64 results about "Sparse learning" patented technology

Provided is a process that includes training a computer-vision object recognition model with a training data set including images depicting objects, each image being labeled with an object identifier of the corresponding object; obtaining a new image; determining a similarity between the new image and an image from the training data set with the trained computer-vision object recognition model; and causing the object identifier of the object to be stored in association with the new image, visual features extracted from the new image, or both.

The invention discloses a migration classification learning method for maintaining a sparse structure of image classification. The method includes the steps of finding two different source and targetdomains with similar distribution, the source domain containing label data, firstly, training a classification classifier on the source domain by using a supervised classification method, and predicting a pseudo label of target domain data by using the classifier; secondly, constructing edge distribution and conditional distribution terms of the source and target domain data respectively by usingthe maximum mean difference, and combining the both to form a joint distribution term; thirdly, constructing a sparse representation matrix S on all the data by using an effective projection sparse learning toolkit, to construct a sparse structure preserving term; fourthly, constructing a structural risk minimization term by using the structural risk minimization principle; and fifthly, combiningthe structural risk minimization term, the joint distribution term, and the sparse structure preserving term to construct a uniform migration classification learning framework, and substituting into the framework using a classification function representation theorem including a kernel function to obtain a classifier that can be finally used to predict the target domain category.

The invention discloses a mutual conversion method of visible light and near-infrared human face images. The method comprises the steps of converting a near-infrared human face image or a visible light human face photo into an initial visible light human face photo or an initial near-infrared human face image by use of a method based on sparse learning, and converting the initial visible light human face photo or the initial near-infrared human face image into a high-definition detail photo of the visible light human face photo or the near-infrared human face image by use of a method based on multi-characteristic selection. According to the mutual conversion method, heterogeneous human face images are fitted in a stratified manner by use of a method based on sparse regularization and the visible light human face photo is generated from the near-infrared human face image, and therefore, the detail information of the synthetic photo is increased and the problem of heterogeneous human face recognition is solved.

The invention relates to the target tracking technology and discloses a target tracking method based on sparse representation. According to the technical scheme, sparse learning is conducted on obstructions by means of the spatial continuity and prior information of the obstructions, the sparse coefficient is obtained by means of an updated sparse representing model on this basis, the reconstitution residual error is calculated according to the obtained sparse coefficient, real-time updating is conducted on a redundant dictionary with the minimum target of the reconstitution residual error as a tracking target, the position of the target at the next moment is predicted with the particle filter tracking method, an estimation target is obtained, and finally the obtained estimation target and the updated redundant dictionary are fed back to the sparse representing model to conduct repeated iteration. According to the method, the sparse learning idea is introduced to the particle filter tracking algorithm based on sparse representation, sparse learning of the obstructions and establishment of an obstruction model can be conducted under the condition that the obstructions are not sparse, and accurate tracking of the target can be conducted according to the updated sparse representation model.

The invention discloses a functional magnetic resonance image data classification method based on super-network discriminant subgraphs. The method includes the steps of preprocessing a resting-state functional magnetic resonance image, and extracting the average time series of divided brain regions; using a sparse linear regression method and sparse learning to optimize an objective function to generate a super-network; extracting each super-edge in the super-network as a sub-graph, calculating the frequency of each sub-graph, selecting a frequency threshold, filtering the frequent sub-graphs, and taking a frequent sub-graph mode as a feature; in a training set, adopting a frequent score feature selection method, and then obtaining an optimal feature subset and a regularization parameter C based on the performance of a test set; adopting a classification algorithm based on a graph kernel and using discriminant subgraphs as features for classification; and quantifying the importance and redundancy of the selected features. For the diagnosis of brain diseases, the method of the invention retains the integrity of an original network topology, without the loss of discrimination of the features, and shows a higher level and more complex interaction among the brain regions.

