68 results about "Tucker decomposition" patented technology

The invention discloses a deep network compression method based on dimensional adaptation adjusting Tucker decomposition. The method comprises a dimensional adaptation adjusting process and a dimensional adaptation weight tensor decomposition process. Through adaptively adjusting a size of each dimensions of a tensor, a new any order tensor is generated. Through a learnable nuclear tensor and a transfer matrix, tensor decomposition is realized so that a purpose of network optimization compression is reached. Compared to an existing low rank compression method, by using the method of the invention, under the condition that network performance is maintained, a network parameter quantity has a greater compression multiple and a higher compression multiple can be acquired. Simultaneously, a guiding position of a nonzero element does not need to be stored, an index does not need to be recorded and a storage space can be effectively used.

The invention provides a lithium battery unit defect detection method. The main technical scheme is that images of the front side and reverse side of a lithium battery unit are acquired, a Tucker machine supporting method is applied, defected and non-defected lithium battery unit image sets which are preprocessed are trained to obtain weight tensor, core tensor and other parameters, then the unit images to be detected are judged by applying the Tucker machine supporting method according to the parameters obtained through training to complete defect classification, and accordingly detection results are obtained. The lithium battery unit defect detection method utilizes the machine, uses tensor type of weight parameters obtained through Tucker decomposition and can completely retain information and structures of image data, and classification accuracy is improved. The lithium battery unit defect detection method can be widely applied to image based defect detection.

The invention discloses a face feature extraction method based on heterogeneous tensor decomposition. The method includes: using an arbitrary camera array to collect color views of different faces at different visual angles, and obtaining an initial view set of the faces after image processing such as gray-scale transformation and normalization; extracting images of the initial view set, and carrying out sequential stacking to form a third-order tensor, wherein a third-order dimension of the formed third-order tensor corresponds to the total number of the face images; carrying out TUCKER decomposition on the formed third-order tensor to obtain a kernel tensor, a first pattern factor matrix, a second pattern factor matrix and a third pattern factor matrix, and updating the same; judging whether the kernel tensor converges; and decomposing the obtained third pattern factor matrix Z, wherein high-dimensional face data are mapped to the pattern factor matrix of low-dimensional feature subspace, and thus the third pattern factor matrix Z is finally extracted face features. According to the method, automatic extraction of the face image features is realized, tedious steps of traditional feature extraction are avoided, and feature extraction speed is increased.

The invention discloses a remote-sensing image digital watermark embedding and extracting method based on the quaternary wavelet. The embedding process comprises carrying out quaternary wavelet decomposition on R, G and B wave bands of an original color remote-sensing image, making amplitude low frequency coefficients in the three wave bands form a third-order tensor, and then carrying out Tucker decomposition to obtain a core tensor; carrying out Arnold scrambling transformation on an original watermark image; embedding the scrambled watermark information into the core tensor, reconstructing the tensor and carrying out quaternary wavelet inverse transformation, and obtaining a color remote-sensing image including a watermark. The extracting process comprises carrying out quaternary wavelet decomposition on R, G and B wave bands of the color remote-sensing image including the watermark, making amplitude low frequency coefficients in the three wave bands form a third-order tensor, and then carrying out Tucker decomposition to obtain a core tensor; calculating the singular value of the core tensor to extract embedding information; and obtaining watermark information by means of inverse transformation of the Arnold scrambling transformation. According to the invention, invisibility and anti-attack capability of the watermark can be improved.

The invention discloses a tensor space-based no-reference high dynamic range image objective quality evaluation method. From image data, a colorful high dynamic range image is represented with a three-order tensor; then a distorted high dynamic range image is subjected to tensor decomposition through a Tucker decomposition algorithm in tensor decomposition to obtain three characteristic images, namely, a first characteristic image, a second characteristic image and a third characteristic image are obtained, wherein the three characteristic images fuse color information of the high dynamic range image; a manifold structure characteristic of the first characteristic image and perception detail contrast characteristics of the second and third characteristic images are extracted; and an objective quality evaluation value of the distorted high dynamic range image is calculated by utilizing a support vector regression method in machine learning, so that the objective quality evaluation of the no-reference colorful high dynamic range image is realized, the evaluation effect is remarkably improved, and the correlation between an objective evaluation result and subjective perception is effectively improved.

