2648 results about "Feature matrix" patented technology

The invention discloses a short text classification method based on a convolution neutral network. The convolution neutral network comprises a first layer, a second layer, a third layer, a fourth layer and a fifth layer. On the first layer, multi-scale candidate semantic units in a short text are obtained; on the second layer, Euclidean distances between each candidate semantic unit and all word representation vectors in a vector space are calculated, nearest-neighbor word representations are found, and all the nearest-neighbor word representations meeting a preset Euclidean distance threshold value are selected to construct a semantic expanding matrix; on the third layer, multiple kernel matrixes of different widths and different weight values are used for performing two-dimensional convolution calculation on a mapping matrix and the semantic expanding matrix of the short text, extracting local convolution features and generating a multi-layer local convolution feature matrix; on the fourth layer, down-sampling is performed on the multi-layer local convolution feature matrix to obtain a multi-layer global feature matrix, nonlinear tangent conversion is performed on the global feature matrix, and then the converted global feature matrix is converted into a fixed-length semantic feature vector; on the fifth layer, a classifier is endowed with the semantic feature vector to predict the category of the short text.

A system and method of providing personalized item recommendations in a communication system comprising a server and a plurality of client devices. At the server, a plurality of user rating vectors are received from a plurality of client devices and aggregated into a rating matrix that is factorized into a user feature matrix and an item feature matrix, with the product of the user feature and item feature matrixes approximating the user rating matrix. The factorization comprises the steps of the ALS1 or the IALS1 algorithm including: initializing the user feature matrix and the item feature matrix with predefined initial values; alternately optimizing the user feature matrix and the item feature matrix until a termination condition is met. The item feature matrix is transmitted from the server to at least one client device, and a predictive rating vector is generated as the product of the associated user feature vector and the item feature matrix. At least one item is selected for recommendation to a user from the items associated with the predictive rating vector.

Disclosed herein are a method and system for training a big data machine to defend, retrieve log lines belonging to log line parameters of a system's data source and from incoming data traffic, compute features from the log lines, apply an adaptive rules model with identified threat labels produce a features matrix, identify statistical outliers from execution of statistical outlier detection methods, and may generate an outlier scores matrix. Embodiments may combine a top scores model and a probability model to create a single top scores vector. The single top scores vector and the adaptive rules model may be displayed on a GUI for labeling of malicious or non-malicious scores. Labeled output may be transformed into a labeled features matrix to create a supervised learning module for detecting new threats in real time and reducing the time elapsed between threat detection of the enterprise or e-commerce system.

The invention provides a method of constructing a text classification model. The method comprises the steps of: constructing a training sample set according to structure features of characters, wordsand sentences of text information, wherein each piece of sample data in the training sample set corresponds to a feature matrix A of a piece of text information about the words and a feature matrix Babout the characters and a category vector O corresponding to the piece of text information, and the dimension number of O is the same as a category number; and using the feature matrix A about the words and the feature matrix B about the characters in the training matrix set as input and the corresponding category vector O as output to train a deep learning model to obtain the text classificationmodel. Classification is carried out according to the classification model constructed by the method, an accuracy rate of text classification can be improved, and the method is particularly suitablefor use in short-text classification.

The invention is applicable to the technical field of identity authentication, and provides a voiceprint recognition method and device. The method comprises the steps of preprocessing input voices toobtain a valid voice in each voice; extracting MFCC acoustic features of each voice and outputting first and second feature matrices including the MFCC dimension and the voice framing number; constructing a short- and long-time recurrent neural network model, and taking the first feature matrix as the input; training feature extraction matrices by using training parameters of the neural network model and speaker characteristics of the voices, each feature extraction matrix corresponding to a speaker model; and selecting the speaker model matching the second feature matrix, with the speaker output corresponding to the matched speaker model as a voiceprint recognition result. The method and device in the invention adopt a supervised learning method to train a voiceprint background model andcan excavate more suitable acoustic features from the training voices so as to be able to distinguish the difference characteristics of speakers more accurately and learn a more robust speaker model for a better voiceprint recognition result.

