159 results about "Nonnegative matrix factorisation" patented technology

The invention discloses a face identification method based on non-negative matrix factorization and a plurality of distance functions. The method comprises the following steps of: extracting face image characteristics, performing the non-negative matrix factorization on all face images in a learning image library to obtain a corresponding basic image and a weight vector, namely a characteristic vector corresponding to each learning image; performing characteristic extraction on all of test images by utilizing a non-negative matrix factorization algorithm to obtain a characteristic vector corresponding to each test image; calculating a mean characteristic vector Hm corresponding to each type of training sample image set with known identity by utilizing the obtained characteristic vectors; and calculating similarity between each test image and the mean vector corresponding to each type of training sample image set, obtaining a quantized numerical value of the similarity by combining the different distance functions, finding a training image type which is proximate to the test image, namely the nearest neighbor point, and classifying the test image into a type with the nearest neighbor point according to a nearest neighbor classification method to finish the identification of all the test images.

The invention discloses a self-adaptive hyperspectral image unmixing method based on region segmentation. In consideration of coexistence of linear mixing and bilinear mixing, the method is implemented by adopting the following steps: inputting a hyperspectral image; estimating the number of end elements with a minimum error based hyperspectral signal recognition method; extracting end element matrixes with a vertex component analysis algorithm; clustering hyperspectral data with a K-means clustering method, and segmenting the image into a homogeneous region and a detail region; adopting a linear model for the homogeneous region and performing unmixing with a sparse-constrained non-negative matrix factorization method, and adopting a generalized bilinear model for the detail region and performing unmixing with a sparse-constrained semi-non-negative matrix factorization method. According to the method, characteristics of the hyperspectral data and abundance are combined, the hyperspectral image is represented more accurately, and the unmixing accuracy rate is increased. The sparse constraint condition is added to the abundance, the defect of high probability of local minimum limitation of the semi-non-negative matrix factorization method is overcome, more accurate abundance is obtained, and the method is applied to ground-object recognition for the hyperspectral image.

The invention belongs to the technical field of statistical model identification and machine learning, and in particular discloses a nonnegative matrix factorization-based dimensionality reducing method used for clustering. The method comprises the following steps of: adopting a KL distance; adding a data normalization constraint; directly discovering the internal relation among data dimensionalities by minimizing an object error function between data compression and reconstruction; obtaining a mapping matrix; and projecting high-dimensional data to a low-dimensional subspace by using the mapping matrix so as to perform effective data analysis such as clustering and the like. A simpler iterative formula compared with an original factorization method is obtained by the method; and normalization can be maintained naturally in each iterative updating. The normalization makes a final mapping matrix have higher sparsity compared with the original factorization method. In the obtained low-dimensional space, a clustering result shows that a more effective low-dimensional data characteristic can be obtained by the method; and an algorithm is simple and effective.

A process discovery system that includes an offline system training module configured to cluster similar process log traces using Non-negative Matrix Factorization (NMF) with each cluster representing a process model, and learn a Conditional Random Field (CRF) model for each process model and an online system usage module configured to decode new incoming log traces and construct a process graph in which transitions are shown or hidden according to a tuning parameter.

The invention discloses an image retrieval method based on sparse non-negative matrix decomposition. It includes the following steps: 1) Query and extract the image and accompanying text of the retrieval results under two different image data sources; 2) Extract the tags in the accompanying text, and filter the results according to word frequency to form a vocabulary; 3) For each Image set, using the association relationship between labels and images to form an association matrix between labels and images; 4) Using sparse non-negative matrix decomposition to analyze the association matrix obtained in step 3), to obtain the shared subspaces of different sources of data and their corresponding independent subspaces Subspace; 5) The user sends a retrieval request for images on a data source, forms a query vector and maps it to the corresponding subspace of the data source, calculates the similarity with all images and sorts them, and returns the top N most similar images. The present invention makes full use of the associated knowledge of tags and images under multiple data sources, performs migration learning through sparse non-negative matrix decomposition, and improves the accuracy of image retrieval on target data sources.

The invention provides a non-negative matrix factorization face recognition method and system based on kernel machine learning, which comprises five steps. The invention has the following beneficial effects: the invention avoids the learning of the inaccurate pre-image matrix by directly learning two kernel matrices, Kwx and Kww, and avoids the derivation of the kernel function in the iterative formula by changing the learning object, so that there is no limit to the selection of kernel function and a general algorithm for any kernel function is obtained.

