Remote Sensing Image Feature Expression And Segmentation Algorithms Based On Riemannian Manifold

Most image segmentation algorithms are based on the properties of images in the spectral measurement space to partition the image into classes.Unfortunately,spectral measurement space only contains spectral information of the detected image.The expression is considered incomplete and inefficient.The segmentation ability of image segmentaion algorithms,which are based on the spectral measurement space,are restricted,leading the algorithm being sensitive to noise and outliers.Based on information geometry theory,an image feature space on Riemannian manifold is constructed.The new constructed feature space contains both the spectral information and the interactions between neighbor pixels.That means the expression of properties in Riemannian manifold space is more compeletely than that in the spectral measurement space.Then data and parameter submanifolds are defined to characterize the image features and segmentation results based on the Riemannian manifold space.Since image information is fully utilized,the recognition ability and robustness of noise and outliers are improved.To obtain parameters in the proposed model,a manifold projection operation is designed to project points on the data submanifold to active points on the parameter submanifold.Then the active points are updated according to the projection result.When the algorithm converges,the active points on the parameter submanifold stand for the classes in the image segmentation result and the projection represent for the segmentation.To further improve the seperability of image data in Riemannian manifold space,kernel function is employed to map points on low dimensional manifold to high dimensional space.Thus,the projection criterion defined by geodesic kernel function is introduced to improve the accuracy of segmentation results.In addition,the Riemannian manifold algorithm based on Riemannian manifold space and the clustering algorithm based on spectral measurement space are demonstrated to be equal in processing.To qualitatively and quantitatively analyze the segmentation accuracy of the proposed algorithms,two synthetic images with simple and complex scenes are employed.User accuracies,product accuracies and overall accuracies of all the classes are calculated according to the confusion matrix.At the same time,8 small scale and a large scale remote sensing image are used to validate the effectiveness of the proposed algorithms.Qualitative and quatitative analysis of the experimental results implies that the Riemannian manifold algorithms are more accurate than the clustering algorithms in identifying the objects and resisting noise and outliers.In addition,it is demonstrated that the proposed Riemannian manifold algorithms are equivalent to the clustering algorithms in processing.While comparative experiments on the Riemannian manifold image sementaiton algorithms and the equvalent clustering algorithms validate the effectiveness of the proposed image segmetation model and the advantages of expression ability of the proposed Reimannian manifold space.Besides,the equivalency indicates that the proposed Riemannian manifold algorithm could construct an image segmentation system on the Riemannian manifold space just as the clutering algorithm does on the traditional spectral space.