Research On Scene Understanding And Recognition Of High Spatial Resolution Remote Sensing Images

As one of the primary means of earth observation,high-spatial resolution remote sensing image plays an important part in applications such as land use management and monitoring,and smart city project,due to its detailed description of geometric,texture and structural features of objects.The startup of "Major Project for High-Resolution Earth Observation System" pushes the research of high-resolution remote sensing image processing to a new climax.Image analyses based on pixel level and object-oriented level can no longer satisfy the application demands for the complexity of object description caused by increasing spatial resolution.Hence,some newly developed research interests become hotspots,including mining semantic information,designing semantic representation and inference model between objects,and interpreting higher-level and more abstract image information such as scene recognition.Aiming at semantic scene understanding and recognition,this dissertation focuses on several main problems in high-spatial resolution remote sensing image processing,which makes considerable contributions to semantic information mining and land use automatic interpretation of high spatial resolution remote sensing images.The content of this dissertation includes the following four parts.First,large data volume and complex ground elements and structure make it difficult to process the original images directly.Aiming at this,this dissertation studies region of interest(ROI)extraction method for large-scale remote sensing image,for the purpose of improving effectiveness and efficiency of subsequent operations.In this part,a modified model is proposed after studied several visual attention models.Most isual attention models usually focus on spectral information,and spatial information is overlooked.From the spatial distribution perspective,ROI extraction based on geometric distribution is implemented by constructing the edge-corner joint density distribution index.On the basis of this,a ROI hierarchical extraction method based on visual saliency and centroid density distribution index is proposed.The preliminary ROI extraction is implemented using the previously modified visual attention model.Then centroid density distribution index is constructed to refine the ROI extraction.The experimental results prove that the proposed method can extract different types of ROI and reduce the false alarm effectively.Second,aiming at the problem that the traditional low-level visual features cannot represent scene-level semantic information of remote sensing images effectively,this dissertation studies bag-of-visual-words(BOVW)method which specifically focuses on semantic gap.The variables of each step in the BOVW,for instance,segmentation of image patches,selection of feature descriptor,setting of the numbers of visual words,and selection of classifier kernel,are compared and discussed to find proper representation of semantic information.Based on this,spatial information is introduced into BOVW.Combined with optimal scale segmentation,a mid-level feature based semantic classification and scene recognition method is proposed.The experimental results prove that the modified BOVW can improve the accuracy of scene recognition effectively.Moreover,incorporating the semantic information of mid-level features can also improve the classification accuracy.Aiming at the problem that labeled samples cannot be shared between datasets,a remote sensing scene recognition method based on scene transfer learning is proposed.On one hand,taking full advantages of different remote sensing scene recognition datasets,the idea of adversarial learning is introduced into the training process of Variational Auto-Encoders.Information transfer between different remote sensing scene datasets is realized by the adversarial training,using the unlabeled remote sensing scene images in source domain and target domain.On the other hand,since there are plenty of labeled natural scene recognition datasets in the field of computer vision,the cross-view image transfer learning method based on canonical correlation analysis is discussed to eliminate the view discrepancy,and thus realize information transfer between natural scene images and remote sensing scene images.The experimental results show that the high spatial resolution remote sensing image scene recognition based on scene transfer can make up to the overfitting and underfitting phenomenon to some extent and improve the accuracy of scene recognition.Finally,focusing on the insufficiency of labeled samples in remote sensing scene recognition,a small-sample-size high spatial resolution remote sensing scene recognition method based on semi-supervised learning and active learning is proposed.By studying the distribution of unlabeled samples and combining a small amount of labeled samples,a sample selection method integrating density peaks clustering and graph theory is proposed.The method enlarges the training set by uncertainty strategy in active learning and giving pseudo labels for some of the unlabeled samples using density peak clustering and graph theory.In this way,small-sample-size high spatial resolution remote sensing scene recognition is achieved.The experimental results show that the proposed method yields satisfactory results and improves the scene recognition accuracy significantly,especially when the training samples are relatively insufficient.