The invention provides a tensor decomposition cutoff remote sensing hyperspectral image compression method based on fast optimal core configuration search and relates to a hyperspectral image processing method. Aiming at the problem that a compression method based on the tensor decomposition cannot easily and fast obtain the optimal tensor core configuration under the requirement of setting the compression quality and the compression ratio, the tensor decomposition cutoff remote sensing hyperspectral image compression method based on the fast optimal core configuration search is provided. The method has the following steps that hyperspectral images are subjected to complete Tucker decomposition; spectrum dimension search starting points are calculated, iterative search is started, and the spectrum dimension optimal configuration is obtained; then, the trimming iteration is carried out, and the space dimension optimal configuration is obtained; and finally, complete decomposition results are intercepted, and final compression results are obtained. The method can be applied to satellite-bone or ground hyperspectral image compression, the compression recovery quality is ensured, and meanwhile, the calculation quantity of the compression method can be effectively reduced.

The invention provides a user behavior data mining-based electric energy experience analysis method which comprises the following steps: historic electricity consumption data of all users on a user side; a three dimensional tensor model is built while date information, time information and user information are used as coordinate axes; a Tucker decomposition method is adopted for decomposing the three dimensional tensor model; two dimensional characteristic matrixes that respectively correspond to the data information, the time information and the user information; a distribution law of electricity consumption loads of users is determined; cluster analysis is conducted, and user electricity consumption mode types are built according to all kinds of cluster results; all levels of user sensitivity are determined and combined with electricity consumption modes, an electricity consumption mode prediction model is established, current actual electricity consumption data of object users is obtained and imported into the electricity consumption mode prediction model, and current electricity consumption mode types of the object users and corresponding current sensitivity levels are obtained. Via use of the analysis method, user electric energy experience analysis results can be obtained from a user angle based on user side load sensitivity.

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.

A video sequence classifying method based on a tensor time domain association model comprises the steps of representing an original video sequence in a three-grade video tensor manner; performing tensor Tucker decomposition on the three-grade video tensor for obtaining a latent kernel tensor; applying an autoregression model on the time domain of the obtained latent kernel tensor for establishing relevance between adjacent time slices; dynamically studying the front process, updating the result until an algorithm is convergent, and obtaining an optimal result. The video sequence classifying method ensures time domain relevance and dependence of the video sequence after dimension reduction through limiting the time domain of the video sequence. The video sequence classifying method has advantages of sufficiently utilizing latent useful information in the video, eliminating redundant information in the video, ensuring high continuity of the video sequence in time domain, reducing classification difficulty of the video sequence, and improving classification accuracy of the video sequence. The video sequence classifying method is better than a traditional video sequence classification method and greatly improves classification precision.

The invention discloses a method for detecting and repairing traffic loss data by using an iterative tensor decomposition (ITD) algorithm. The method comprises a step of constructing traffic data including lost data into a three-dimensional tensor of position*date*time, a step of determining the size of a core matrix by using a truncated singular value decomposition (TSVD) method based on a Tuckerdecomposition method and obtaining an initialized core matrix and three factor matrices, a step of establishing a target equation with the tensor as input and with the initialized core matrix and thethree factor matrices as initial values, solving the target equation by using a gradient descent method, updating the core matrix and the three factor matrices, and obtaining an initial repair tensorby using an SVD-combined tensor decomposition (STD) method, and a step of detecting and repairing the lost data in the traffic data by using an ITD algorithm with the initial repair tensor as input.