The invention belongs to the technical field of image processing, and particularly discloses an SAR image target recognition method based on non-negative matrix factorization of sparse constraint. According to the method, more effective features are extracted by improving a non-negative matrix factorization method to improved recognition precision, and the problems that features extracted in the prior art are not typical and recognition precision is not high are solved. According to the method, a training sample image and a test sample image are pre-processed in the same way, logarithmic transformation is carried out, a pre-processed training sample set is factorized through non-negative matrix factorization of sparse constraint to obtain a basis matrix and a coefficient matrix, a test sample set is projected in a sub space of a basis matrix structure, classification is carried out through an SVM after a feature matrix is obtained, and then classification accuracy is obtained ultimately. Compared with the prior art, the extracted features are more effective and recognition precision can be effectively improved.

The invention provides a feedback density peak value clustering method and a system thereof. A problem that one type is divided into a plurality of types when a plurality of density peak values are generated in one type in an original density peak value algorithm is solved. Simultaneously, accuracy of an original algorithm in a high dimension data set is increased. The method comprises the following steps of 1, using non-negative matrix factorization to carry out characteristic extraction on a data set; 2, according to an original density peak value clustering algorithm, drawing a decision graph and selecting a plurality of clustering centers; 3, using a ''nearest neighbor'' algorithm to distribute the rest of points and removing a noise point; 4, using a SVM to feed back a clustering result between two types; and 5, according to a feedback result, merging the types which can be merged. By using the method, robustness of a density peak value algorithm can be effectively increased, clusters with any shapes can be well discovered, high dimension data can be effectively processed, and a good clustering effect is possessed.

Supervised nonnegative matrix factorization (SNMF) generates a descriptive part-based representation of data, based on the concept of nonnegative matrix factorization (NMF) aided by the discriminative concept of graph embedding. An iterative procedure that optimizes suggested formulation based on Pareto optimization is presented. The present formulation removes any dependence on combined optimization schemes. Analytical and empirical evidence is presented to show that SNMF has advantages over popular subspace learning techniques as well as current state-of-the-art techniques.

The invention discloses an off-line handwritten signature recognition method based on non-negative matrix factorization (NMF). The method comprises the needed steps that 1, signature data are collected, all signature images are preprocessed, and a data matrix A of all preprocessed training set signature images and a data matrix B of testing set signature images needing recognizing are obtained; 2, features of the training set signature images are extracted, wherein NMF is performed on the A, and a corresponding base image matrix W and a feature matrix H formed by weight vectors which are namely the feature vectors and correspond to all the signature images are obtained; 3, features of the testing set signature images are extracted, wherein the data matrix B of the testing set signature images are projected on a base image, and a projection coefficient matrix H' is obtained; 4, training and classification decision are carried out through a support vector machine, and whether the signature is true or false is judged.

Non-negative matrix factorization, NMF, is combined with identification of a maximum margin classifier by minimizing a cost function that contains a generative component and the discriminative component. The relative weighting between the generative component and the discriminative component are adjusting during subsequent iterations such that initially, when confidence is low, the generative model is favored. But as the iterations proceed, confidence increases and the weight of the discriminative component is steadily increased until it is of equal weight as the generative model. Preferably, the cost function to be minimized is:minF,G≥0X-FG2+γ(w2+C∑i=1nL(yi,w·gi+b)).

The invention discloses a non-negative matrix factorization-based heart and lung sound separation method. The method comprises the steps of firstly filtering a mixed signal of heart and lung sounds in a time domain to obtain a priori heart sound signal and a priori lung sound signal; secondly performing short-time Fourier transform and sparse non-negative matrix factorization on the priori heart sound signal and the priori lung sound signal; and finally obtaining a basis matrix of the mixed signal. The sparse non-negative matrix factorization process is equivalent to a learning process of the basis matrix, and has a clustering effect; the basis matrix can be mined effectively from the priori heart sound signal and the priori lung sound signal; and a heart sound signal and a lung sound signal are accurately separated from the mixed signal of the heart and lung sounds in the time domain according to the basis matrix in the subsequent steps. The invention furthermore discloses a non-negative matrix factorization-based heart and lung sound separation apparatus, which has the abovementioned beneficial effects.