The invention discloses an emergent event abstract extracting method based on a sparse learning theory. The method comprises the following steps: acquiring emergent event topics, and performing query extension on the emergent event topics to obtain an extension topic lexical item set of the event topics; retrieving events according to extension topic lexical items after the query extension of the events to obtain document sets relevant to the event topics; performing feature selection and semantic clustering on the document sets of the topics in sequence by a non-negative matrix decomposing method to obtain topic clustering results of the events; and extracting a representative sentence from the clustering results to serve as a finial abstract result according to a maximal marginal relevance (MMR) method. Through the technical scheme provided by the invention, when the emergent events burst, latest condition information of event development is provided for a user in the presence of massive redundant news report streams.

Systems and methods for executing an unsupervised feature selection algorithm on a processor which directly embeds feature selection into a clustering algorithm using sparse learning are disclosed. The direct embedding of the feature selection, via sparse learning, reduces storage requirement of collected data. In one method, unsupervised feature selection may be accomplished through a removal of redundant, irrelevant, and / or noisy features of incoming high-dimensional data.

The present invention provides a block sparse compressive sensing based infrared image reconstruction method the method comprises : the step of block sparse representation, the continuous infrared small target image sequence is shown having a block sparse characteristics of the signals; the time correlation modeling step in bayesian framework infrared small target for continuous the data between the image sequence for modeling time correlation; step of outputting a signal reconstruction, sparse learning algorithm is adopted to realize the reconstruction of the infrared small target image, and a final estimated value of the output signal of the present invention also provides a block sparse compressive sensing based infrared image reconstruction system. The technical scheme provided by the invention in the compressed sensing reconstruction process, utilizing a continuous signal in time correlation signals are reconstructed, the single frame signals can be improved, and the reconstruction accuracy significantly reduces the run time of the algorithm.

The invention discloses a super-resolution reconstruction method of multi-frame images. A steering kernel regression method is introduced separately to a clustering phase of an adaptive sparse learning model and a regularized reconstruction process; the two methods are enabled to complement each other's advantages and disadvantages, forming a reconstruction algorithm of better performance. Motionregistration is not required, and therefore, reconstruction errors due to registration errors are avoided and the quality of a reconstructed image can be significantly improved; the method is applicable to places including any motion mode. Compared with existing learning-based multi-frame super-resolution reconstruction methods, the simple efficient integrated reconstruction model that integratessteering kernel regression and sparse learning is provided which can make use of both global structural self-similar priori constraints and sparse constraints to carry out regressive estimation, so that edge details of an image can be better maintained, and image high-frequency distortion can be relieved; the method is not limited to motion amplitude or application scenes and accordingly is adaptive to complex environments.

The invention discloses a sparse learning RCS sequence feature extraction method, belonging to the neighborhood of radar target recognition technology. At first, sparse learning is carried out on thetrain sample, Using the sparse coefficient as the adjusting factor of the distance between the features of samples, A sparse learning projection matrix is established, which automatically selects a small number of neighborhood samples through sparse learning, and can better preserve the local structure information of the sample neighborhood, thereby improving the performance of target recognitionand overcoming the shortcoming of the traditional projection method that only global structure features can be extracted.

The invention provides a method and device for constructing a brain disease classification model, and an intelligent terminal. The brain disease classification model construction method comprises the steps of obtaining Rs-fMRI image data from multiple brain disease data sets; based on a plurality of preset brain templates, preprocessing the Rs-fMRI image data to obtain an average time sequence; constructing a functional connection network corresponding to each brain disease data set and the brain template according to the average time sequence, and extracting features from the functional connection network to obtain a feature matrix; constructing a plurality of initialized brain disease classification models corresponding to each brain disease data set according to a preset sparse learning model, a regularization item and a feature matrix; and performing integrated learning on the plurality of initialized brain disease classification models, and determining a target brain disease classification model corresponding to each brain disease data set. According to the method, the constructed brain function connection network has more physiological significance, the problem of high heterogeneity in data can be effectively solved, and the universality of a brain disease classification model is improved.

