46 results about "Subspace model" patented technology

The invention provides a method for detecting and tracing three-dimension face action based on video stream. The method include steps as follows: detecting face and a key point position on the face; initializing the three-dimension deformable face gridding module and face texture module used for tracing; processing real time, continuous trace to the face position, gesture and face action in follow video image by using image registering with two modules; processing evaluation to the result of detection, location and tracking by using a PCl face sub-space, if finding the trace is interrupted, adopting measure to restore trace automatically. The method does not need to train special user, has wide head gesture tracing range and accurate face action detail, and has certain robustness to illumination and shelter. The method has more utility value and wide application prospect in the field, such as human-computer interaction, expression analysis, game amusement.

The invention discloses a local spline embedding-based orthogonal semi-monitoring subspace image classification method. The method comprises the following steps of: 1) selecting n samples serving as training sets and the balance serving as testing sets from image data sets, wherein the training sets comprise marked data and unmarked data; 2) building an extra-class divergence matrix and an intra-class divergence matrix by using the marked data; (3) training data characteristic space distribution by using a whole and building a Laplacian matrix in a local spline embedding mode; 4) according to a local spline, embedding an orthogonal semi-monitoring subspace model, and searching a projection matrix to perform dimensionality reduction on the original high dimension characteristic; 5) building a classifier for the training samples after the dimensionality reduction by using a support vector machine; and 6) performing the dimensionality reduction on the testing sets by using the projection matrix and classifying the testing sets after the dimensionality reduction by using the classifier. In the method, the information, such as image sample marking, characteristic space distribution and the like, is fully utilized; potential semantic relevance among image data can be found out; and image semantics can be analyzed and expressed better.

The invention proposes an end-to-end three-dimensional human face reconstruction method based on a depth neural network, and the method mainly comprises the contents: a 3D face shape subspace model, the depth neural network (DNN) configuration and end-to-end training. The process comprises the steps: employing the 3D face shape subspace model, and enabling a 3D face to serve as a linear combination of shapes and mixed-shape base lines; adding a sub-convolution neural network (fusion CNN) to a face model based on a VGG network for regression of expression parameters; adding a multitask learning loss function for identity parameter prediction and expression parameter prediction, wherein the input of the DNN in the end-to-end training is a two-dimensional image, and the output consists of an identity parameter vector and an expression parameter vector. According to the invention, the method solves a problem of impact caused by the posture, expression and illumination changes in a face image, and avoids the loss of the depth information in a process of image collection. Meanwhile, the method simplifies the frame, reduces the calculation cost, and improves the reconstruction precision and the recognition robustness.

The invention discloses a layered integrated Gaussian process regression soft measurement modeling method used for a complex changeable multi-stage chemical process. The layered integrated Gaussian process regression soft measurement modeling method is an on-line multi-model strategy. A Gaussian mixture model is employed to identify different stages of the process, principal component analysis is carried out on data in each stage, on the basis of the contribution degree of each auxiliary variable in the principal element space, data in each mode is divided into several subspaces, and a corresponding Gaussian process regression soft measurement model is established. When new data comes around, variable selection is carried out by means of subspace PCA, and on the basis of the soft measurement model which is established off line, the prediction output of each model can be obtained. By carrying out mean value fusion on outputs of subspace models, first layer integrated output, i.e., local prediction output in each mode can be obtained, finally new data obtained according to calculation is attached to the posterior probability of each different stage, and local prediction in each mode is fused by means of the posterior probability to obtain second layer integrated output. Key variables can be accurately predicted, and therefore the product quality is improved, and the production cost is reduced.

A new method is developed to accelerate high-resolution magnetic resonance spectroscopic imaging (MRSI). The method is built on a low-dimensional subspace model exploiting the partial separability of high-dimensional MRSI signals and uses this subspace model for data acquisition, processing, and image reconstruction. Specifically for two and three dimensional MRSI with one spectral dimension, this method sparsely samples the corresponding (k,t)-space in two complementary data sets, one with dense temporal sampling and high signal-to-noise ratio but limited k-space coverage and the other with sparse temporal sampling but extended k-space coverage. The reconstruction is then done by estimating a set of temporal/spectral basis functions and the corresponding spatial coefficients from these two data sets. The proposed subspace model can be further extended to incorporate multiple signal components for nuisance signal removal in ¹H-MRSI and more generalized reconstruction methods. The resulting imaging technique can be used for high-resolution MRSI of different nuclei. It will be useful for high-resolution metabolic imaging with many exciting applications.