The invention provides an engine fault diagnosis method based on tensor tucker decomposition fuzzy control. Data in an existing diesel engine ECU are integrated to form third order tensioners; a tensor tucker decomposition thought is utilized, feature extraction is carried out by using Tucker decomposition of a tensor of HOSVD-ALS to obtain a core tensor and a horizontal slice matrix Bm thereof; and data normalization processing is performed on the data of the core tensor horizontal slice matrix Bm, reasoning operation is performedon the two processed tensors by adopting a fuzzy method to obtain an output quantity matrix, and defuzzification is performed on the output quantity matrix to obtain an accurate output quantity matrix, thereby realizing fault diagnosis on short circuit, open circuit and failure of the diesel engine sensor.

The invention provides a motion recognition method based on sparse coding tensor decomposition. An original video sequence is expressed as a three-order video sequence tensor A belongs to R<I<1>xI<2>xT>, wherein T refers to the length of the video sequence, and I<1>xI<2> refers to the size of the video frame; Tucker decomposition is performed on the three-order video sequence tensor A belongs to R<I<1>xI<2>xT> so that a nuclear tensor of which the spatial domain dimension is reduced is acquired; the video sequence tensor is zoomed to the same scale; and the result is updated by dynamically learning the process until the algorithm convergence result achieves the optimum. According to the motion recognition method based on sparse coding tensor decomposition, the video sequence can be processed into the unified length-the sparse coding tensor decomposition technology. The frames of the most information are adaptively selected from a tensor decomposition framework in the process to construct a new video sequence having the unified video sequence length. According to the method, the difficulty of gesture recognition can be reduced and the accuracy of gesture recognition can be enhanced so that the great conditions can be provided for subsequent video sequence classification, and the accuracy of video sequence classification can be enhanced.

The invention discloses a Tucker decomposition-based spectral tensor dimension reduction and classification method. The method comprises the steps of constructing a three-order tensor by taking factors influencing a spectral feature of a ground object as intra-class factors and taking the intra-class factors, classes and pixel spectrums as modes respectively, and performing low-rank tensor decomposition-based dimension reduction on the three-order tensor; performing low-rank tensor decomposition on the three-order tensor to obtain a nuclear tensor Z, a class space matrix Uclass, an intra-class factor space matrix Uwithin-class and a pixel spectral matrix Upixels; and performing classification on unclassified test hyperspectral images d by adopting a supervised classifier. According to the method, the hyperspectral images can be classified after model building, and adjustment does not need to be carried out; for other tensor modeling methods, the best classification effect can be achieved by repeated setting and adjustment of parameters; and all pixel spectrums in a class are mapped to a same coefficient vector, so that the influence of various factors is reduced to minimum, the classification precision is improved, and a result is stable.

The invention discloses a tensor decomposition method for non-equilateral video gesture identification. The method comprises the steps of representing an original video as a three-order tensor A, wherein I1, I2 and T represent the width, the height and a video frame number of a video respectively, and I1XI2 represents the size of a video frame; performing Tucker decomposition on the three-order tensor A to obtain a nuclear tensor A' reduced in dimension; calculating the nuclear tensor, and updating a result through dynamic learning until an algorithm convergence result reaches the best; repeating the processes for representing all videos as nuclear tensors same in size, namely, zooming video sequences to be same in dimension; and realizing tensor decomposition under a non-equilateral condition. According to the method, the difficulty in gesture identification is lowered, the accuracy of gesture identification is improved, good conditions are provided for subsequent video sequence classification, and the accuracy of video sequence classification can be greatly improved.