The invention relates to a multi-feature and group sparse-based visual object tracking method. The technical feature is that the method comprises the following steps of carrying out multi-feature extraction in a target in the current frame of a video; constructing a learning dictionary under different features by utilizing multi-feature information; carrying out particle sampling in a new video frame; removing unqualified particles by adoption of boundary particle re-sampling, and solving a sparse optimization equation for the remaining particles; updating templates, investigating a cosine similarity between a local frame result and a template with the maximum coefficient; if the similarity is less than a certain value, replacing a template with the minimum coefficient by the current template; and if the video is unfinished, carrying out re-sampling. According to the method, multi-feature, particle filtration and group sparse learning technologies are fused; through tracking multiple features of objects, the constructed dictionary contains richer target information; the tracking precision of the overall algorithm is increased; the stability of the tracking result is improved; and a good visual object tracking result is obtained.

The invention discloses an efficient deep convolutional neural network pruning method, which mainly solves the problem that the existing deep convolutional neural network consumes a large amount of storage resources and computing resources, and the implementation scheme is as follows: a scaling factor is optimized through a sparse learning method based on an ADMM algorithm, the deep convolutional neural network is trained, and the network structure is sparse; a genetic algorithm is used to search the tailoring rate suitable for each layer of the trained deep convolutional neural network, and the optimal tailoring rate meeting the requirement is automatically searched under the guidance of a fitness function; and each layer of the network after sparse learning training is cut by using the optimal cutting rate to obtain the convolutional neural network with the optimal efficiency. According to the method, the precision loss of the pruned convolutional neural network can be greatly reduced, the consumption of storage resources and computing resources by the convolutional neural network is greatly reduced by reducing the parameter quantity of the network, and the method can be used for compression of the deep convolutional neural network.

The invention discloses a multi-channel adaptive detection method for resisting suppression-type hydroacoustic interference. The multi-channel adaptive detection method comprises the steps that to-be-detected data received by a multi-array-element linear array and obeying independent identical distribution are obtained; on the basis of the to-be-detected data, parameters of a pre-established binary sparse hypothesis test are estimated; and the estimated parameters are input into a pre-established detector for resisting the suppression-type hydroacoustic interference based on sparse learning, and adaptive detection of a target is completed. According to the method and a system, sparse information is fully mined, the to-be-detected data subjected to sparse representation are solved by adopting an iterative minimization sparse learning algorithm, thus the hydroacoustic interference target is effectively detected, the key parameters such as interference energy, the number and the angle ofarrival can further be accurately estimated, and actual application is facilitated more.

The invention discloses a communication fingerprint identification method integrating multi-layer sparse learning and multi-view angle learning, which comprises the following steps: 1) adopting a sparse automatic encoder to suppress noise for an original steady-state signal and an original transient-state signal; Carrying out bispectrum analysis and cyclic spectrum analysis on the de-noised signal, and obtaining characteristics on a transform domain by using a sparse coding method based on an over-complete signal dictionary; 2) for the second-order matrix form features on the transform domain,adopting a sparse coding method to obtain low-dimensional features which describe the fine features of the signal more simply and accurately; 3) for the radio station with multiple frequency points and multiple modulation modes, in order to comprehensively extract the common characteristics of the radio station under different working carrier frequencies and modes, adopting tree structure sparsecoding, and 4) from the characteristics of different visual angles, adopting multi-visual-angle canonical correlation analysis to carry out fusion of multiple sparse coding characteristics, and adopting a full connection neural network to carry out classification.

The invention disclose an infrared spectrum identification method and system for polluted gases. The method comprises steps of: extracting characteristics of target gas spectrum under a preset condition, forming a target gas spectrum characteristic data set, and recording the data set as a standard spectral library data set; extracting characteristics of actually measured spectrum of a target gasbased on sparse learning, so as to obtain actually measured spectrum characteristics; selecting spectrum matching the actually measured spectrum characteristics in the standard spectral library data set so as to reconstruct the actually measured spectrum; and determining components of the target gas according to the reconstructed actually measured spectrum. The infrared spectrum identification method and system for polluted gases have characteristics of highly adapting to complex environment under the actual measurement condition and identifying components of mixed gases highly precisely.

