303results about How to "Improve sparsity" patented technology

The invention discloses a human behavior recognition method based on an attention mechanism and a 3D convolutional neural network. According to the human behavior recognition method, a 3D convolutional neural network is constructed; and the input layer of the 3D convolutional neural network includes two channels: an original grayscale image and an attention matrix. A 3D CNN model for recognizing ahuman behavior in a video is constructed; an attention mechanism is introduced; a distance between two frames is calculated to form an attention matrix; the attention matrix and an original human behavior video sequence form double channels inputted into the constructed 3D CNN and convolution operation is carried out to carry out vital feature extraction on a visual focus area. Meanwhile, the 3DCNN structure is optimized; a Dropout layer is randomly added to the network to freeze some connection weights of the network; the ReLU activation function is employed, so that the network sparsity isimproved; problems that computing load leap and gradient disappearing due to the dimension increasing and the layer number increasing are solved; overfitting under a small data set is prevented; and the network recognition accuracy is improved and the time losses are reduced.

The invention discloses a voice secret communication system design method based on compressive sensing and information hiding, comprising the following steps: embedding secret voice into carrier voice by an embedded system to obtain mixed voice; designing a compressive sensing overcomplete dictionary aiming at the voice signal; sampling the secret voice by a compressive sensing self-adaption observation matrix to obtain a observation vector for reducing dimensions; quantizing the observation vector by an LBG (Linde-Buzo-Gray algorithm) vector, taking the quantized observation vector to serve as secret information to embed into the carrier voice, and carrying out two-stage transform on the carrier voice to obtain mixed voice; extracting the secret voice from the mixed voice by an extraction system; carrying out discrete cosine transform on mixed voice, and improving wavelet transform two-stage transform to obtain a wavelet transform coefficient; obtaining a secret bit stream by a scalar Costa decoding algorithm; obtaining a reconstructing observation vector by an LBG vector quantization decoder; reconstructing the secret voice by a compressive sensing orthogonal matching pursuit algorithm; and improving the quality of the reconstructed secret voice with a wavelet denoising method.

The invention relates to a method for re-identifying persons on the basis of deep learning encoding models. The method includes steps of firstly, encoding initial SIFT features in bottom-up modes by the aid of unsupervised RBM (restricted Boltzmann machine) networks to obtain visual dictionaries; secondly, carrying out supervised fine adjustment on integral network parameters in top-down modes; thirdly, carrying out supervised fine adjustment on the initial visual dictionaries by the aid of error back propagation and acquiring new image expression modes, namely, image deep learning representation vectors, of video images; fourthly, training linear SVM (support vector machine) classifiers by the aid of the image deep learning representation vectors so as to classify and identify pedestrians. The method has the advantages that the problems of poor effects and low robustness due to poor surveillance video quality and viewing angle and illumination difference of the traditional technologies for extracting features and the problem of high computational complexity of the traditional classifiers can be effectively solved by the aid of the method; the person target detection accuracy and the feature expression performance can be effectively improved, and the pedestrians in surveillance video can be efficiently identified.

The invention discloses a bioluminescence tomography reconstruction method. The method includes the steps: acquiring signal distribution conditions of a bioluminescent light source in a small animal body on a body surface by the aid of a bioluminescence tomography data acquisition platform and acquiring anatomical structure information in the small animal body by the aid of a Micro-CT (computed tomography) system; and constructing a system equation based on a three-order simplified spherical harmonic approximation model of a radiative transport equation and a finite-element mesh, and solving the system equation by a compressive sensing sparse regularization method to obtain accurate positioning information of the light source in the small animal body. By the aid of the bioluminescence tomography reconstruction method, three-dimensional light source distribution and the positioning information in the small animal body can be inversed through the two-dimensional light source distribution and the anatomical structure information of the small animal body surface.

