54 results about "Higher-order singular value decomposition" patented technology

The invention is directed generally to providing methods and systems for trend extraction and analysis. Embodiments include methods and systems for trend extraction and analysis of information extracted from dynamically changing data included in computer systems and / or networks. Various exemplary embodiments are provided that may generate characteristic indicators for trend(s) and / or distribution(s) for one or more data sources by use of, for example, temporal indicators derived through analysis of the difference in contribution separate portions of the data to the whole data set being considered, contribution of individual sources, and / or the interaction of the separate portions of the data with one another. Some exemplary approaches may include the use of singular value decomposition (SVD) and higher-order singular value decomposition (HOSVD) data extraction and analysis techniques. One use of these techniques is in the analysis of the dynamic data contained in Weblogs and the blogosphere.

The invention discloses a method for identifying multi-view human face based on nonlinear tensor resolution and view manifold, which comprises the following steps: normalizing the size of the multi-view human face; dividing multi-view human face images into a test set and a training set by adopting a method of leaving one out; arraying the human face images in the training set into a form of tensor along the direction of identity, view and pixel information change, resolving tensor data by using high-order singular values to obtain a coefficient matrix of identity, view and pixel factors of the human face images; using a data-concept driving mode to array and interpolate view coefficients to obtain the view manifold of human face; according to the rotating objective sequence of the human face, generating the view manifold through a concept driving mode; using the nonlinear tensor resolution to map the view manifold to a data space of the multi-view human face, obtaining a modular matrix of identity coefficient, and establishing a model of the multi-view human face; and adopting an iterative algorithm based on EM-like to solve a model parameter, and achieving identification by the parameter meeting the minimum reconstructed error criterion. The method has the advantages of high accuracy and high speed, and can be used for complex human face retrieval and identification under different view angles in the field of biological characteristic identification.

The invention discloses a tensor decomposition based context-dependent place recommendation method and mainly solves the problem of poor place recommendation quality in the prior art. The method is implemented by the steps of 1, constructing a three-dimensional sign-in tensor A and a user similarity matrix B by utilizing sign-in data of all users of a to-be-recommended city; 2, obtaining a three-dimensional tensor A by using a high-order singular value decomposition algorithm; 3, obtaining the current position of a to-be-recommended user c; and 4, according to the three-dimensional tensor A, performing place recommendation on the to-be-recommended user c. According to the method, the communication traffic between the user and a recommendation system is reduced by utilizing tensor decomposition, and the effectiveness and reliability of a place recommendation result in a data sparsity scene is ensured in combination with a time context and historical user data; and the method can be applied to position based place recommendation services in a social network.

The invention discloses a MIMO (Multiple-input Multiple-output) radar angle estimation algorithm based on tensor subspace and rotation invariant. Through building a three-order tensor model for receiving data, a higher-order covariance tensor model for tensor data is further built, and the internal correlation structure of array signals is thoroughly exploited; HOSVD (Higher Order Singular Value Decomposition) is then carried out on the tensor data, new signal subspace is built, and high-precision noise subspace is thus acquired; and finally, by using the rotation invariant of array data, a DOD (Direction of Departure) and a DOA (Direction of Arrival) which are paired are acquired through a constrained optimization method and a lagrangian multiplier method, and further pairing calculation is not needed. According to the MIMO radar angle estimation algorithm of the invention, by using the internal correlation structure of the received signals, the angle estimation precision is higher, the target DOD and the target DOA which are more accurate can be acquired, a more reasonable reference is provided for further processing on a detected target, spectral peak search is not needed, and the computation complexity is low.

The invention discloses an earthquake data compression method based on tensor adaptive rank truncation. According to the method, based on given compression conditions and a different-dimension singular value setting truncation rank acquired through instant higher-order singular value decomposition, tensor decomposition is carried out further according to singular value distribution determination truncation rank magnitude; on the condition that a compression rate is guaranteed, a compression peak value signal to noise ratio is improved, and the compression effect is improved; through designing the tensor decomposition flow based on adaptive rank truncation, manual parameter regulation and control in the compression process can be reduced, and data compression work is simplified. On the basis of the singular value distribution setting truncation rank magnitude, data anisotropy characteristics are utilized, a singular value standard deviation is introduced to calculate weight of a truncation rank corresponding to different dimensions, the compression peak value signal to noise ratio is improved, and the data compression effect is further improved.