The invention discloses a robust graph convolutional network method based on space-time sparse learning. According to the method, spatial sparsity is realized on each node through a TopK function, and a time sparsity-based attention mechanism is provided, namely, different weights are allocated to each dimension of a feature space according to different activation frequencies. The invention provides an improved graph convolutional neural network, the original network precision is maintained, the robustness is high, and the anti-interference capability of the model to noise is improved.

The invention relates to a visible light and infrared image fusion method based on structure group double sparse learning. The method comprises the following steps: (1) carrying out sliding window processing on input visible light and infrared images, searching similar blocks of original image blocks, carrying out group vectorization, and establishing an image similar structure group matrix; (2) taking the image similar structure group matrix as a training sample, forming a base dictionary by utilizing a Kronecker product of shear wavelets, obtaining a sparse dictionary through online learning, and performing linear reconstruction on the base dictionary and the sparse dictionary to obtain a final double sparse dictionary; and (3) in combination with the double sparse dictionaries, performing group sparse solution on the image similar structure group by adopting SOMP to obtain a group sparse coefficient, and obtaining a final fused image through image reconstruction by adopting a maximum fusion rule. The method solves the problem that the existing sparse fusion algorithm ignores the correlation between the image blocks, the dictionary adaptability is poor, and the image fusion quality is low, and can be applied to the fields of remote sensing detection, medical diagnosis, intelligent driving, safety monitoring and the like.

The invention discloses a transformer area topological structure verification method based on sparse learning, and relates to the field of electric power operation and maintenance. At present, in a method for carrying out topological error correction on users in a transformer area by utilizing mass power data, the more the power consumption data of the users at different moments is, the more accurate the estimation result is, but more time is needed for calculation of mass data. According to the technical scheme, on the basis of the power consumption data, collected by the power consumption information system, of the transformer area users, a parameterized model of the power consumption of the transformer area is constructed, a sparse self-adaptive parameter estimation method is provided,model parameters representing the topology of the transformer area users are identified, and the users with wrong statistics of the topological structure of the transformer area are further identifiedthrough threshold detection. The technical scheme has higher precision ratio, recall ratio and higher convergence rate, can carry out calculation on line according to the power consumption data of the users, can capture the change condition of the network topology in real time, and saves a large amount of manual on-site troubleshooting cost.

The invention discloses a Baijiu age detection method based on a 0-order sparse learning TSK fuzzy model, wherein the method comprises the following steps: 1) detecting Baijiu volatile substances of to-be-detected Baijiu with known Baijiu age by an electronic nose, to obtain an original frequency signal; 2) processing the frequency signal obtained in the step 1) by the 0-order sparse learning TSKfuzzy model based on fuzzy subspace cluster, and establishing a Baijiu identification model; and 3) detecting the Baijiu volatile substances of the to-be-detected Baijiu by the electronic nose, and judging the Baijiu age by the identification model based on the 0-order sparse learning TSK fuzzy model and obtained in the step 2). The Baijiu age detection method based on the 0-order sparse learningTSK fuzzy model can quickly, conveniently and accurately judge the Baijiu age, has high detection sensitivity, can overcome the limitations of traditional physical and chemical analysis and artificialsensory evaluation, and provides an effective and simple new method for Baijiu quality evaluation.

The invention discloses a highway electromechanical system fault classification method based on a deep sparse memory model. The method comprises the steps: acquiring data collected by a data collection and monitoring system SCADA of an expressway electromechanical system, and carrying out the data preprocessing; constructing a tripping fault detection model based on a double-layer stack type sparse auto-encoder and an SVM, performing feature extraction and fault detection on the electrical quantity parameters of the highway electromechanical system, and outputting a fault moment; and constructing an improved FSS-LSTM network model, performing fault state classification on the fault moment data of the system, and outputting fault classes. Through close combination of sparse learning and the deep neural network, classification efficiency is improved, classification of various faults is realized more accurately, and technical support is provided for operation and maintenance of an electromechanical system in a highway scene.

