47 results about "Laplacian regularization" patented technology

Embodiments of the invention provide a significant tag sorting-based image significant target detection method. The method mainly comprises the following steps of: segmenting each image in an image sample set into a plurality of image areas by using an SLIC segmentation method, and extracting visual features and background contrast features of each image area; forming a training set and a test set according to the visual features and background contrast features of each image area and significant value tags, and learning a significant value of each image area in each image by using a significant tag sorting-based algorithm; and carrying out significant image recovery on each image according to the significant value of each image area by utilizing a low-rank matrix recovery theory, and detecting a significant target in the image. According to the method provided by the invention, nuclear norms of matrixes are fully utilized to control the complexity of models, similarity of the visual features and semantic tags are combined, and the correlation between Laplacian regularization constraints on images is utilized, so that the problem that the significant tags are relatively large in space but quantity of training images is limited is effectively solved.

The invention discloses an unsupervised cluster characteristic selection method based on Laplace regularization. The unsupervised cluster characteristic selection method comprises the following steps: (1) constructing a sample characteristic matrix, (2) calculating a Laplace matrix, and (3) extracting the characteristics of the sample characteristic matrix. The unsupervised cluster characteristic selection method disclosed by the invention selects the characteristics through directly measuring the variance of follow-up study prediction results, and can directly enhance the follow-up study prediction results. Influence of the selected characteristics to predicted values of the study problems is taken into the consideration in the characteristic extraction process, so that the follow-up study efficiency can be efficiently improved. In addition, the modeling of data of the unsupervised cluster characteristic selection method disclosed by the invention is on the basis of a Laplace method of manifold geometry of the data. The unsupervised cluster characteristic selection method can efficiently reflect distribution information of the data in the space so as to calculate the maximum dimensionality of the information amount.

The invention discloses a multi-modal data analysis method and system based on high Laplacian regularization and low-rank representation and belongs to the field of multimodal data analysis. The invention aims to capture the global linear structure and nonlinear geometric structure of multimodal data. The method includes the following steps that: multi-modal data are processed, so that a pluralityof data matrices are obtained; low-rank representation and Laplacian regularization term are combined so as to construct a non-negative sparse hyper-laplacian regularization and low-rank representation model, the non-negative sparse hyper-laplacian regularization and low-rank representation model is made to learn each data matrix, so that a high Laplacian regularization and low-rank subspace is obtained; learning is performed on the basis of the high Laplacian regularization and low-rank subspace and a support vector machine, so that a plurality of classifiers can be obtained; and voting is performed for the classifiers, so that a final classifier is obtained. The structure of the system includes a data processing module, a data analysis module, a classification module, and a voting module. With the method adopted, the global linear structure and nonlinear geometry of the multimodal data can be captured.

According to the Chinese tourism field named entity identification method based on the graph convolution neural network, the graph convolution neural network comprises an input layer, an embedded layer, a graph convolution layer and a hierarchical structure, and an input body comprises a named entity and a non-entity; the method comprises the following steps: S1, simultaneously expanding to two sides by taking any non-entity of a tourism domain text as a center until a single character in a complete sentence is traversed; S2, extracting character features; S3, extracting character features; S4, inputting and training; S5, optimizing a graph convolution layer; S6, labeling all named entities in the tourism field text data, introducing a Laplace regularization loss function into the graph convolution layer so as to mine node internal structure information and extract character features; and S7, obtaining a hierarchical relationship between the named entity and the non-entity. According to the method, the character feature extraction method is constructed by using the graph convolution neural network, and semantic modeling is performed on the character features so as to realize correct identification of the named entities in the text.

The invention provides a personalized adverse drug reaction prediction method, system and device and a medium, and belongs to the technical field of biomedicine. The invention provides a multi-task learning model (KEMULA) based on multi-kernel function learning so as to replace traditional learning methods of universal application and complete individuation. More specifically, the model calculatesand ranks the risks of ADR development of patients by learning a constrained personalized ADR ranking function by assuming a sharing function of the model. This function is referred to as a personalized ADR ranking function, which is a linear combination of several scoring functions that calculate the risk of developing related ADR of patients. In addition, the model is also combined with Laplacian regularization to ensure that variable information trained by personADRank functions of similar patients is close, so that the causal relationship (true positive) of the model to the association between a given patient and the corresponding ADR can be improved. Therefore, the method has good practical application value.

The invention discloses an image feature extraction method based on Laplace operator. The method in the invention introduces Laplace regularization least square factor in the process of feature extraction, using its characteristics to take into consideration potential manifold structure in data and distinctiveness of data. Meanwhile, in the method, to extract features that minimize covariance matrix of a result set, two different measuring methods can be used--trace optimization (A-optimality) of covariance matrix and determinant optimization (D-optimality) of covariance matrix, and two different algorithms are produced correspondingly, i.e., Laplace regularization A-optimal feature selection (LapAOFS) and Laplace regularization D-optimal feature selection (LapDOFS), respectively. The advantages of the method are that distinctive structure and geometric structure are considered at the same time, a proper feature subset can be extracted, and performance of subsequent learning process can be improved to the greatest extent while potential manifold structure of data is maintained.