The invention discloses an MIMO radar angle estimation algorithm based on tensor space and spectral peak search. Through building a three-order tensor model for receiving data, a higher order covariance tensor model for tensor data is further built, and the internal correlation structure of array signals is fully dug; HOSVD (Higher Order Singular Value Decomposition) is then carried out on the tensor data, new signal subspace is built, and high-precision noise subspace is thus acquired; and finally, the orthogonality properties between a steering vector and the noise subspace are used for spectral peak search to acquire a paired DOD (Direction of Departure) and DOA (Direction of Arrival), and further pairing calculation is not needed. According to the MIMO radar angle estimation algorithm, the internal correlation structure of received signals is used, GPE does not need to be estimated firstly, the angle estimation precision is higher, the reliability is stronger, the target DOD and the DOA which are more accurate can be acquired, and more reasonable reference is further provided for related processing on detection targets.

Computer-readable media having computer-executable instructions and apparatuses categorize documents or corpus of documents. A Tensor Space Model (TSM), which models the text by a higher-order tensor, represents a document or a corpus of documents. Supported by techniques of multilinear algebra, TSM provides a framework for analyzing the multifactor structures. TSM is further supported by operations and presented tools, such as the High-Order Singular Value Decomposition (HOSVD) for a reduction of the dimensions of the higher-order tensor. The dimensionally reduced tensor is compared with tensors that represent possible categories. Consequently, a category is selected for the document or corpus of documents. Experimental results on the dataset for 20 Newsgroups suggest that TSM is advantageous to a Vector Space Model (VSM) for text classification.

Disclosed is a higher-order singular value decomposition based magnetic resonance image denoising method. The method includes: (1) subjecting an original image to variance stability transformation; (2) subjecting the transformed image to first higher-order singular value decomposition denoising, specifically, (a) searching similar blocks of reference blocks one by one to form a higher-dimensional array for the reference blocks corresponding to pixels of each object, (b) subjecting the higher-dimensional array to higher-order singular value decomposition transformation to form coefficients and adaptive bases, (c) performing first threshold operation, and (d) performing higher-order singular value decomposition inverse transformation; (3) performing weighted average and pixel binning; (4) performing weighted average to obtain an image after first weighted average; (5) subjecting the image after first weighted average to second higher-order singular value decomposition denoising to obtain an image after second denoising; (6) subjecting the image after second denoising to variance stability inverse transformation to obtain a filtered image. By the method, magnetic resonance image noise can be suppressed effectively, and image quality is improved.

The invention discloses a video fusion performance evaluating method based on high-order singular value decomposition. The video fusion performance evaluating method based on the high-order singular value decomposition mainly solves the problem that the fusion performance of video images containing noise cannot be evaluated in the prior art. The video fusion performance evaluating method based on the high-order singular value decomposition comprises the implementation steps that two registered reference videos and a registered fusion video are input respectively; a four-order tensor is formed by the input videos, the high-order singular value decomposition is carried out on the four-order tensor, and respective space geometrical characteristic background images and respective time motion characteristic images are obtained; the time motion characteristic images are divided into motion target areas and noise areas through a thresholding method; then, different evaluation indexes are designed respectively to evaluate all the areas; lastly, an overall performance evaluation index is structured through power exponent multiplication, and thus evaluation on the whole fusion performance of the video images is achieved. The video fusion performance evaluating method based on the high-order singular value decomposition can effectively, accurately and objectively evaluate the fusion performance of the videos in a noise environment, and can be used for monitoring fusion video image quality.