The invention relates to an online target tracking method based on increment bilateral two-dimensional principal component analysis (Bi-2DPCA) learning and sparse representation, and provides an increment Bi-2DPCA learning algorithm capable of rapidly and accurately updating a target sub-space model to reflect an appearance change of a target in a tracking process. Aiming at the problem that the target is frequently blocked and polluted by noise in the tracking process to cause that the tracking effect becomes worse, according to the method disclosed by the invention, a Bi-2DPCA-based sub-space model is embedded below a sparse representation frame, so that interferences to target positioning and target sub-space model updating caused by blocking and noise are furthest removed. Meanwhile, a novel method for calculating visual similarity is used. By adopting the method, energy distribution of Bi-2DPCA when an image is presented is considered, and is more accurate in comparison with a classical reconstruction error; tracking is achieved under a Bayesian inference framework; the target state is estimated by using a particle filtering algorithm.

Described is a system for anomaly detection to detect an anomalous object in an image, such as a concealed object beneath a person's clothing. The system is configured to generate a subspace model for a normal class using training images. The normal class represents normal objects in a common class. The system receives a novel image having an object in the common class. A set of geometric landmarks are identified in the object in the novel image for use in registering the image. The novel image is registered by warping the image so that the geometric landmarks coincide in the novel image and the training images, resulting in a warped novel image having an object. Thereafter, the system determines if the object in the warped novel image is anomalous by measuring the distance of the warped novel image from the subspace model. Finally, if anomalous, an operator is notified accordingly.

The invention discloses an industrial data driving prediction control method based on subspace identification. The industrial data driving prediction control method includes steps of (1) calculating the predication model output and obtaining the predication output expressed by the predication increment according to the subspace model identification; (2) for the nonlinear time-varying characteristic, adopting the adaptive prediction control method of an online recursion identification model and updating the predication error of the before-after predication output and the process output through comparison to determine whether the predication model needs to be updated; and 3) performing constraint processing and for the physical constraints in the system, solving the problem by means of standard secondary programming and reducing the solution computational complexity by means of Lagrange daily functions. The industrial data driving prediction control method has the advantages of good practicality, high control precision, and simplified calculation.

The invention discloses a dynamic function connection local linear embedded feature extraction and brain state classification method and system. The dynamic function connection local linear embedded feature extraction method comprises the following steps: collecting functional magnetic resonance imaging data in a resting state; after preprocessing, extracting an average time sequence signal of each brain region through a brain template; calculating and constructing a dynamic function connection matrix by using a sliding time window, and taking the dynamic function connection matrix as to-be-processed high-dimensional brain dynamic description original data; and carrying out manifold learning on the dynamic function connection matrix by using a local linear embedding algorithm to obtain a low-dimensional manifold subspace model, and extracting a feature part in the low-dimensional manifold subspace model to obtain a dynamic function connection local linear embedding feature. For the dynamic function connection local linear embedded feature extraction method, the feature extraction method is rapid in calculation and ideal in data processing effect, can construct significant crowd feature description, does not depend on the absolute value of the amplitude of an imaging signal, is migratable between different MRI machines, is excellent in classification and discrimination performance, and can conveniently utilize a machine learning model to realize brain state classification.

The invention relates to a method for establishing an OPC model and an optical proximity correction method for user target graphs, wherein the method for establishing the OPC model comprises the following steps: providing a test graph having several sprite graphics and an initial OPC model, wherein the sprite graphics can be divided into several fragments, establishing the initial OPC model according to an optical system parameter and a lithography parameter; simulating the test graph according to the initial OPC model to obtain the space light intensity parameter of each fragment; grouping the fragments according to the distribution condition of the space light intensity parameter of each fragment; respectively establishing a corresponding sub-space model according to space light intensity parameter of the fragment corresponding to each group, by aiming at different space light intensity parameters of the sprite graphics in the test graph, respectively employing the different sub-space models for performing optical proximity correction to effectively increase the subsequent formed graphics quality on wafer.