A three-dimensional extraction method of human face wrinkles shown in point cloud data includes the steps of simulating a generation process of human face wrinkles, setting a wrinkled area as the superposition of a smooth human face curved surface with sunken wrinkle textures, setting human face wrinkles shown in point cloud data as an object of study, remaking the smooth human face area with a computer, fitting the smooth human face area through a triplet spline curved surface fitting method, constraining triplet spline curved surface fitting, setting a smooth human face fitting process as a solving process of a regular problem, performing iterative solution through a Laplace regularization method to obtain a preliminary result epsilon 0 of human face wrinkle extraction, projecting the human face wrinkle extraction to an X-Y plane, performing binarization and 8 neighborhood communication region calculation on a projection image, setting a communication region 30% smaller than the largest communication region as noise, and removing points corresponding to the noise from the epsilon 0 to obtain a final three-dimensional extraction result of the human face wrinkles.

The invention discloses a semi-supervised classification method capable of simultaneously learning an affinity matrix and a Laplacian regularized least square, which mainly comprises the following steps: firstly, a joint model capable of simultaneously learning the affinity matrix and the Laplacian regularized least square is established according to a training sample; secondly, the block coordinate descent method is used to optimize all kinds of variables in the model; and finally, the soft label of the sample is obtained by a Laplacian regularized least square classifier, and the dimension with the largest element in a label vector is selected as the category of the sample. The invention effectively fuses the sparse self-representation problem of samples and the Laplacian regularized least square classifier, and realizes the simultaneous optimization and mutual improvement of the sample affinity matrix and the Laplacian regularized least square classifier in the learning process. Theinvention has an explicit classifier function, so that the problem of an external sample can be effectively handled. Compared with other semi-supervised classification methods, the method has more accurate classification accuracy and good application prospects.

The embodiment of the invention provides a data classification method and device based on a unified optimization target framework graph neural network. The method comprises the steps of obtaining description information of to-be-classified objects and relationship information between the to-be-classified objects; generating a feature matrix based on the description information, and generating an adjacent matrix based on the relationship information; inputting the feature matrix and the adjacent matrix into a pre-constructed and trained graph neural network to obtain a classification result of each to-be-classified object; wherein the graph neural network is constructed according to a predetermined feature propagation equation, the feature propagation equation is obtained by performing graph filter assignment on the basis of a preset optimization target equation, and the optimization target equation comprises a feature fitting constraint term and a graph Laplace regularization constraint term. A unified optimization target equation of the graph neural network is proposed, assignment is performed on the graph filter to obtain the feature propagation equation, the graph neural network is constructed according to the feature propagation equation, the to-be-classified objects are classified according to the constructed graph neural network, and the classification accuracy can be improved.

The invention discloses an image compression method based on active learning and semi-supervised learning, which selects color information of some representative pixel points in an image to learn a model which can predict colors of the remaining pixel points in the image on the basis of the active learning. The image compression method mainly includes the following steps that all pixel points of the image are provided, the image is segmented into a plurality of rectangular areas simply, one pixel point is respectively and randomly selected from each rectangular area, the pixel points are learned actively so as to select the most representative pixel points, grey level values and color values of the most representative pixel points are recorded, grey level values of the remaining pixel points of the image are only recorded, and a compression process is completed. During decompression, Laplace regularized least squares (LapRLS) serving as a semi-supervised learning algorithm is used for predicting and restoring colors of all pixel points.

The invention discloses an anchor graph structure-based semi-supervised data classification method of double Laplacian regularization. The method mainly comprises the following steps: firstly, carrying out clustering on a data set to obtain anchor point data which can approximatively indicate the entire data set, and calculating linear reconstruction weights between sample points and adjacent anchor points thereof through an FLAE method; then respectively constructing Laplacian regularization terms on the anchor points and Laplacian regularization terms on the sample points on the basis of a weight matrix between the sample points and the anchor points, and establishing an anchor graph structure-based semi-supervised classification model of double Laplacian regularization; and finally, using a zero-gradient method to parse and solve the model to obtain category soft-labels of the anchor point data, and using feature codes of unlabeled samples to linearly combine the category soft-labels of the anchor points, and discriminating categories of the unlabeled samples. Double-Laplacian-regularization constraints established by the method can better describe graph structure information among the samples, and thus realize higher classification and discriminating ability, and the method has very good application prospects.

The invention discloses a product quantification method based on semi-supervised learning, which is an improved algorithm based on a common Cartesian K-means algorithm, namely a semi-supervised Cartesian K-means algorithm. In the algorithm, a traditional minimum square loss function in a quantization step needs to be replaced with an optimal reverse prediction loss function. According to the traditional semi-supervised learning, the marked data can be directly used for model training, and different from a traditional semi-supervised learning model, the marked data can be used for model training only through Laplacian regularization.