The invention relates to a high-order singular value decomposition (HOSVD) algorithm based video compression system and method. The video compression system comprises a camera device and a corresponding sensor. The camera device and the corresponding sensor are respectively connected with a complex programmable logic device (CPLD) processor and a first in first out (FIFO) memory. The CPLD processor and the FIFO memory are connected with a digital signal processor (DSP). The CPLD processor is connected with the FIFO memory and an audio video standard (AVS) coder separately. The AVS coder is connected with a read only memory (ROM) video memory. According to a tensor decomposition theory, hardware and software are combined, so that time domain and spatial domain redundant information is compressed, data storage capacity is compressed furthest, and effective storage of color videos and minimum information loss are realized.

The invention discloses a coupling tensor decomposition-based multispectral image and hyperspectral image fusion method, and mainly aims at solving the problems of space information fuzziness and spectrum distortion in the prior art. The method comprises the following steps of: (1) respectively inputting a low-spatial resolution hyperspectral image, a high-spatial resolution multispectral image, aspectrum degradation matrix and a regular parameter; (2) carrying out spatial up-sampling on the low-spatial resolution hyperspectral image; (3) carrying out Tucker decomposition on the up-sampled low-spatial resolution hyperspectral image by utilizing a high-order singular value decomposition algorithm; (4) carrying out Tucker decomposition on the high-spatial resolution multispectral image; and(5) multiplying a factor matrix obtained in the step (3) by a core tensor obtained in the step (4) along each model so as to obtain a high-spatial resolution hyperspectral image. The method is capable of obtaining correct space and spectrum information and can be applied to the remote sensing fields such as target detection and recognition and terrain classification.

The invention provides an angle estimating method for a bistatic MIMO radar non-circular target under mutual coupling conditions. The method includes the following steps: a receiving array element receives mutually orthogonal non-circular signals emitted by an emitting array element, wherein, both the receiving array element and the emitting array element are uniform linear arrays with half wavelength; matched filtering is performed on the received non-circular signals through a matched filter of a receiving array element receiver; decoupling operation is performed on matched signal data, andthe decoupling operation is performed in a tensor domain; augmented tensor is constructed by utilizing a non-circular and multidimensional structure of the decoupled signal data in the tensor domain;high-order singular value decomposition is performed on the augmented tensor so that signal subspace can be obtained; and according to the obtained signal subspace, estimation is performed on the emitting angle and receiving angle of a target by combining a rotation invariant matrix, so that the angle estimation of the target can be obtained.

Computer-readable media having computer-executable instructions and apparatuses categorize documents or corpus of documents. A Tensor Space Model (TSM), which models the text by a higher-order tensor, represents a document or a corpus of documents. Supported by techniques of multilinear algebra, TSM provides a framework for analyzing the multifactor structures. TSM is further supported by operations and presented tools, such as the High-Order Singular Value Decomposition (HOSVD) for a reduction of the dimensions of the higher-order tensor. The dimensionally reduced tensor is compared with tensors that represent possible categories. Consequently, a category is selected for the document or corpus of documents. Experimental results on the dataset for 20 Newsgroups suggest that TSM is advantageous to a Vector Space Model (VSM) for text classification.

The invention discloses a motion detecting method which comprises the following steps: expressing a current training area sequence of a training background image sequence into a high-level tensor; analyzing a high-level singular value of the tensor to obtain a tensor subspace of the tensor; and detecting the motion of a corresponding area to be detected of the current image to be detected by using the tensor subspace. Moreover, the invention also discloses a motion detecting device, a background model establishing method and a background model establishing device. The disclosed technical scheme of the invention can increase motion detecting accuracy.

The invention provides a radio frequency interference suppression method for a high frequency surface wave radar based on tensor subspace projection. Depending on the distance correlation and direction characteristics of radio frequency interference, joint estimation is performed on a radio frequency interference subspace from multiple dimensions by tensor decomposition. The implementation processcomprises the following steps: performing Doppler sliding window processing based on radar data to finish data segmentation; performing radio frequency interference detection; judging whether interference frequency points exist in a currently-selected data block or not; if radio frequency interference exists, constructing a third-order tensor by using long-distance metadata without sea echo of multiple antenna channels to serve as a training tensor; performing tensor decomposition on the training tensor by using high-order singular value decomposition; estimating a radio frequency interference subspace; and performing distance sliding window and subspace projection processing to obtain data without radio frequency interference. The method can suppress radio frequency interference to a great extent, improve data quality and retain a target signal completely, has a good application effect, improves the detection performance of a radar system, and has important practical application value.