The invention discloses a WAMS-based low-frequency oscillation decentralized controller design method considering interaction which can be applied to low-frequency oscillation control of a power system. The method comprises the steps of: acquiring modal information by integrating multiple identification results; identifying controllability and observability by adopting a state subspace model identification method (N4SID), and determining installation positions of controllers and candidate feedback signals; transferring functional matrix definition according to MIMO for identification and solution row by row, calculating a highly-controllable low-interaction control loop combination by utilizing a branch and bound method, and converting a complex multi-controller coordination design problem into a simple decentralized controller independent design problem; designing a PSS for a region pattern by adopting an identification-based pole placement method, and designing the controllers for an interval pattern by adopting a model prediction control method. The good control effects of the two types of independently designed controllers verify effectiveness of the WAMS-based low-frequency oscillation decentralized controller design method. The WAMS-based low-frequency oscillation decentralized controller design method provides a new idea for designing a multi-damping controller based on identification coordination.

The invention relates to an extended target sensitivity detector and a system during subspace signal mismatch, and the method comprises the steps: firstly constructing a to-be-detected data matrix, asignal matrix and a training sample matrix, then constructing a sampling covariance matrix through the training sample matrix, and then constructing a whitening matrix through the sampling covariancematrix; whitening the to-be-detected data matrix and the signal matrix by utilizing the whitening matrix, constructing a detection statistic by utilizing the whitened data matrix, determining a detection threshold by utilizing the detection statistic and a false alarm probability, finally comparing the detection statistic with the detection threshold, and judging whether a target exists or not. Onthe basis of the self-adaptive detection idea, mismatch sensitive target detection under signal mismatch is achieved, the method is suitable for extended target detection based on a subspace model, and the detector has the constant false alarm characteristic and does not need independent constant false alarm processing.

The invention discloses a sparse subspace clustering algorithm based on semi-supervision. The sparse subspace clustering algorithm comprises the steps that data prior information is converted into a constraint matrix suitable for a sparse subspace model in the form of point pair constraint; interference of flag-free bits is eliminated in the form of Hadamard product, the state of the coefficient represented by different constraint conditions is also considered and corresponding constraint terms are established; and a semi-supervised sparse subspace model of two hard threshold and soft threshold forms is established by using the constraint terms, and a semi-supervised framework is accordingly established on the sparse subspace clustering algorithm. The clustering accuracy of the sparse subspace algorithm can still be maintained by the algorithm without prior information. Meanwhile, the performance advantages of the sparse subspace clustering algorithm are also absorbed so that the high-dimensional clustering problem containing interference information data can be directly and effectively processed, the clustering performance is ensured to be effectively enhanced under the condition of less known prior information and thus the algorithm applicability can be increased.

The invention relates to an SDN-based flow detection method. The method comprises the following steps: obtaining the network flow detection rule; obtaining attribute information of the message of a first data stream, judging whether the attribute information is matched with the obtained network flow detection rule, if the attribute information is matched with the safety rule, executing the next step, if the attribute information is not matched with the safety rule, determining that the network flow is normal; building a sub-space model, and computing a flow threshold value E according to the sub-space model; accessing the information item of the OpenFlow exchanger flow list through an SDN network controller, and obtaining the communication flow Q arriving the exchanger; and comparing the flow threshold value E with the communication flow Q, if E is smaller than or equal to Q, presenting that the network flow is normal, otherwise the network flow is abnormal. The SDN-based flow detection method is capable of detecting the safety of the internal flow of the network, and effectively improving the detection accuracy for DDoS attack generated in the SDN network.

The invention discloses a feature subspace integration method for biological cell microscope image classification. The method comprises the following steps: features of biological cell microscope images to be classified are extracted; a feature subspace model is constructed for three extracted image features of the biological cell microscope images by using the kernel principal component analysis (KPCA) to make each type of biological cell microscope images has three feature subspaces; three image features are reconstructed for the biological cell microscope images to be classified by adopting each type of three trained feature subspaces to obtain the reconstruction result of each feature subspace in each type on the image features to be classified, and the classification confidence of the classified images on each type is obtained through the comparison between the reconstruction result of each feature subspace and the originally-extracted image feature vectors; and the images to be classified are classified into the type having the highest confidence. According to the method, the feature dimension can be effectively reduced, the diversity of integration classifiers can be improved, and the classification effect can be further improved.