The invention relates to a vehicle-mounted high-quality imaging method based on two-step recovery. The method comprises following steps of: S1: expressing the degenerative process of a vehicle-mounted image as the formula shown in the description, wherein f (x, y) is the original clear image, h1 (x, y) is the fixed point spread function of a vehicle-mounted environment imaging system, h2 (x, y) is a motion point spread function under the vehicle-mounted environment, n (x, y) is the additive noise, and g (x, y) is the degenerative image; S2: using a "cepstrum method" to estimate the motion point spread function h2 (x, y) of the g (x, y), and using ultra Lapras regularization ultra resolution algorithm to recover a moving fuzzy image so as to obtain g1 (x, y); and S3: through a "knife edge method", measuring the point spread function h1 (x, y) of a vehicle-mounted optical imaging system, using an improved L-R algorithm to recover the g1 (x, y) fuzzy image and finally recovering a quite clear image f1 (x, y). According to the invention, the image is recovered through two steps, so disadvantages in current methods can be well overcome and high-quality image recovery is achieved.

The invention discloses a multi-source time sequence missing data recovery method based on matrix decomposition, and aims at the multi-source time sequence missing data recovery problem, the method considers the data characteristics of a multi-source time sequence under two angles of time and a sensor on the basis of multi-source time sequence matrix decomposition, and the prior information is fully mined. For the time sequence, constraining is carried out by utilizing second-order difference regularization; for multi-sensor data, a basic principle of a graph theory is introduced, and a correlation measurement method of double Pearson coefficients is adopted to obtain a Laplace matrix representing a data relationship of each sensor. Finally, graph Laplace regularization and second-order difference regularization are fused into a matrix decomposition framework, and optimization of a target function is achieved by utilizing a gradient descent method. According to the missing data recovery method provided by the invention, data priori is fully utilized, two regularization constraint conditions are fused, and the missing data recovery method is still effective under the condition of relatively high missing rate.

The invention relates to a joint Laplacian regularization and adaptive feature learning-based water flow image clustering method. The method includes the following steps of: step 1, preprocessing: the pixel values of water flow images are expanded in columns and concatenated into column vectors, and the column vectors, adopted as elements, constitute the feature matrix X=[x1,x2, ... ,xn] of a water flow image data set, wherein xi belongs to R<d*1>; step 2: joint Laplacian regularization and adaptive feature learning-based data clustering; and step 3, water flow image clustering: the water flow image data set is divided into c clusters according to a similarity matrix S having c block diagonal structures according to flow velocity features.

The invention discloses an aluminum electrolysis superheat degree identification method based on a semi-supervised extreme learning machine, and the method comprises the steps: collecting aluminum electrolysis real-time production data, and carrying out the normalization processing and standardization processing of the collected data; Constructing a Hessian regularization operator; Calculating hidden layer output of the extreme learning machine ELM model; And constructing a loss function of the ELM model, and solving an output weight matrix. According to the method, hidden information in the label-free data can be fully explored, the problem of weak inference force of traditional Laplacian regularization is solved, the training speed is higher than that of a traditional method, and the superheat state of an industrial system can be accurately monitored in real time.

The invention discloses a zero-sample classification method based on low-rank representation and manifold regularization. The method comprises the steps of calculating a mapping relation between visual features and semantic features of samples in a visible data set; calculating semantic representations of samples in an invisible data set; introducing sparse constraints and in combination with Laplacian regularization constraints, calculating low-rank representations of the samples in the invisible data set; calculating a weight matrix and a Laplacian matrix; introducing the manifold regularization, and removing noises of the semantic representations in the invisible data set; and predicting labels of the samples in the invisible data set, thereby realizing sample classification. Accordingto the zero-sample classification method based on the low-rank expression and the manifold regularization, the classification method effectively overcomes the limitation of low classification precision under the conditions of few samples, sample label information loss and the like in a traditional classification method; the more accurate semantic representations in the invisible data set are obtained; the description capability of data features is enhanced; and the precision of zero-sample classification can be effectively improved.

The invention relates to a prior knowledge fault diagnosis method based on the Tennessee Eastman process. The method comprises the steps that the offline historical data of the Tennessee Eastman process are acquired; an adjustment parameter matrix that U belongs to R<nxn> and k of a KNN algorithm are selected; an adjacent matrix W is constructed on an existing weighted undirected graph, a matrix D is accordingly calculated, a Laplacian matrix L=D-W is defined, and the Laplace regular term L<~> is calculated according to a Laplace regularization algorithm; the local regular term (I-A)<T>(I-A) is calculated according to a local regularization algorithm; a tag matrix is calculated according to F<*>=(UD<~>+L<~>+(I-A)<T>(I-A))<-1>UD<~>Y; and the unmarked samples are marked according to f=arg maxF<*><ij>, 1<=j<=c, and fault classification information of the industrial process is obtained after normalization. Characteristic information of the marked samples and the unmarked samples is fully mined and utilized to establish a fault diagnosis model and verification is performed by using the Tennessee Eastman process data, and the classifier is improved in the final classification phase so that the classification accuracy can be enhanced, and the classification error rate of the samples and the sample separation degree and other verification standards can be improved.