The invention provides a radar angle and distance estimation method based on tensor high-order singular value decomposition, and the method comprises the following steps: constructing an MIMO radar receiving array comprising k subarrays, obtaining the receiving data of a target echo through the MIMO radar receiving array, and constructing a third-order tensor signal model; decomposing the three-order tensor signal model by adopting a high-order singular value to obtain tensor-based signal subspaces; extracting a receiving matrix of the signal subspace to realize MIMO radar target DOA parameter estimation; and extracting a signal subspace and an emission matrix corresponding to each subarray, eliminating the phase ambiguity problem caused by the coupling of the target DOD parameter and the distance, achieving the automatic pairing of the target DOD and the distance with the DOA parameter, and finally achieving the estimation of the target DOD parameter and the distance parameter of the MIMO radar.

The invention discloses a context recommendation method and device based on high-order singular value decomposition. The method comprises the steps of obtaining grading information of a user on a project and corresponding context information; constructing a three-order tensor corresponding to each piece of context according to different context types; expanding each three-order tensor according toan expansion rule to obtain three second-order matrices; according to all the second-order matrices, determining a central tensor dimension by utilizing the singular value decomposition, and constructing a new third-order tensor; calculating the weight of each piece of context; constructing an N-order tensor according to the new third-order tensor and the corresponding context weights; finding the position corresponding to the target user on the N-order tensor according to a target user ID and a project ID, and generating a recommendation list for the target user. By means of the method, thecontext information is fused into recommendation generation, and by calculating the context weights and determining the central tensor dimension, the accuracy of a recommendation result is greatly improved.

The invention discloses a hyperspectral anomaly target intelligent detection method based on a robust spectral covariance distance, and the method comprises the following steps: constructing a spatialdimension factor matrix according to high-order singular value decomposition, so as to fully extract the spatial dimension information of a hyperspectral image; classifying all pixels of the hyperspectral data into k categories by using a clustering algorithm, removing pixels with the number of pixels less than P in a clustering cluster, and calculating kernel space anomaly indexes of the remaining pixels according to each cluster, so that the first P pixels with the maximum superposition sum are finally selected as constituent atoms of a spectral dimension factor matrix; establishing a hyperspectral image anomaly intelligent detection model based on robust spectral covariance distance regularization, constructing a Lagrange equation, iteratively solving a certain variable step by step while fixing other variables, solving an anomaly detection model, and obtaining an abnormal target according to the obtained solution. The method can achieve the intelligent detection of the abnormal target in a hyperspectral remote sensing image, and effectively reduces the false alarm rate.

The embodiment of the invention provides a clustering method of multi-source heterogeneous data, which comprises the following steps of: aiming at the multi-source heterogeneous characteristic of a characteristic space, fusing the characteristic space to construct more than one object tensor and characteristic space combination vector; obtaining a corresponding feature transfer tensor according tothe object tensor; processing the feature transfer tensor by adopting a preset multi-relation attribute combination ranking algorithm to obtain a corresponding attribute combination ranking tensor; decomposing the object tensor and the attribute combination ranking tensor by adopting a preset high-order singular value decomposition algorithm to obtain a corresponding core tensor and a factor matrix; calculating according to the feature space combination vector, the core tensor and the factor matrix to obtain a corresponding object similarity matrix; and performing clustering according to theobject similarity matrixes under different feature spaces to obtain a multi-clustering result. The method solves the problem that the clustering result is difficult to interpret in the prior art.

The invention provides a silicon material melting process monitoring method based on high-order singular value decomposition. A CCD camera is used for real-time image sequence collection in a silicon material melting process first; a plurality of two-dimensional image sequences previous to the current moment are used for forming a three-dimensional tensor, and the tensor is subjected to high-order singular value decomposition (HOSVD) to obtain base images used for representing a previous melting state; then, a correlation between images collected in the current silicon material melting process and the base images containing main formation is calculated, and whether the silicon material melting state changes is judged according to the correlation. By means of the silicon material melting process monitoring method based on high-order singular value decomposition, the change of state of silicon materials can be detected under the condition that the background of the silicon material melting state is relatively complicated in an initial stage and an intermediate stage, the change in the silicon material melting state and noise of molten liquid movement can be distinguished easily in a later stage of melting, and therefore the silicon material melting process is monitored accurately and effectively.

