Research On Remote Sensing Image Object Recognition Method Based On Tensor Pattern

Object recognition is a promising research topic for remote sensing interpretation.Considering the actual situation that remote sensing satellites observe the land,on the one hand,some types of objects,especially military objects,may have only a small number of training samples.On the other hand,since different satellites have different orbits and revisit periods,the available object data can be represented as arbitrary combinations of heterogeneous remote sensing data acquired from satellites with different resolutions and imaging angles.Facing this situation,the existing widely used deep learning-based object recognition methods rely on a large number of samples,and their processing manners are not suitable for the heterogeneous structure of multi-satellite data,let alone processing the input from different satellite combinations.The machine learning-based object recognition methods generally follow the procedure,i.e.,object detection stage,feature extraction stage,and object type classification stage,and offer the dominant performance for small sample situations,making them inherently suitable for processing remote sensing data with multi-array structures.In particular,the tensor-pattern-based machine learning methods,due to their ability to directly process remotely sensed images characterized as multiplexed data,have received considerable attention in remote sensing in recent decades.However,the existing tensor-pattern-based machine learning methods can only accept homogeneous,fixed-satellite remote sensing data as input and fail to adapt to heterogeneous,different satellite combinations of remote sensing data as input to support object recognition for multisatellite remote sensing.To this end,based on tensor theory,this thesis fully excavates the advantages of tensors in the measurement,characterization,and processing and profoundly explores the intrinsic structure and attribute information of multi-satellite heterogeneous remote sensing data.In this way,the traditional vector-based and homogeneous tensor-based object recognition frameworks are extended to a tensor-pattern-based object recognition framework that integrates object detection,feature tensor processing,and object type recognition in multi-satellite remote sensing images,which effectively adapts to small sample and input of heterogeneous multi-satellite data situation.The main work of this thesis is given as follows.For the object detection stage,to attack the tackle that typical object detection methods rely on a large number of samples and are susceptible to object rotating and scaling,the rotation and scale-invariant tensored generalized Hough transformation method is proposed to detect objects of different orientations and scales in remote sensing images effectively.This method uses a tensor-based reference table to replace the discrete table function to capture the characteristic of template objects in different orientations and scales.In addition,combined with the analytic tensor-based voting mechanism and multi-order binary search tree,the space complexity is reduced for processing object detection.Furthermore,by converting the object detection problem into an analytic optimization problem,two types of false alarm removal strategies are derived to improve the precision of detection results.For the feature processing stage,multi-attribute feature tensors are first constructed based on the detected object slices.On the one hand,considering that the object is observed by a single satellite,the coupled heterogeneous Tucker decomposition-based heterogeneous feature tensor transfer method is proposed to transfer the heterogeneous feature tensor with different resolutions and angles to the same feature space,and thus reduce the feature distribution discrepancy between multi-satellite data to achieve the reuse of samples across satellites,which breaks through the limitation that the classical vector-based and homogeneous tensor-based feature transfer technology cannot adapt to the complex structure of multi-satellite heterogeneous data.On the other hand,considering that the object is observed by multiple satellites,to integrate the complementary information of multi-satellite heterogeneous data to enhance the subsequent recognition results,a graph-based and classifier-oriented multi-satellite heterogeneous feature tensor fusion method is proposed,which uses a self-updating graph embedding mechanism to explore the local relationships hidden in the heterogeneous data during the feature fusion process.In addition,combining with classifier-oriented feature discriminative terms,the proposed method can increase the support tensor machine-based hyperplane classification margin to output the fused features tensors for subsequent object type classification.For the object type recognition stage,to remedy the problem that typical classifiers can only accept homogeneous tensors from fixed satellites as input and fail to handle heterogeneous tensors from different satellite combinations,the adaptive heterogeneous support tensor machine method is proposed to exploit the different satellite combinations represented as multiple heterogeneous feature tensor to jointly learn the shared classification hyperplane of multiple heterogeneous tensor spaces,achieving the effective classification for different combinations situations under the single trained classifier.Additionally,to attack the problem that typical support tensor machines can only handle binary classification and objects described by fixed-size slices and fails to process multiclass classification and objects described by different sizes,the multiclass multiscale support tensor machine is proposed to break this limitation,which utilizes the multiscale projection tensor to map objects of different categories and sizes to the category space to achieve object type classification.To verify the effectiveness of the proposed tensor-pattern-based remote sensing images object recognition framework,we have evaluated the performance of the object detection method module,feature transfer module,feature fusion module,and object type classification modules adequately,respectively,using multiresolution and multiangle optical datasets and the other publicly available remote sensing images.The experimental results demonstrate that the proposed framework outperforms typical machine learning-based and deep learning-based methods under small sample situations for the different combinations of multi-satellite remote sensing data.