Scene classification of high-resolution remote sensing (HRRS) image is an\nimportant research topic and has been applied broadly in many fields. Deep\nlearning method has shown its high potential to in this domain, owing to its\npowerful learning ability of characterizing complex patterns. However the\ndeep learning methods omit some global and local information of the HRRS\nimage. To this end, in this article we show efforts to adopt explicit global and\nlocal information to provide complementary information to deep models.\nSpecifically, we use a patch based MS-CLBP method to acquire global and local\nrepresentations, and then we consider a pretrained CNN model as a feature\nextractor and extract deep hierarchical features from full-connection\nlayers. After fisher vector (FV) encoding, we obtain the holistic visual representation\nof the scene image. We view the scene classification as a reconstruction\nprocedure and train several class-specific stack denoising autoencoders\n(SDAEs) of corresponding class, i.e. , one SDAE per class, and classify the test\nimage according to the reconstruction error. Experimental results show that\nour combination method outperforms the state-of-the-art deep learning classification\nmethods without employing fine-tuning.
Loading....