The invention relates to a tumor neoantigen prediction method and prediction device based on chromatin high-level conformation and deep sparse learning, and a storage medium, and the method comprisesthe steps: building a deep neural network prediction model based on group selection, and carrying out the training of the model through training data to obtain an immunogenicity predicted value of a to-be-predicted object (namely, a potential tumor neoantigen peptide), wherein each sample adopted when the deep neural network prediction model is trained comprises chromatin 3D conformation information and features generated based on a polypeptide amino acid sequence. Compared with the prior art, the method has the advantages of high prediction precision, convenience in prediction and the like.

The invention discloses a graph clustering method based on robust rank constraint sparse learning. The method is used for solving the technical problem that an existing graph clustering method is poorin robustness. According to the technical scheme, a data similarity graph S is learned through a sparse representation method, and an initial target function is obtained in combination with an L2, 1norm; constructing an initial graph by adopting a k-nearest neighbor method, and constraining the obtained similar graph S in the neighborhood of the initial graph; and laplace rank constraint is added, so that the rank is equal to the number of the data points minus the number of the connected regions of the similar graph, and a final target function is obtained. Converting the constrained optimization problem of the target function into an unconstrained optimization problem by applying an augmented Lagrange multiplier method; alternately optimizing variables contained in the target function;updating parameters contained in the augmented Lagrange multiplier method at the end of each iteration; and after iteration reaches a termination condition, decomposing the similarity matrix according to the obtained solution to obtain a final clustering result. According to the method, the robustness of the method is improved by constructing the graph with high quality and utilizing the L2, 1 norm.

The invention discloses a static-dynamic collaborative awareness complex industrial process operation state evaluation method. The method comprises the following steps that: a KPI-Driven SFA algorithmis used for conducting static-dynamic characteristic cooperative sensing information mining, so that an offline evaluation model of an operation state is established; a sliding window technology is introduced, the similarity between the score vector and first-order difference of online data and the score vector and first-order difference of each state level is calculated, and static and dynamic evaluation indexes are calculated according to the similarity; effective evaluation rules are formulated, and online identification of the process operation state is realized according to the static evaluation indexes; and online identification of a process operation state change trend is realized according to the dynamic evaluation indexes, and comprehensive evaluation of each state and a transition process is completed, aiming at a non-optimum variable, a data-driven fault diagnosis method based on sparse learning is used to reduce the influence of irrelevant variables, and the accurate position of the non-optimum variable is traced according to group contribution GWC. The method can effectively guarantee the quality of industrial products.

The invention relates to a method for computing cloud cover in cloud atlas based on a multi-layer unsupervised sparse learning network. The method comprises the following steps: performing unsupervised layer-by-layer feature coding on a picture by utilizing a forward layer-by-layer sparse automatic coding machine so as to obtain high-order semantic information; dividing the cloud atlas into thickcloud, thin cloud and clear sky by utilizing the high-order semantic information; and finally, computing the total cloud cover in the cloud atlas by utilizing a 'spatial correlation method'. Comparedwith the traditional satellite cloud atlas cover computing, the method disclosed by the invention is high in accuracy, and the sample training time and cloud cover computing time can be greatly reduced under the same hardware condition.

The invention discloses a functional magnetic resonance image data classification method based on a supernetwork discriminant subgraph, which preprocesses the resting state functional magnetic resonance image, and extracts the average time series of each segmented brain region; uses sparse linear regression Method and sparse learning to optimize the objective function and generate a hypernetwork; extract each hyperedge in the hypernetwork as a subgraph, calculate the frequency of the subgraph, select the frequency threshold, filter the frequent subgraph, and use the frequent subgraph pattern as a feature ;Use the frequent score feature selection method on the training set, and then get the optimal feature subset and regularization parameter C based on the performance of the test set; use the classification algorithm based on the graph kernel, and use the discriminative subgraph as a feature to classify ; Quantify the importance and redundancy of the selected features. For the diagnosis of brain diseases, it not only retains the integrity of the original network topology, but also does not lose the discriminability of features, and presents higher-level and more complex interactions between brain regions.