The invention relates to a method for monitoring health of a rotary machine suitable for a working condition changing condition. The method comprises the steps that (1) a monitor model is constructed, wherein a relevance vector machine is used for fitting the function relation of health characteristic parameters and a working condition, the function relation is used as the parameters of a self-adaption threshold model, and the self-adaption threshold model is constructed; (2) a health state is monitored, wherein through test signals of the rotary machine to be detected, the constructed self-adaption threshold model is used for detecting whether test data exceed a threshold value or not, the machine is judged to be healthy if the test data do not exceed the threshold value, and otherwise, the machine is judged to break down. According to the method, the relevance vector machine is used for fitting the mean value of health characteristics and the function relation of standard deviation changing along with working condition parameters, and the method has the advantages that the relevance vector machine has strong learning capacity, the problems of the local minimum, the over-fitting and the under-fitting of a neural network can be solved, the relevance vector machine has better sparsity than a support vector machine, and obtained results are simpler and more practical. The method has the advantages of being high in monitor precision and capable of being used under rotating speed changing and load changing conditions.

The invention discloses a high-performance large-scale MIMO downlink transmission channel estimation method, which mainly comprises the following steps: a base station generates a large-scale beam setto cover the whole cell through beam shaping, and all antennas of the base station synchronously send pilot sequences to a target user; the base station generates a large-scale beam set to cover thewhole cell. After acquiring the pilot information, the user constructs the compressed sensing channel estimation problem according to the designed pilot matrix and dictionary. According to the block sparsity of the channel gain function after over-complete dictionary sampling, the compressed sensing recovery algorithm based on block sparsity is used to estimate the channel vector, and the estimation result is the required channel vector. The invention improves the sparsity of the channel vector under the overcomplete dictionary by using the multi-band modulation DPSS matrix, improves the performance of the channel recovery, reduces the operation complexity of the algorithm, and reduces the pilot overhead through the use of the non-orthogonal pilot matrix.

The invention provides a point cloud data sparse representation method based on compressed sensing. Under the premise of ensuring certain precision, massive point cloud data is compressed, thus the sparsity of the point cloud data is greatly raised, and a good foundation is laid for point cloud data compression and reconstruction based on compressed sensing. The method comprises the steps of (1) point cloud data normalization, (2) over-complete dictionary sparse representation based on a K-SVD algorithm, (3) normalized point cloud data observation, transmission and storage, (4) point cloud data reconstruction based on l1 norm minimization, and (5) normalized point cloud data recovery.

The invention discloses an overall situation reconstitution optimization model construction method for image block compressed sensing. The procedures of a collecting end include that firstly, an image x is divided into n B*B small blocks xi, wherein x and xi are pulled to be column vectors in a raster scanning manner; secondly, an independent identically distributed gauss random matrix phi B with the size of MB*B2 is generated; thirdly, incoherent measuring is performed on each block xi to obtain observed value vectors yi which is equal to phi B xi; fourthly, the observed value vectors yi and a seed for generating the gauss random matrix are sent to a reconstruction end. The procedures of the reconstruction end include that firstly, the received observed value vectors yi of all the blocks are accumulated to be y=[yi; y2;..., yn] in columns; secondly, an overall situation reconstitution measurement operator theta ( ) is constructed, wherein the input of the overall situation reconstitution measurement operator is an image x, the corresponding output of the overall situation reconstitution measurement operator is y, and the overall situation reconstitution measurement operator is composed of a block measurement matrix set phi and a ranking operator P ( ); thirdly, an overall optimization reconstitution model is set up, and the image is recovered with a corresponding compressed sensing reconstitution algorithm. The overall situation reconstitution optimization model construction method for image block compressed sensing can effectively eliminate the block effect in the prior art, and strong robustness on variation of the block size B is achieved.

The invention relates to a compressed sensing magnetic resonance fast imaging method, which comprises the following steps of: adopting a quadruple sampling module to sample a K space to obtain a K space undersampled signal y; obtaining a Fourier undersampled signal of a wavelet subband through the signal y, wherein the Fourier undersampled signal of the wavelet subband comprises a low-frequency subband Fourier undersampled signal uLL and a high-frequency subband Fourier undersampled signal un; rebuilding the low-frequency subband Fourier undersampled signal uLL to obtain a low-frequency subband to build a mathematic model, and rebuilding the high-frequency subband Fourier undersampled signal un to obtain a high-frequency subband N, which belongs to an element of a set (HL,LH,HH), according to the mathematic model; and carrying out wavelet inverse transformation on the low-frequency subband and the high-frequency subband N which belongs to the element of the set (HL,LH,HH) so as to obtain a rebuilt image. According to the compressed sensing magnetic resonance fast imaging method, the signal sparsity is improved through dictionary learning, and through utilizing a relationship between the wavelet subband and the K space, the traditional problem of image rebuilding in compressed sensing magnetic resonance is optimized to be a problem which has a smaller computing scale and can be solved through parallel computing, so that a final image can be rebuilt quickly.

