic in the field of computer vision. Under the premise of image segmentation, this paper proposes a novel scene classification algorithm, combining pixel location, color characteristics, Gabor features, and local binary features (LBP) to form a covariance descriptor, and then converting it to the Sigma-point characteristics into a European Space, to complete the scene description and SVM training. To compare performance with some of the classic classification algorithms, we simulate the algorithm on standard Image SUN Database, and besides we construct data set with noise to validate their tolerance in dealing with noise and robustness. The results show that the proposed algorithm not only has a strong advantage on computation time, feature dimension and classification performance, but also has good fault tolerance and robustness.