The invention provides a hyperspectral image target prior optimization method based on multitask sparse learning. The method includes performing pre-detection and set operation; obtaining an over-complete target dictionary and a to-be-reconstructed target pixel; building a multi-task sparse representation model, performing complementary learning on sparse representation models of different targetpixels by utilizing spectral similarity of the target pixels, and using an average sparse coefficient of target atoms as a reconstruction weight of the target atoms for an optimal target prior spectrum, so that a target prior optimization effect is improved. The invention has the advantages that background pixels can be prevented from participating in the target priori optimization process, the spectral similarity of the target pixels is fully utilized, and the optimization effect of the target priori is improved.

The invention discloses an electroencephalogram channel optimization method based on combination of deep learning and sparse learning. The method comprises the steps that firstly, a model driving experiment is used for collecting a data set, samples in a source domain and a target domain tend to be balanced through data expansion, sparse learning and domain adversarial learning are conducted on the balanced samples, and the design initial intention is to minimize loss values of a label predictor and a discriminator at the same time; based on the purpose, an objective function is designed, themodel has the feature selection capability by adding L21norm, in addition, GAN is used, and the robustness and generalization capability of the model are improved to a certain extent. And finally, inthe experimental evaluation stage, on one hand, the performance of the method is independently evaluated, and on the other hand, the performance is compared with other channel optimization algorithm items, and unique advantages are obtained. On the other hand, on the premise that the accuracy is guaranteed, the number of channels can be effectively reduced, and therefore the burden and expenditureof the system are reduced.

The invention discloses an anti-machine-impacting method and device for a sixth-shaft movement gripper of a robot. According to the method, a current sensor is used for monitoring the current changing relations when various shaft motors normally work, sparse learning is performed, the current information of the shaft motors at the maximum load and at the work extreme position is obtained, and the current information is recorded as a robot current analysis expert database; meanwhile, a precision position sensor is adopted for detecting the gripper position information; finally, a field measuring and control computer is combined with the information of the current sensor and the precision position sensor to perform fuzzy logic control, and the action safety control pre-warning information is output; and rigid supporting columns capable of achieving strain are added at the mechanical action position of the gripper. The device comprises a base (1); the sixth-shaft movement gripper (2) of the robot is connected to the position above the base (1) through springs (3) and the supporting columns (4); and a weak point (5) is arranged on the middle portion of each supporting column (4). By means of the anti-machine-impacting method and device for the sixth-shaft movement gripper of the robot, machine impacting of the sixth-shaft movement gripper of the robot can be prevented, and the maintaining cost can also be reduced.

The invention provides an abstract method based on a social media microblog specific topic. The method comprises the following steps of (1) obtaining W according to a following formula in the description, wherein the formula merges a microblog abstract optimization model of group sparsity study and social regular term parameters; the S is a text matrix; a data set is obtained through the calculation from TF-IDF; the W is a refactoring coefficient matrix; Lambda is a group sparsity regular term parameter; the L is D-T and is an Laplace matrix; (2) obtaining the importance Score (i) of the i-thmicroblog by calculating the normal formulas of the i-th line of the W shown in the description; sequencing the microblog according to the importance; further screening the front k microblogs to be used as the abstract, wherein the Score (i) is shown in the description. The abstract method is based on the basic framework of the sparse reconstruction and merges the social media content and a socialnetwork structure; the obtained microblog abstract is more similar to an expert mutual evaluation result in aspects of three evaluation indexes of ROUGE-1, ROUGE-2 and ROUGE-SU4 through being compared with the existing model.