The invention discloses a gesture classification method based on tensor decomposition; the gesture classification method comprises the following steps: using a third-order tensor to model gesture videos; using a modified higher-order singular value decomposition method to decompose each gesture video; carrying out visualized presentation and parsing the tensor decomposition result; respectively using a typical angle to classify the gesture videos through a K neighbor classifier and a support vector machine classifier; changing a factor matrix column vector number so as to carry out a contrast experiment. The method can visualize the tensor decomposition result, and the result can be presented for classification, thus better understanding the tensor decomposition physics meanings.

The invention discloses a multipath data floating point processor prototype. A framework is characterized by aligning structured data, half-structured data and non-structured data to be three paths of arrays; applying a high singular value decomposition, and decomposing the three arrays to be a two-order tensor matrix mode; transforming the matrix mode to a sparse domain, and performing blocked floating point quantization treatment. Finally, the multipath data floating point processor prototype is structured.

The invention relates to a missing travel time filling method based on tensor decomposition. The missing travel time filling method comprises the following steps: step 1, preprocessing original bayonet data; and step 2, filling the missing travel time, wherein the step 2 comprises the following sub-steps: S201, performing tensor construction; S203, calculating an initial solution for the input ofthe HOOI method by adopting high-order singular value decomposition; S202, performing Tucker decomposition of the travel time tensor; and S204, iteratively calculating a final solution by adopting anHOOI method. Through the missing travel time filling method, the filling accuracy is improved while the filling efficiency is improved.

The invention provides a high-precision positioning algorithm and device for a bistatic EMVS-MIMO radar, and the algorithm comprises the steps: constructing a sample matrix after the matched filteringof a receiving array, and calculating a covariance matrix estimation value of a received signal; constructing a high-order received signal covariance tensor model by using the Tucker tensor model; performing high-order singular value decomposition on the high-order covariance tensor model to obtain a new signal subspace and a new noise subspace; obtaining estimation of an azimuth angle and a pitch angle of the target and estimation of a two-dimensional polarization angle by using a rotation invariant technology and a vector cross product technology; and pairing of DODs and DOAs is realized byusing a subspace orthogonal principle. According to the method, a tensor structure for receiving array signals is considered, and high-precision target multi-parameter estimation can be obtained based on a tensor subspace algorithm.

The invention provides a filtering method and apparatus for a compressed video. According to the method and apparatus, through high-order singular value decomposition processing, the non-local spatialcorrelation in a video frame can be effectively utilized, the quality of the compressed video is improved, and the video compression performance is further enhanced. Moreover, the filtering method isapplied to a video encoding loop so that the quality of a reference frame can be improved, the prediction efficiency is further improved, and the encoding performance is enhanced. The filtering method for the compressed video based on the non-local spatial correlation is proposed, and the corresponding apparatus is also proposed.

The invention relates to the field of magnetic resonance imaging, and discloses a rapid myocardial perfusion magnetic resonance imaging method based on low-rank tensor coding, which comprises the following steps of: (a) searching three-dimensional similar blocks in a myocardial perfusion image frame according to a certain similarity criterion by adopting a k-nearest neighbor method, and clustering to construct a three-dimensional low-rank tensor; (b) performing high-order singular value decomposition on the constructed three-dimensional low-rank tensor to utilize space-time redundancy of the myocardial perfusion sequence image; (c) introducing an augmented Lagrange technology, and establishing an image reconstruction model based on low-rank tensor coding; and (d) solving the reconstruction model by using an alternating direction multiplier method to obtain an imaging result. The space-time redundancy of the myocardial perfusion magnetic resonance image is mined through low-rank tensor coding, so that the image data can be better sparsely represented, and a satisfactory image reconstruction effect is obtained under over-sampling and under-sampling, so that the speed of myocardial perfusion magnetic resonance imaging is increased.