The invention discloses a gray scale video watermarking algorithm based on core tensor. The gray scale video watermarking algorithm comprises the following steps: S11, carrying out Arnold transformation on a video watermark; S12, decomposing each video tensor through Tucker and obtaining a core tensor; S13, quantifying and correcting the obtained core tensor; and S14, reconstructing the quantizedand corrected core tensor to generate a reconstructed watermark video. According to the invention, the grayscale video is represented by three orders as tensors, and the correlation between adjacent frames of the video is fully considered. By quantifying the core tensor, the watermark information is uniformly dispersed into each frame of the video, so that the non-perceptibility of the algorithm and the common video attack are improved.

The invention discloses a total variation remote sensing image stripe removing method based on tensor decomposition. A remote sensing image is decomposed into superposition of a target image, strip noise and random noise, and different regular term constraints are adopted for each component to form a target model; in order to ensure the similarity between recovery data and original data, a non-convex data fidelity term is introduced; meanwhile, a total variation regular term is introduced to constrain the piecewise smooth characteristic of the target image; and finally, low-rank Tucker decomposition and L21 norm joint constraint are adopted for strip noise, a detailed algorithm is given by adopting an alternating direction multiplier method to solve the model, a target image is recovered from an original image, and a simulation experiment result is given to verify the feasibility and effectiveness of the method.

The invention discloses a tensor trajectory path planning method based on context. The invention discloses a tensor trajectory path planning method based on context. The method comprises the steps oftensor modeling; the obtained GPS path data are processed, and a third-order tensor is obtained through tensioning, and the three dimensions represent a driver, a road segment and a time period respectively; Based on known GPS path data, extracting track data in L recent time periods and expanding the track data into a third-order tensor; Each value in the tensor s shown in the specification, andthe tensor is shown in the specification, wherein the tensor is shown in the specification; wherein in the kth time period, the travel time spent by the ith driver on the jth road is a, and L is thelast time period of statistics. The method has the beneficial effects that for Tucker decomposition, the algorithm of the method can be improved; An initial iteration value of an iteration algorithmof the system is randomly initialized, and a tensor unknown solution position is filled with a 0 value and then decomposed. And the value iterated by using the context-aware tensor method is used as the initialization of the algorithm.

The invention relates to an urban dynamic analysis method of non-negative tensor decomposition based on context awareness, tensor factor decomposition is adopted as the basis of a model, and the modelnot only can find human activity space and time modes in a city, but also can simulate the interaction among the modes; In addition, urban context information is introduced into the tensor factor decomposition model, so that the model is more accurate; The invention further provides an assembly line initialization method for analyzing the tensor sequence, so that the non-negative tensor decomposition model can analyze the long-term evolution of city dynamics. According to the method, a context-aware non-negative tensor decomposition model (cNTF) is provided, resident flow data and urban environment information are utilized, and potential modes of human activities in cities are discovered based on tensor Tucker decomposition. According to the method provided by the invention, the problemsthat an existing model cannot perform interaction between fitting modes and cannot analyze long-term evolution of a dynamic mode are solved, and meanwhile, urban context information is introduced, sothat the result is more accurate.

The invention relates to the field of signal identification, in particular relates to the field of blind source / signal separation and specifically relates to an underdetermined blind identification method based on general covariance and tensor decomposition. The underdetermined blind identification method based on general covariance and tensor decomposition comprises establishing kernel functional equation sets according to the general covariance matrix of observation hybrid sampled data, stacking the kernel functional equation sets into a three-dimensional tensor model, and finally gaining a factor matrix by virtue of tensor Tucker decomposition, thereby identifying underdetermined hybrid system characteristics or hybrid matrixes. The underdetermined blind identification method based on general covariance and tensor decomposition has the technical advantages of effective performance improvement and relatively low computation complexity in contrast with a mixed matrix identification method in traditional underdetermined blind source separation; the underdetermined blind identification method provides a technical foundation for the blind anti-jamming technique of array signal processing and satellite communication.