An abnormal sound detection device includes a first learning unit that inputs a normal operating sound of the machine equipment, and registers a normal sound base spectrum from an amplitude spectrogram, a new sound / new vibration extraction unit that executes supervised nonnegative matrix factorization (NMF) using the normal sound base spectrum as teacher data on an acoustic signal input during diagnosis, and outputs a nonnegative matrix not allowed to be approximated with a low rank in the normal sound base spectrum, a second learning unit that extracts a new sound / new vibration component from the normal operating sound of the machine equipment, learns a normal sound model from the new sound / new vibration component, and registers the normal sound model, and an abnormality detection unit that computes a probability that the new sound / new vibration component extracted from the acoustic signal of the machine equipment during diagnosis is generated from the normal sound model.

The invention relates to a non-negative matrix factorization and convolutional neural network-based color image processing method. The method comprises a step of obtaining a color image by acquisitionequipment and preprocessing the color image, wherein the preprocessing comprises the following steps of: extracting image features by adoption of a non-smooth non-negative matrix factorization algorithm; and combining a convolutional neural network to optimize a color image classification and reconstruction process: firstly reading an image data matrix to obtain a row number and a column number,initializing corresponding parameters, and combining the convolutional neural network to complete image classification and reconstruction. In color image processing, the image processing method is more accurate in feature extraction, few in reconstruction errors and high in classification efficiency.

The problem that in the prior art, the face recognition accuracy is low is solved. The invention provides an adaptive graph regularization non-negative matrix factorization method for face recognition. The method comprises the following steps: vectorizing a face picture; decomposing the data matrix X into a product of a base matrix and a coefficient matrix; adding a non-negative constraint to eachelement in the basis matrix and the coefficient matrix; constraining a coefficient matrix of the data matrix X; obtaining a data matrix X, obtaining a processed similarity matrix, processing the similarity matrix S by utilizing a diagonal matrix and a Laplace matrix, introducing a weight constraint item Lambda to obtain an adaptive graph, and finally introducing the adaptive graph as a regularization constraint item into a non-negative matrix factorization objective function of the data matrix X to identify the face photo. According to the method, the weight constraint term is introduced, theweight matrix is not kept unchanged any more, a large amount of noise and outliers are removed, local neighborhood structure information is complete, and the recognition accuracy is improved.

The invention discloses a point cloud level fusion method of laser radar data and a hyperspectral image. The method mainly aims at heterogeneous remote sensing data, namely the laser radar data and hyperspectral image data, and is characterized in that firstly unmixing is conducted on the hyperspectral data on the basis of a non-negative matrix decomposition framework added with regular terms to obtain a hyperspectral image data abundance matrix and an end member matrix; according to abundance information of a final abundance matrix of the hyperspectral image data, an abundance matrix of the laser radar data is obtained through a bilinear interpolation method; and finally the abundance matrix of the laser radar data is fused with a final hyperspectral end member matrix to obtain a hyperspectral point cloud. According to the method, the hyperspectral high-resolution spectral information and the small-spot laser radar elevation information are fused at the point cloud level, and the capabilities of carrying out spectral classification and 3D structure segmentation are achieved at the same time.

The invention discloses a method for recognizing upper limb and hand rehabilitation training action of a stroke patient. The method comprises the following steps: performing blind source separation onelectromyographic signal data by using a non-negative matrix factorization model, and removing non-stationary muscle activation information to obtain a stable time-varying blind source separation result; performing further pattern recognition by using the factorized time-varying blind source separation result data to improve the stability and accuracy of recognition; and making learning featuresmaintain time and space characteristics at the same time through a CNN-RNN model. The CNN-RNN model does not require manual data feature extraction and screening, directly processes the data, automatically extracts the features and completes classification recognition, can realize end-to-end rehabilitation training action recognition analysis, and is combined with an attention layer for attentionweighting of a hidden state of a second layer in a two-layer two-way GRU layer to give a greater weight to data with a greater contribution degree, so that the data can play a greater role, and the accuracy of classification recognition is further improved.

The invention relates to a mixed pixel unmixing method and system of a space artificial target linear model. Imaging spectral data of a space artificial target are acquired. Superpixel segmentation iscarried out on collected imaging data. Spatial information and spectral information of the imaging spectral data are fully combined, and a superpixel group sparse non-negative matrix mixed pixel unmixing method is introduced into standard non-negative matrix decomposition. The method is based on a linear hybrid model. A space group sparse constraint term and an abundance quadratic function constraint term are introduced into a standard non-negative matrix. The space structure information of the sample data is considered in the space group sparse constraint item. The quadratic function constraint term of abundance better describes the distribution rule of the surface material forming the space artificial target, so that the unmixing precision of the mixed pixel is improved. The existence of a pure pixel does not need to be assumed. The end member and abundance of the mixed pixel of the space artificial target can be obtained at the same time, and the robustness to noise is better.