The invention provides a rolling bearing fault diagnosis algorithm based on time-frequency domain multidimensional fault feature fusion. Aiming at the respective features of vibration signals of a rolling bearing in a normal state, a roller fault state, an inner ring fault state and an outer ring fault state in a time-frequency domain, through extraction of time domain and frequency domain features, redundancy removal and re-fusion, fault features are described in an optimal way to obtain an intelligent judgment result. First, wavelet de-noising is performed on extracted original rolling bearing vibration data; then, time domain feature vectors are extracted to form a time domain feature matrix, and coefficient energy moments after wavelet packet decomposition and reconstruction are extracted to form a frequency domain feature matrix; and the time and frequency domain matrixes are further fused to obtain a time-frequency domain multidimensional fault feature matrix. Redundancy of the multidimensional feature matrix is eliminated to obtain a new multidimensional feature matrix. Then, information of multidimensional features is fused with a weighted feature index distance, and a state judgment result of the rolling bearing is obtained through the feature index distance obtained through fusion.

The invention discloses a classified processing method of internet public information. The method comprises the following steps: selecting a classified public information text as a training text, and parsing words; selecting and screening nouns and verbs, acquiring feature words by extraction, vectorizing the training text, then acquiring a PCA transformation feature matrix, a BP neural network model, and a decision tree rule; performing dimension reduction on vectors of the vector matrix of the public information text to be classified by the PCA transformation feature matrix, and transforming the vectors by the BP neural network model to obtain an output vector which has the same number of dimensions as the classified number, and then performing matching by the decision tree rule, and determining that the public information text to be classified belongs to the public information category marked by the rule if the matching is successful. As the PCA transformation converts a feature word space related to a high dimension into a low-dimensional orthogonal feature space, the disadvantage of inaccurate classification is solved; meanwhile, the decision tree rule is used for classification without data similarity comparison so that a plurality of data sources can be processed in a short time.

The invention discloses a network text named entity recognition method based on neural network probability disambiguation. The method includes: performing word segmentation on an unlabeled corpus, utilizing Word2Vec to extract a word vector, converting a sample corpus into a word feature matrix, windowing, building a deep neural network for training, adding a softmax function into an output layer of the neural network, and performing normalization to acquire a probability matrix of named entity type corresponding to each word; re-windowing the probability matrix, and utilizing a condition random field model for disambiguation to acquire final named entity annotation. A word vector increment learning method without changing structure of the neural network is provided according to the characteristic that network words and new words exist, and a probability disambiguation method is adopted to deal with the problem that network texts are nonstandard in grammatical structure and contain a lot of wrongly written or mispronounced characters, so that the method has high accuracy in network text named entity recognition tasks.

The invention discloses an abnormal vibration analysis-based GIS (gas insulated switchgear) mechanical fault diagnosis method and system. The GIS (gas insulated switchgear) mechanical fault diagnosis method includes the following steps that: vibration signals on a GIS are acquired; de-noising processing is performed on the acquired signals through using a threshold de-noising method; various kinds of feature information contained in the signals are extracted by using a narrowband noise aided multivariate empirical mode decomposition (EMD) method; the power feature of the signals is extracted by using a power density function, and the maximal amplitudes of the power spectra of all IMF of each channel are calculated so as to form power feature matrixes under measured conditions; and a power feature matrix under a normal condition and / or power feature matrixes under various kinds of fault conditions which are obtained through tests are adopted as judgment criteria of faults. Compared with a traditional method, the method and system of the invention have no influence on the normal operation of a whole power system, and can monitor the running state of the GIS safely and reliably, and are suitable for being applied to GIS substations of various voltage levels.

System and method for performing Collaborative Filtering while preserving complete user anonymity are provided. Each of a group of client devices sends a rating vector anonymously to a server. The cells in each rating vector correspond to a set of items, and selected cells have ratings provided by the user associated with the corresponding client device for the corresponding items. The server aggregates all the rating vectors into a rating matrix, and factorizes the rating matrix into a user feature matrix and an item feature matrix through approximation, such that the rating matrix equals the product of the user feature matrix and the item feature matrix. The item feature matrix is sent to the client devices. Each of the client devices calculates its own user feature vector based on its rating vector and the item feature matrix, and provides personalized recommendations on selected items based on the client's user feature vector and the item feature matrix.