The invention provides a magnetoencephalogram source positioning method and device based on Tucker decomposition and a ripple time window. The device comprises a magnetoencephalogram sensor which is used for obtaining a first magnetoencephalogram signal of a user; a ripple detection unit which is used for detecting the ripple time window in the first magnetoencephalogram signal through a root-mean-square method to serve as a time window for source positioning and obtaining the first magnetoencephalogram signal in the ripple time window to serve as a second magnetoencephalogram signal; a high-order orthogonal iteration based Tucker decomposition unit which is used for carrying out Tucker decomposition on the original tensor of the second magnetoencephalogram signal to calculate an estimated value of the original tensor; and a source positioning unit which is used for calculating a covariance matrix for the estimated value and calculating a source position corresponding to the second magnetoencephalogram signal through an LCMV inverse problem solving method in a beamforming method. According to the method, the influence of noise signals is eliminated, the calculation complexity is reduced, the consistency of each calculation result is ensured, and the positioning accuracy of the epileptogenic region is improved.

The embodiment of the invention relates to the technical field of knowledge graphs, and discloses a knowledge graph embedding method, system and device based on tensor decomposition. The method comprises the steps of firstly obtaining a tensor corresponding to a current knowledge graph; carrying out Tucker decomposition on the tensor to obtain a tensor after decomposition; determining a fact scorecorresponding to the decomposed tensor; and updating the embedded vector corresponding to the preset knowledge graph embedded model according to the fact score, and performing an embedding operationof the current knowledge graph through the updated embedded vector. Obviously, the embodiment of the invention provides a better embedding mode for the N-element knowledge graph, specifically, the tensor decomposition operation is combined on the knowledge graph embedding operation, and finally, the accuracy of the embedding operation can be improved; meanwhile, the parameter complexity is greatlyreduced, and the operation efficiency is improved. Moreover, the embodiment can perform joint decomposition based on the Tucker and the tensor loop expression.

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 present invention discloses a hyperspectral image noise removing method and apparatus. The method comprises: performing Tucker decomposition on a hyperspectral image to obtain a core tensor; calculating an energy information vector of the core tensor in three dimensions; calculating an energy-to-noise ratio of the hyperspectral image in the three dimensions; calculating a tensor rank according to the energy information vector and the energy-to-noise ratio; and decomposing the hyperspectral image according to the tensor rank to remove noise.

The invention discloses a wireless sensor network high-dimensional data completion method based on Hankel transformation. The method comprises the steps of S1, modeling received wireless sensor network high-dimensional data into a tensor form chi; S2, counting effective element positions of the tensor data chi to form a tensor data support set omega; S3, performing TUCKER decomposition on the tensor data chi to obtain a kernel tensor of the tensor data chi, wherein chi = C * 1U (1) *... * iU (i)... * NU (N); S4, establishing a simple kernel tensor nuclear norm tensor completion model; and S5,based on the simple kernel tensor nuclear norm tensor completion model, carrying out Hankel transformation and variable separation, carrying out calculation filling on missing data through optimization of an alternate iteration method, and outputting a completion result. According to the method, data complementation can be rapidly and accurately carried out.

The invention provides a nuclear magnetic resonance FID signal noise suppression method based on multilinear singular value tensor decomposition, and the method comprises the steps: obtaining a multi-channel signal: enabling one channel to start sampling at an interval of delta t through delay sampling, and obtaining the multi-channel signal; converting each channel signal into a Hankel matrix, forming a third-order tensor, performing Tucker decomposition and processing on the third-order tensor to obtain a third-order tensor, recovering the third-order tensor into a multi-channel signal, performing CP tensor decomposition processing on a second-order tensor X formed by the multi-channel signal, and finally obtaining a high signal-to-noise ratio signal xnew after fusion of multiple communication signals. The method has the advantages that the limitation of a current algorithm under strong noise interference is effectively overcome, meanwhile, the instantaneity and universality of the algorithm are guaranteed, the method can be suitable for instruments such as a proton magnetometer and a nuclear magnetic resonance water exploration instrument, the frequency domain measurement precision and the hydrological parameter accuracy of the instruments are improved, and the precision and accuracy of related geosciences are effectively improved.