The invention provides a voice enhancement method and device based on multi-frame spectrum and non-negative matrix decomposition and belongs to the voice enhancement and non-negative matrix decomposition field. The method comprises steps that pure voice, noise and noise-contained voice are pre-processed to acquire short-time spectrum which is converted into multi-frame spectrum; the multi-frame spectrum of the noise and the pure voice is converted into products of corresponding base matrixes and corresponding coefficient matrixes, and a base matrix of the multi-frame spectrum of the noise and a base matrix of the multi-frame spectrum of the pure voice are solved; the two base matrixes are synthesized to form a base matrix of the multi-frame spectrum of the noise-contained voice, the multi-frame spectrum of the noise-contained voice is converted into a product of a base matrix and a coefficient matrix, a coefficient matrix of the multi-frame spectrum of the noise-contained voice is acquired, and an initial estimate of the multi-frame spectrum of the noise and enhanced voice is acquired; through a Wiener filtering method, the multi-frame spectrum of the enhanced voice is acquired and is transformed into a time domain signal, and enhancement voice is lastly acquired. The method is advantaged in that the special voice information is kept, the voice is better reduced, and the voice enhancement effect is improved.

The invention proposes an NMF image fusion-based SAR (Synthetic Aperture Radar) image change detection method, which is used for solving the technical problems of low accuracy and indefinite physical significance in an existing image fusion-based SAR image change detection method. The method is implemented by the following steps of inputting two original SAR images same in region and different in time, and constructing two basic difference graphs Xl and Xm after de-noising processing; performing neighborhood feature description on Xl and Xm to obtain feature description matrixes XL and XM; constructing a change detection feature matrix X by utilizing the two feature description matrixes; performing NMF non negative matrix decomposition on X to obtain a decomposed coefficient matrix H; obtaining a post-fusion difference graph by utilizing the coefficient matrix; and clustering the post-fusion difference graph into two classes by utilizing a clustering algorithm to obtain a change detection result. According to the method, detail information of the original images is better utilized; and the image fusion method has a more remarkable physical significance and can be used for analysis and labeling of SAR image change information.

The invention relates to the field of network structure deanonymization, in particular to a network structure deanonymization system and method based on matrix decomposition, and the method comprisesthe steps: taking a data source which comprises a topological structure of a relation between networks and a topological structure of a relation between users in the networks; dividing the data set ofthe dynamic social network into T static network data sets according to the time, and obtaining an adjacent matrix corresponding to each piece of static network structure information; inputting the adjacent matrix, performing training according to a low-rank representation model, and calculating and obtaining a target matrix; inputting the target matrix into an improved non-negative matrix decomposition model for training and prediction to obtain a de-anonymized matrix of the target matrix. According to the method, low-rank training is carried out on network data, a prediction model of non-negative matrix factorization is combined, structure information and time sequence evolution of the network are considered in a combined mode, and the prediction accuracy is improved.

The invention provides a multi-view clustering method based on multi-manifold dual graph regularized non-negative matrix factorization. The multi-view clustering method comprises the following steps of: S10, acquiring views to be clustered; S20, constructing an adjacency matrix of a data graph and an adjacency matrix of a feature graph for each view to be clustered; S30, acquiring a target function of multi-manifold dual graph regularized non-negative matrix factorization through a consistency coefficient and multi-view local embedding; S40, conducting iterating a preset number of times by using an iterative weighting method according to the target function, and updating the adjacency matrix of the data graph of each view to be clustered, the adjacency matrix of the feature graph of each view to be clustered and graph regular terms to obtain a feature matrix of each view to be clustered; and S50, analyzing the feature matrix of each view to be clustered by using a k-means clustering algorithm to realize multi-view clustering. Compared with a traditional multi-view clustering method, the clustering method has the advantages that structural information and features contained in viewdata are more effectively utilized, clustering effect is greatly improved, and better clustering performance is brought.