The invention an electric-energy quality data compression and reconstruction method based on self-adaptive dictionary learning, and self-adaptive dictionary learning is used for sparse representation of electric-energy quality data. A lot of different types of electric energy quality data form a training sample set, and the completeness and redundancy of the training sample set are ensured; basic atom sparse codes which can represent the electric energy quality data is extracted from the training sample set adaptively, and a self-adaptive dictionary is obtained via iteration; a random Gauss matrix is used to carry out dimension-reducing measurement on electric energy quality test signals, and compression sampling is realized; and on the basis of compressed perception theories, the self-adaptive dictionary is used for sparse solution to obtain a sparse representation matrix of test signals, and the original signals are decoded and reconstructed. Thus, the electric energy quality data is compressed and sampled simply and reconstructed accurately, the sampling efficiency of the electric energy quality data is improved, and storage of redundant data is reduced.

The invention provides a subspace clustering method based on high-dimensional overlapping data analysis and relates to the technical field of machine learning. The method builds a weighted mixed normsubspace representation model for a data matrix that needs clustering; obtaining an optimized coefficient matrix in the weighted mixed norm subspace representation model by using a linear alternatingdirection method; establishing a similarity matrix based on the optimized coefficient matrix; dividing the similarity matrix into subspaces by using a spectral clustering algorithm to obtain an initial clustering results; establishing an overlapping probability model of the subspaces; applying the overlapping probability model to an initial subspace division result to determine the overlapping ofthe subspaces; verifying a subspace clustering result to obtain a final clustering result. The subspace clustering method based on high-dimensional overlapping data analysis can improve the density ofthe same subspace data and the sparsity of different subspace data, and improve the accuracy of clustering.

The invention discloses a method for SPEED rapid magnetic resonance imaging based on wavelet domain sparse representation. The method for SPEED rapid magnetic resonance imaging based on wavelet domain sparse representation comprises two steps of regular under-sampled data collection and SPEED image reconstruction based on wavelet domain sparse representation. The regular under-sampled data collection comprises two steps of k-space regular under-sampling and k-space central domain complete sampling. The SPEED image reconstruction based on wavelet domain sparse representation comprises seven steps of zero filling Fourier reconstruction, discrete wavelet transformation, wavelet domain double-layer sparse ghost model construction, overlapped ghost separation based on a least square error method, discrete wavelet reverse transformation, rectification summation of a plurality of ghost images and image reconstruction. The discrete wavelet transformation is used for SPEED data sparse representation, the sparseness of the image is improved through discrete wavelet transformation, and further the quality of the SPEED reconstruction image is improved.

The invention discloses an MIMO (Multiple Input and Multiple Output) radar single measurement vector DOA (Direction of Arrival) estimation method based on iterative weighted near-end projection. The method comprises the following steps of: vectorizing a covariance matrix of received data after dimensionality reduction; constructing a weighting matrix by using high-order power of a covariance inverse matrix after dimensionality reduction so as to perform proper weight constraint on a sparse vector; establishing a weighted near-end function optimization model to represent a non-convex and non-smooth sparse optimization problem in MIMO radar single measurement vector DOA estimation; and finally, obtaining a near-end operator through an SCAD (Smoothly Clipped Absolute Deviation Penalty) function in an iteration process, and projecting the near-end operator to a feasible set to solve the weighted function optimization model so as to obtain a sparse solution, and obtaining a real target DOA estimation value by searching the position of a spectral peak. Compared with a reweighted l1-SVD algorithm and a weighted SL0 (Smoothed l0norm) algorithm, the method can obtain the better DOA estimation performance, and the prior information of the number of the targets is not needed to be known in advance.