The invention discloses a modeling method for a multi-cell LPV system of an elastic hypersonic aircraft. The method comprises the following steps: establishing an elastic hypersonic aircraft nonlinearlongitudinal model, obtaining corresponding aerodynamic parameters, analyzing the relationship between the elastic deformation of an aircraft and a rigid body of the aircraft, and establishing a mathematical model; establishing a three-input three-output multi-cell LPV system model and a two-input two-output multi-cell LPV system model for the nonlinear longitudinal model of the hypersonic aircraft based on Jacobian linearization and tensor product transformation; to be specific, selecting a scheduling variable, and performing Jacobian linearization at a balance point to obtain a continuous LPV model; discretizing the continuous LPV model and storing the discretized continuous LPV model in a tensor; performing high-order singular value decomposition on the tensor to extract a core LTI vertex system; and constructing a continuous weight function of a core LTI vertex system. According to the method, the multi-cell LPV model capable of reflecting the elasticity of the hypersonic flight vehicle is established, the characteristics of the large flight envelope of the hypersonic flight vehicle are covered, the modeling calculation amount is reduced, and meanwhile the modeling precision is kept.

The invention discloses a multi-exposure image fusion method based on high-order singular value decomposition. The multi-exposure image fusion method includes the steps: dividing a brightness channelimage of an exposure image into overlapped brightness blocks; obtaining a kernel tensor and a third mode factor matrix of the brightness blocks by utilizing high-order singular value decomposition; obtaining a characteristic coefficient and an activity level measure of the brightness blocks, obtaining a fused brightness block according to the first mode factor matrix and the second mode factor matrix of the brightness block and the characteristic coefficient and the activity level measure, and performing linear transformation on the obtained brightness channel image to obtain a fused brightness channel image; obtaining a fused first chroma channel image by calculating a fusion coefficient of pixel points in the first chroma channel image of the exposure image; obtaining a fused second chroma channel image by calculating a fusion coefficient of pixel points in a second chroma channel image of the exposure image; and obtaining a fused image according to the fused images of the three channels. The multi-exposure image fusion method has the advantage that better detailed textures and rich color information can be obtained.

The invention discloses a tensor-based background subtraction method and system. The method comprises the steps of obtaining an observation tensor; carrying out high-order singular value decompositionon the observation tensor to obtain an initialized core tensor, a factor, an initialized auxiliary tensor, a Lagrange multiplier and the number of iterations; decomposing the observation tensor intoa background tensor and a foreground tensor; constraining a background tensor through a tensor sparse detection method based on Kronecker-based representation, and constraining a foreground tensor through a norm to obtain a minimization problem; obtaining an iterative core tensor, an iterative factor, an iterative auxiliary tensor and an iterative foreground tensor by solving a minimization problem; judging whether the change between the iterative foreground tensor and the foreground tensor is smaller than a preset threshold value or not; if not, updating the Lagrange multiplier; if so, obtaining an iterative background tensor according to the iterative core tensor and the iterative factor; and completing foreground and background separation according to the iterative background tensor andthe iterative foreground tensor. The invention can quickly and accurately separate the background from the foreground.

The invention discloses a prediction method based on high-dimensional data structure relations. The prediction method comprises the following steps of obtaining posts and replies of a Sina share postbar and an Eastern Fortune share post bar through a crawler technology and obtaining specific news of Baidu's advanced search; conducting natural semantic processing on the posts and news to obtain features and combining with formula calculations to obtain technical indicator features; constructing a tensor of the three features, and reconstructing the tensor by high-order singular value decomposition to achieve the purpose of denoising and strengthening the relationship between various factors; obtaining tensor information bodies of the above steps every day, and reconstructing data by a reconstruction algorithm proposed by us by using the correspondence relationship with lifting price information to obtain a new tensor sequence, wherein the information bodies that have the similar levelof lifting or the same direction of lifting are similar through reconstruction restrictions; conducting optimized tensor ridge regression for the new tensor sequence; using a regression prediction asa secondary trading system.