The invention discloses an electrocardio T-wave feature extraction method based on fractional Fourier transform and tensor decomposition, and belongs to the field of electrocardio signal processing. According to the method, electrocardio of a plurality of heart beats is continuously processed, fractional Fourier transform is adopted to perform dimension expansion on T-wave data of each heart beatto obtain a T-wave matrix, the T-wave matrix of the continuous heart beats is used to construct a third-order tensor, projection components in a third direction are obtained through Tucker decomposition, entropy values of the projection components are taken as features of electrocardio T-waves of the section, electrocardiosignal classification can be carried out through combination with machine learning, and particularly, a TWA phenomenon can be detected.

The invention discloses an abnormal resident travel mode mining method based on taxi OD data, and belongs to the field of intelligent traffic and data mining. In order to better mine taxi passenger travel rules and deeply mine abnormal modes existing in resident travel, the invention provides a method based on high-dimensional sparse tensor decomposition, namely, multi-dimensional information including time, longitude and latitude, functional area attributes and the like is organized as a tensor model, and low-rank sparse decomposition is performed on the tensor model. Therefore, the key technical problem to be solved comprises the following steps: dividing functional areas for a research area and classifying corresponding data into the corresponding functional areas; organizing corresponding data such as time, longitude and latitude, functional area attributes and the like to form a tensor model; performing low-rank sparse decomposition on the tensor model, respectively extracting a low-rank model and a sparse model, and performing Tucker decomposition; and visualizing a basis matrix obtained through decomposition, and displaying a passenger travel mode visually.

The invention discloses a hyperspectral image denoising method based on non-local low-rank tensor decomposition of a subspace, a high-dimensional hyperspectral image has a low-rank characteristic in the subspace, and for a subspace feature map of the hyperspectral image, combined low-rank constraint is carried out on the subspace feature map by using spatial non-local self-similarity and tensor Tucker decomposition. And the denoising capability of the mixed noise is further improved. The technical scheme is mainly used for solving the problem of noise interference in the hyperspectral image, and Gaussian noise and sparse noise contained in the hyperspectral image can be removed. A high-dimensional hyperspectral image has a low-rank characteristic in a subspace, and for a subspace feature map of the hyperspectral image, combined low-rank constraint is performed on the subspace feature map by using spatial non-local self-similarity and tensor Tucker decomposition, so the denoising capability of mixed noise is further improved.

The invention discloses a deep network compression method of Tucker decomposition based on dimension adaptive adjustment, which includes a dimension adaptive adjustment process and a dimension adaptive weight tensor decomposition process, and generates a new tensor by adapting to adjust the size of each dimension of the tensor Arbitrary order tensors, and then realize tensor decomposition through learnable kernel tensors and transfer matrices, thereby achieving the purpose of network optimization and compression. Compared with the existing low-rank compression method, the present invention has a larger compression factor for network parameters while maintaining network performance, and can obtain a higher compression factor; at the same time, it does not need to store the guide position of non-zero elements , does not need to record the index, and can use the storage space more effectively.

A tensor decomposition method for non-equal-length video gesture recognition, including: representing an original video as a third-order tensor where I 1 , I 2 and T represent the width, height and video frame number of a video respectively, and I 1 × I 2 Represents the size of the video frame; perform Tucker decomposition on the third-order tensor Α to obtain a nuclear tensor with reduced dimensions to solve the nuclear tensor, and update the result through dynamic learning until the algorithm convergence result reaches the optimum; repeat the above process to represent all the videos as the same The size of the kernel tensor, that is, to scale the video sequence to the same scale; realize tensor decomposition under non-equal length conditions. The invention reduces the difficulty of gesture recognition, improves the accuracy of gesture recognition, provides good conditions for subsequent video sequence classification, and can greatly improve the accuracy of video sequence classification.