The invention belongs to the technical field of estimation of the direction of arrival of radar and discloses a low-altitude target direction-of-arrival estimating method for radar. The low-altitude target direction-of-arrival estimating method for the radar comprises the steps that an antenna array of the radar receives echo signals, the echo signals are sampled to obtain a data vector matrix, singular value decomposition is conducted on the data vector matrix to enable a corresponding dimensionality-reduction data vector and a corresponding noise subspace matrix to be generated, a guide vector matrix is obtained, a corresponding multi-path attenuation vector is obtained according to the noise subspace matrix and the guide vector matrix, a composite guide vector is constructed, a sensing matrix is constructed, a regularization parameter is obtained, and the direction of arrival is obtained by solving a corresponding mathematical model.

The invention discloses a method for extracting a brain function network of an individual based on the analysis of multiple tested brain function data. The method comprises the following steps of: calculating the tested independent components of the individual, having correspondence among the different tested, based on the tested brain function data of the individual; using a provided algorithm for analyzing the independent components with reference signals based on a multi-target function optimization framework, meanwhile, optimizing the correspondence between the tested independent components of the individual and the reference signals and the independence among the different tested components of the individual, wherein the reference signals are obtained by jointly analyzing the independent components of the tested brain function data of the individual, and can also be obtained from a brain network pattern and the like obtained through the brain network analysis or meta analysis of other modal imaging data; after the tested independent components of the individual are obtained, using a provided time sequence calculation method to calculate a time sequence corresponding to each independent component; and judging the obtained independent components to obtain a brain function network, wherein the time sequence corresponding to the independent component is an activating mode corresponding to the brain function network.

The invention discloses a wireless sensor networks WSNs signal processing method based on a sparse dictionary. Firstly, in a training stage, the sparse dictionary of WSNs signals is constructed through comprehensive adoption of a K-SVD (K-Means Singular Value Decomposition) algorithm and a DCT (Discrete Cosine Transform) initialization dictionary, thus obtaining sparse representation of the WSNs signals; then in a sampling stage, a Gauss random matrix is taken as an observing matrix to carry out CS (compressed sensing) sampling to the WSNs; and finally, based on the sparse dictionary K-SVD-DCT and the Gauss random observing matrix, approximate original WSNs signals are reconstructed through adoption of an l1 norm minimization method. The sparse dictionary K-SVD-DCT constructed in the method has good self-adaptability; the good sparse representation of the WSNs signals can be realized in the sparse dictionary K-SVD-DCT; and based on the sparse dictionary K-SVD-DCT, the WSNs signal compressed sensing reconstruction success rate is high, and the reconstruction precision is good.

The invention discloses an over-complete dictionary constructing method applicable to voice compressed sensing. The over-complete dictionary constructing method includes firstly, constructing over-complete linear prediction dictionary by a great quantity of training voices during training, subjecting voice signals to CS (compressed sensing) sampling by utilizing a random Gaussian matrix as an observation matrix in a real testing stage; and finally reconstructing the voice signals in high quality by adopting the BP algorithm on the basis of the linear prediction dictionary. Without a prediction system for testing voice, the over-complete dictionary constructing method is simple in construction, the voice signals are good in sparsity in the over-complete linear prediction dictionary, and voice compressed sensing reconstruction signals based on the over-complete linear prediction dictionary are excellent in performance and have better robustness.

The invention discloses an underdetermined blind source separation method applying multiple constraints, which is improved on the basis of a traditional blind source separation method to realize a better signal separating effect. The method disclosed by the invention aims to separate mixed signals and comprises the steps of firstly carrying out equalization and whitening to improve the robustness of initial conditions, carrying out multi-constraint limit on a non-negative matrix factorization algorithm and carrying out optimization on an objective function and further improving the signal separating property through a feedback mechanism finally. Therefore, the method disclosed by the invention has the advantages of good factor interpretability and high separated signal purity.

The invention provides an image fusion method based on multi-scale dictionary learning. The image fusion method comprises the steps of: carrying out multiscale learning to decompose each training dictionary into S subbands each of which correspondingly learns a subdictionary; and then carrying out wavelet transformation for a source image to obtain subbands of all resource images, solving sparse representation coefficients of the subbands according to the SOMP (Space Oblique Mercator Projection) algorithm and carrying out fusion, and finally carrying out inverse wavelet transformation to obtain fused images. With the method, the sparseness and the fitting degree of the image representation coefficient are improved, the detail information of the fused image is enhanced, and the fused image has an excellent fusion effect and better noise suppression ability; and the generalization ability of the dictionaries is enhanced.