A computational method is disclosed for producing a sequence of high-resolution (HR) images from an input sequence of low-resolution (LR) images. The method uses a structured subspace framework to learn pairs of LR dictionaries from the input LR sequence ‘and’ employ learned pairs of LR dictionaries into estimating HR images. The structured subspace framework itself is based on a pair of specially structured HR basis matrices, wherein a HR basis spans any HR image whose so-called polyphase components (PPCs) are spanned by the corresponding LR dictionary.

The invention relates to a subspace model identification prediction control method based on data driving, and the method comprises the steps: carrying out offline identification on a controlled object through a subspace model identification method, and obtaining an offline state space model; taking the obtained off-line state space model as a prediction model, controlling the system based on a model prediction control strategy, changing a characteristic value of a controlled object, and collecting on-line input and output data after the characteristic value of the controlled object is changed; identifying the controlled object online through a subspace model identification method, and obtaining an online state space model; and updating the online state space model until the difference value between the system output value controlled according to the online state space model and the real output value of the controlled object is smaller than a preset value. According to the method, under the condition that a system model is unknown, the coefficient of a state space system matrix is obtained through the subspace model identification method, on this basis, the controlled object is controlled, the model is updated, finally, the model of the controlled object is identified accurately and operates stably, and a good operation state is achieved.

The invention discloses an anomaly monitoring method for a non-stationary nonlinear industrial process, and belongs to the field of industrial process anomaly monitoring. On the basis of probabilistic stationary subspace analysis, a kernel probability stationary subspace analysis method is provided by using a kernel skill, and non-stationary and nonlinear characteristics of a complex industrial process can be processed at the same time. The method comprises the following steps: firstly, mapping nonlinear data to a high-dimensional feature space, and establishing a linear model in the high-dimensional feature space; secondly, estimating parameters of the model by using an expectation maximization algorithm; by introducing a kernel skill, realizing a parameter learning process by using a kernel function without knowing an explicit expression of nonlinear mapping. Based on a kernel probability stationary subspace model, two detection indexes are proposed for monitoring a non-stationary nonlinear industrial process. Compared with an original probabilistic stationary subspace analysis method, the non-linear relation in the measured variables can be effectively extracted, and therefore the method is more suitable for monitoring the non-stationary industrial process with the non-linear characteristic at the same time.

The invention provides a cartoon animation generation method with the combination of a hand drawing style. The method includes following steps: (1) firstly constructing three-dimensional grids and a skeleton subspace model of a role by employing a hand drawing; (2) then creating a group of viewpoint relevant models matched with multi-viewpoint hand drawings with the combination of posture reconstruction and a differential domain grid deformation algorithm; and (3) finally combining the hand drawing style with an input motion sequence through motion annotation and a heuristic generalization algorithm to enable the generated animation to maintain original characteristics of the motion sequence. According to the animation generated by employing the method, the original characteristics of themotion sequence are maintained, the drawing style of hand drawing is reflected, and the method can effectively generate cartoon role animation representing the hand drawing style.

A new method is developed for ultrafast, high-resolution magnetic resonance spectroscopic imaging (MRSI) using learned spectral features. The method uses Free Induction Decay (FID) based ultrashort-TE and short-TR acquisition without any solvent suppression pulses to generate the desired spatiospectral encodings. The spectral features for the desired molecules are learned from specifically designed “training” data by taking into account the resonance structure of each compound generated by quantum mechanical simulations. A union-of-subspaces model that incorporates the learned spectral features is used to effectively separate the unsuppressed water/lipid signals, the metabolite signals, and the macromolecule signals. The unsuppressed water spectroscopic signals in the data can be used for various purposes, e.g., removing the need of additional auxiliary scans for calibration, and for generating high quality quantitative tissue susceptiability mapping etc. Simultaneous spatiospectral reconstructions of water, lipids, metabolite and macromolecule can be obtained using a single ¹H-MRSI scan.