The invention discloses a multi-target pedestrian detecting and tracking method in monitoring video, comprising the steps that a target detection network based on deep learning is adopted for detecting a first frame of pedestrian image, and an initial rectangular area having one or a plurality of corresponding pedestrian targets can be obtained; based on the initial target area information, the Histogram of oriented gradients feature of a target can be extracted, and kernel function autocorrelation calculating of Fourier expansion domain can be conducted, and the tracking model is initializedbased on the calculating result; based on the target area information of the tracking model, a multi-dimensional construction of a pyramid will be carried out from the second frame of pedestrian image, and the extracting of the Histogram of oriented gradients feature matrix and the kernel function autocorrelation calculating of Fourier expansion domain can be conducted on each scale of the pedestrian rectangular area; the returned check condition is determined, and the identity re-verification and the updating of the tracking model can be conducted on the pedestrian target having returned check. The invention is advantageous in that the problem of drifting models can be resolved; a more accurate pedestrian moving track can be obtained; real-time performance is good.

The invention relates to a method and a system for network protocol recognition based on tri-classifier cooperative training learning. The method comprises the following steps: carrying out IP (Internet Protocol) regrouping and TCP (Transmission Control Protocol) traffic reduction on network original traffic, and stipulating the unit of network data from original packets to flow; extracting each message of unidirectional flow feature information and vectoring to build a feature matrix; building a tri-classifier cooperative training classifier with few identified samples; judging whether a classifying model of an analyzed protocol exists or not, and utilizing a tri-classifier cooperative training learning method to build a protocol classifier if the classifying model does not exist, otherwise, judging the protocol attributes of data packets; training by a tri-classifier cooperative training learning algorithm based on J48 and obtaining the classifying model of the analyzed protocol; carrying out protocol type judgment on network data packets not identified, and outputting two classes of results: one class refers to the network data packets belonging to the target protocol, and the other class refers to network data packets not belonging to the target protocol. High recognition accuracy and high recalling rate are ensured by the method.

The invention discloses an image understanding method based on a depth residual error network and an LSTM; the method comprises the following steps: firstly building a depth residual error network model so as to extract image abstract features, and storing the features as a feature matrix; using a dynamic attention mechanism in a LSTM model to dynamically form a proper feature vector according to the feature matrix; finally using the LSTM model to form a natural language (English) according to the feature vector. The method uses the advantages of the depth residual error network on image feature extraction and LSTM advantages on time sequence modeling; the depth residual error network and the LSTM model can form an encode-decode framework so as to convert the image content information into the natural language, thus extracting the deep information from the image.

The invention relates to an image processing method and device, a storage medium and computer equipment. The image processing method includes: acquiring a to-be-processed image; determining a modal category to which the to-be-processed image belongs; encoding the to-be-processed image as a semantic segmentation feature matrix through a machine learning (ML) model corresponding to the modal category; decoding the semantic segmentation feature matrix to obtain a semantic segmentation image, wherein pixel points in the semantic segmentation image have pixel values indicating classification categories to which the same belong, and correspond to pixel points in the to-be-processed image; and determining a target image region according to pixel points belonging to a target classification category. The solution provided by the application improves image processing accuracy.

The invention discloses an image detection method and system based on a convolutional neural network, and the method comprises the following steps of collecting and obtaining a preset number of labeled sample images, and carrying out the preprocessing of the labeled sample images, and obtaining a denoised target area image with a label; performing feature extraction on each labeled target area image to obtain a 3D feature matrix of each labeled target area; training the 3D convolutional neural network weak entity detection model by using all the obtained 3D feature matrixes to obtain a trained3D convolutional neural network weak entity detection model; and performing preprocessing and feature extraction on the to-be-detected picture to obtain a 3D feature matrix of the to-be-detected picture, inputting the 3D feature matrix of the to-be-detected picture into the trained 3D convolutional neural network weak entity detection model, and outputting a detection result of the to-be-detectedpicture, so that the weak entity target detection can be completed more accurately and quickly.

An embodiment of the invention discloses an information recommending method and device, wherein the method includes: selecting K neighboring users from a social relation network by learning embedding feature vectors of users in the network according to node2vec algorithm, and extracting potential features of a current user from an embedding feature matrix generated from the K neighboring users according to CNN (convolutional neural network) algorithm; subjecting a historical information rating matrix to iterative alternating operation via a preset algorithm according to the potential features so as to obtain a user feature matrix and information feature matrix of the current user; performing information recommending for the current user according to the user feature matrix and / or information feature matrix. