67 results about "Tensor representation" patented technology

The embodiment of the invention provides an image structuring method and device. The method comprises a step of extracting the characteristic of image based on a first neural network, wherein the characteristic is expressed by using a characteristic tensor, a step of determining an area with possible object of interest in the image according to the characteristic tensor based on a second neural network, a step of calculating the characteristic vector of the area with possible object of interest according to the characteristic tensor, and a step of identifying the object of interest in the area with possible object of interest according to the characteristic vector and determining the position of the object of interest and the attribute of the object of interest based on a third neural network. According to the method and device, the introduction of extra errors in the two links of detecting an object and determining an object attribute is avoided. In addition, the interference in analyzing the object of interest by a background is effectively avoided, and the accuracy of image structural analysis is improved.

The invention discloses a hyperspectral image target detection method based on tensor spectrum matched filtering. The invention relates to target detection of a hyperspectral image. The object of the invention is to solve problems of conventional hyperspectral image target detection methods such as low detection precision and inability of carrying out information mining from three-dimensional data integrally. The hyperspectral image target detection method comprises steps that 1, a target and background signal representation model is established under a tensor representation condition; 2, a to-be-detected hyperspectral image is converted into a three-order tensor form based on a predetermine window, and a hollow X-hollow Y-spectrum-sample four-order tensor 4D is established based on a local neighborhood; 3, covariance matrixes in the hollow X direction, the hollow Y direction, and the spectrum direction of the 4D are acquired; 4, a new three-order tensor after mapping is acquired; 5, the inner products of the target spectrum tensor, the hollow X-hollow Y-spectrum three-order tensor, and the new three-order tensor after the mapping are calculated respectively, and whether the pixel of the to-be-detected hyperspectral image is the detection target is determined. The hyperspectral image target detection method is used for the digital image processing field.

The invention discloses a sparse model of a multi-directional model, wherein the sparse model can ensure remarkable improvement on algorithm complexity and storage space without Kronecker product. The sparse model of the multi-directional model is represented by a formula x=B*1D1*2*D2...NDN, wherein ||B||<=K. In the formula, tension x represents a tension product between an N-directional sparse tension B which belongs to R<M1*M2*...*MN and a series of sparse dictionary Dn which belongs to R<In*Mn>, wherein In<=Mn, Dn is defined as the dictionary in an n-th direction, K is sparsity and is used for representing the number of non-zero elements in a sparse coefficient B. The invention furthermore provides a reconstruction method of the sparse model and a dictionary training method of the reconstruction method.

A method including processing social network data to establish a tensor model of the social network data, the tensor model having at least an order of four. The tensor model is decomposed into a plurality of principal factors. A summary tensor is synthesized from a subset of the plurality of principal factors. The summary tensor represents a plurality of relationships among a plurality of entities in the tensor model. A synthesis of relationships is formed and stored. At least one parameter is identified using one of the summary tensor and a single principal factor in the subset. The at least one parameter is selected from the group consisting of: a correlation among the plurality of entities, a similarity between two of the plurality of entities, and a time-based trend of changes in the synthesis of relationships. The at least one parameter is communicated.

The invention discloses a hyperspectral image target detection method based on tensor matched subspace, which relates to a hyperspectral image target detection method and aims at solving problems that the existing hyperspectral image target detection precision is low and the space information utilization rate is low. The method comprises specific steps: 1, signal representation models for a target and a background under tensor representation are built; 2, four-order tensor matrixes for the target and the background are built respectively; 3, an orthogonal projection matrix in three background directions and three target directions of space X, space Y and a spectrum in the four-order tensor matrixes for the target and the background is solved; 4, a to-be-detected signal is mapped to target sample projection subspace and background sample projection subspace obtained in the third step; and 5, whether the to-be-detected signal is a detection target is judged, if a generalized likelihood ratio detection model value under the tensor representation is larger than or equal to eta, the to-be-detected signal is the detection target, or otherwise, the to-be-detected signal is the background target, wherein eta is a threshold. The method of the invention is used in the image detection field.

The invention provides a multi-view spectral clustering algorithm based on tensor expansion. All view data is expressed on the basis of a tensor and is expanded by the n-Mode multiplication of the tensor, the multidimensional constraint relationship (high-dimensional structure information) of multi-view data is analyzed, and key structure information is stored in view of thoughts of low rank representation and sparse representation so as to establish a solving model based on the tensor expansion. In addition, a noise problem in a practical data acquisition process is considered, and the noise tensor is added to carry out noise-proof processing. Since a non-convex low-rank constraint condition is in the presence in an optimization problem, direct solving is difficult, an optimal object needs to be subjected to convex relaxation, and then, and an ADMM (Alternating Direction Method of Multipliers) algorithm is used for carrying out optimal solving. An experiment result of certain real datasets indicates that a multi-view spectral clustering effect can be effectively improved.

The invention discloses a social network graph abstract generation method based on incremental computation, and belongs to the field of social networks. The method comprises the following steps: carrying out tensor representation on a social network graph in a target time period to obtain a Boolean tensor TG; carrying out tensor decomposition on the Boolean tensor TG to obtain decomposed node matrixes N1 and N2, attribute matrixes A1,..., A<h-3> and a time matrix T; clustering the node matrix N1 or N2 to obtain a clustering cluster center and the type to which each node belongs; regarding thecluster center as hyper-points of the graph abstract, calculating hyper-edge weights between the hyper-points to obtain the graph abstract. According to the method, multi-dimensional data fusion is carried out on nodes, node attributes and timestamps of the social network, and unified expression of high-dimensional graph data and Boolean tensor expression of a complex social network are realized based on binaryzation of a social network graph and high-dimensional expression characteristics of tensors. Incremental CP decomposition is introduced, prior information such as the decomposition result of the old graph tensor is fully utilized, the size of the decomposition tensor is reduced, and the decomposition efficiency of the graph abstract is improved.

The invention discloses a dynamic hypergraph structure learning classification method and system based on tensor representation, and the method comprises the steps: 1, extracting a feature vector of sample data in a database, constructing a hypergraph structure according to the feature vector, and representing the connection strength between any point set in the hypergraph structure through the tensor; 2, introducing a potential energy loss function and an experience loss function in a label vector set in the database, a hypergraph structure expressed by tensor and a potential energy of the point set, generating a dynamic hypergraph structure learning model, performing optimization solution on the dynamic hypergraph structure learning model by using an alternating optimization method, andapplying an optimal solution of a label vector set after model solution to data classification. According to the method, the tensor is introduced to serve as the representation form of the dynamic hypergraph structure and the dynamic hypergraph structure learning method, the hypergraph structure and the label vector of the data are alternately optimized, and finally data classification is achievedaccording to the optimal solution of the label vector of the data.

A method and apparatus for characterizing the performance of a printing device comprising printing a target set of patches with the device and measuring the printing device response with the printed target set; compiling a LUT from the printed target set and measured response; and representing the LUT as a tensor. According to one exemplary embodiment, tensor decomposition/parallel factor analysis is employed for compacting the tensor representation of the LUT.

The invention relates to a vehicle flow prediction method in combination with an attention mechanism and a dynamic space-time convolution model, and provides a road-level flow transfer embedding module to learn tensor representation of a flow transfer time sequence between roads in order to solve the problem that an existing model cannot effectively utilize a vehicle flow transfer relation between the roads; aiming at the problem that the existing method cannot capture the spatial dependence of the dynamic change, a dynamic space attention module is provided and is used for calculating a spatial dependence matrix of the dynamic change; aiming at the problem that an existing model cannot effectively model time sequence importance, a dynamic time attention module is provided and is used for calculating a time sequence attention weight. Through the components, the model can effectively utilize road flow transfer data to capture the space-time dependency relationship of dynamic change between roads, so that the accuracy of road electric bicycle flow prediction is improved.

A convex regularized loss function is minimized respective to a prediction tensor of order K to generate an optimized prediction tensor of order K where K>2. The convex regularized loss function comprises a linear combination of (i) a loss function comparing the prediction tensor and an observation tensor of order K representing a set of observations and (ii) a regularization parameter including a K-th order matrix norm decomposition of the tensor trace norm of the prediction tensor. In some such embodiments, the observation tensor of order K represents a set of social network observations and includes at least dimensions corresponding to (1) users, (2) items, and (3) tags. The optimized prediction tensor of order K is suitably used to perform inference operations.

The invention relates to a streaming data increment processing method and device based on tensor chain decomposition, and the method comprises the steps: constructing a high-order unified tensor representation model of multi-source heterogeneous data; representing the original data as an original tensor according to a high-order unified tensor representation model, and performing tensor chain decomposition on the original tensor to obtain a first tensor chain format; representing the newly added data as a newly added tensor according to the high-order unified tensor representation model, and performing tensor chain decomposition on the newly added tensor to obtain a second tensor chain format; and calculating a tensor chain decomposition result of the updated tensor according to the firsttensor chain format and the second tensor chain format. According to the processing method, the original processing result is quickly and accurately updated by utilizing the newly increased data calculation result, the internal relation between the newly increased data and the existing calculation result can be systematically described, meanwhile, the two problems of intermediate result explosionand repeated calculation of incremental processing are solved, and the processing efficiency of big data is improved.

The invention discloses an unconstrained face recognition method based on weighted tensor sparse graph mapping, and relates to the technical field of face recognition methods. In the sparse graph construction stage, training samples (images) are represented by second-order tensors, a supervised over-complete tensor dictionary is constructed, and similar sparse reconstruction coefficients of the samples are optimized and solved; and a more accurate tensor sparse neighbor graph is constructed in a self-adaptive manner. In a bilateral low-dimensional projection stage, low-dimensional tensor subspace distribution is obtained by utilizing identification information implied in sample global distribution. And low-dimensional mapping yWTSGE = UTyV is performed on the to-be-tested sample y by adopting the optimal WTSGE bilateral projection matrixes U and V, and a classifier is trained by using a low-dimensional training sample DWTSGE = UTXV to realize accurate identity authentication of the non-constrained face. According to the method, the complexity of the non-constrained face image data is fully considered, the neighbor distribution diagram of the high-dimensional tensor data is obtainedin a self-adaptive mode through the sparse representation technology, the low-dimensional manifold essential structure of the highly-distorted non-constrained face data is effectively extracted, andthe accuracy of non-constrained face recognition is greatly improved.

The invention discloses a human body behavior recognition method based on sparse tensors. The human body behavior recognition method comprises the following steps: obtaining a human body behavior silhouette sequence construction sample, and representing each sample with a three-order tensor; using sparse tensor local Fisher discriminant analysis to train a sparse projection matrix group; projecting the training sample and the to-be-tested sample to a tensor subspace by using the obtained sparse projection matrix group; and calculating the similarity between the to-be-tested sample and the training sample, and constructing a nearest neighbor classifier based on the tensor distance to identify the to-be-tested sample in the low-dimensional space. According to the tensor local Fisher discriminant analysis algorithm of the invention, the problems of eigenvalue solving and eigenvector solving are converted into a series of linear regression problems, so that the objective of tensor local Fisher discriminant analysis can be satisfied, and the sparsity of an obtained projection matrix can be ensured.

The invention discloses a sleep physiological signal feature extraction method and system based on tensor complexity. The sleep physiological signal feature extraction method comprises the following steps: a sleep physiological signal is collected, and a multi-guide time sequence of the sleep physiological signal is converted into tensor representation, wherein the sleep physiological signal is expressed as an N-order tensor, and the N-order tensor is composed of N-order sub-tensors; and the difference between the elements in each sub-tensor and the elements in the global sub-tensor is judgedto determine the approximate entropy of the N-order tensor, the tensor approximate entropy is used as the feature of sleep physiological signal extraction, the extracted feature can accurately reflectthe internal feature of the sleep physiological signal data, and then the sleep staging is more accurate in the subsequent classification processing.

The invention discloses a lightweight multi-modal sentiment analysis method based on multi-element hierarchical deep fusion. According to the method, the direct correlation among the multi-modal elements is established in a layered manner, and the short-time and long-time dependence among different modals can be captured. In order to avoid reduction of resolution and retain original spatial structure information corresponding to each modal, corresponding attention weights are applied in a broadcast form when multi-modal information interaction is selected and emphasized. Moreover, the invention also provides a new tensor operator, called Ex-KR addition, so that the shared information is utilized to fuse the multi-modal element information to obtain global tensor representation. The methodis an effective supplement for solving the problems that in the current multi-modal emotion recognition field, most methods only pay attention to modeling in a local time sequence multi-modal space, and all complete representation forms participating in modal fusion cannot be clearly learned.

The tensor is capable of sufficiently capturing possible relationships between a plurality of views of data when representing multi-view data, while also avoiding pairwise comparison between views. According to we know, a multi-view outlier detection mode based on tensor representation has not been studied up to now. Existing multi-view outlier detection methods mostly employ cross-view pairwise constraints to obtain new feature representations and define outlier scoring metrics according to the features, which do not make full use of interaction information between views and result in highercomplexity in facing three or more views. In order to make up the defect, the invention provides a new multi-view outlier detection algorithm based on tensor representation. Specifically, firstly, multi-view data is reshaped into a tensor set form, then low-rank representation of the tensor set form is learned, and finally an outlier function under tensor representation is designed to achieve detection.

The invention discloses an image tensor data processing method which can greatly reduce the number of free parameters in a model, is flexible in weight layer limitation and can be suitable for image tensor data of any order. According to the image tensor data processing method, a restricted Boltzmann machine TTRBM model with a tensor structure is introduced, input data and output data of the method are expressed by tensors, weights of an intermediate layer are also expressed by tensors, and the weights are limited to have the structure of a Trustin. The number of free parameters in an intermediate layer is controlled by adjusting the rank of tensor Train decomposition; the rank of TT decomposition is adjusted, and different feature representations of the same size are adjusted.