The invention discloses a post-stroke upper limb intermuscular collaborative coupling analysis method. The method comprises the following steps: firstly, synchronously acquiring and preprocessing multi-channel surface electromyogram signals in an upper limb movement process; then carrying out non-negative matrix factorization on the preprocessed data; presetting the number of decomposition layers,multiplying two decomposed matrixes to obtain a reconstructed data matrix, and calculating an explanatory variance; judging a cooperative and non-cooperative relationship; and finally, calculating the intermuscular coupling strength of the coherence measurement in different frequency bands. From the perspective of motion generation and execution, the coupling strength difference of collaborativemuscles in the muscle collaborative model is discussed, and a physiological basis can be provided for the motion rehabilitation process of stroke patients.

The invention discloses a multi-view data missing complementing method for multi-manifold regularization non-negative matrix factorization, which comprises the following steps of: obtaining manifold and global clustering in which unmarked multi-view data tends to be consistent by utilizing a multi-manifold regularization non-negative matrix factorization algorithm through consistency assumption among multiple views; and constructing a multi-view collaborative discrimination model by adopting a view collaborative improved Gaussian mixture method. Pre-calibration of a cluster to which a sample belongs is realized by calculating the cluster relevancy level of the sample with missing data under a non-missing view angle; and establishing a missing data prediction model under a specific view angle by utilizing the consistency of multiple view angles in a low-dimensional space and adopting a multiple linear regression analysis method, thereby realizing accurate data completion under a multi-attribute missing condition. According to the method, large-scale labeled samples are not needed for training, the pre-defined category relation and related characteristics are avoided, and the understanding and discovering capacity of an existing multi-view mining technology for unlabeled multi-source data is improved.

The invention discloses a data-to-text generation method based on fine-grained topic modeling. The data-to-text generation method comprises the following steps: learning semantic representation of each data record on a coding layer based on a bidirectional long-short-term memory network; learning topic distribution corresponding to each data record and word distribution corresponding to each topicbased on a non-negative matrix factorization method to obtain a topic word table corresponding to each data record; based on the semantic representation of each data record in a decoding layer, carrying out text generation by utilizing a long-term and short-term memory network, an attention mechanism and fine-grained topic representation in combination with a topic word table; and performing model training to obtain an optimal text generation result. According to the method, topic distribution of data and word distribution corresponding to topics are mined by utilizing a non-negative matrix factorization method, so that topic consistency between a generated text and a data table is restrained, and a model is guided to learn a more accurate word use mode; a copying mechanism is introducedin the text generation process, and it is guaranteed that the model can accurately generate numerical description.

A correlation analysis method for identifying Alzheimer's disease related biomarkers comprises the following steps: based on sparse non-negative matrix factorization of joint connectivity, projecting structural magnetic resonance imaging (SMRI), single nucleotide polymorphism (SNP) and gene expression data into a public feature space at the same time; and also adding the connectivity information and genetic data of each region of the brain as priori knowledge to identify a region of interest ROI, risk SNP sites and risk genes associated with Alzheimer's disease patients and early cognitive impairment patients. The method has higher correlation analysis capability and better anti-noise performance and biological interpretation performance.

The invention relates to a tumor genome copy number variation characteristic pattern recognition method and application thereof, and belongs to the technical field of gene diagnosis. According to themethod, systematic classification is carried out on tumor genome copy number variation characteristics, 80 numerical values of eight types of tumor genome copy number variation characteristics are specifically extracted, a copy number variation characteristic mode of a tumor sample is extracted by utilizing a non-negative matrix factorization algorithm, and then the tumor is typed according to thecopy number variation characteristic mode. Therefore, accurate prediction of clinical prognosis of tumors can be realized.

The invention discloses a micro-expression recognition method and device. The method comprises the following steps: solving a climax frame of a preset micro-expression video frame by combining an optical flow method and a dichotomy; obtaining N specific regions in each frame of the high tide feature graph set, wherein the high tide feature graph set is obtained by adding high tide frames and preset frames; obtaining an M-dimensional feature vector of each specific region through an optical flow parameter of each pixel obtained according to an optical flow method; inputting all the obtained feature vectors into a preset GRU model for feature extraction to obtain high tide dynamic features; calculating high tide static characteristics of the high tide frame according to a preset basis matrixand a local non-negative matrix factorization method; and fusing the high tide dynamic features and the high tide static features, and inputting the fused features into a preset classifier for classification to obtain a classification result. According to the method and the device, the technical problem of unreliable identification result caused by the fact that the existing micro-expression identification technology cannot completely obtain the characteristics of each layer of the micro-expression and has redundant information irrelevant to the expression is solved.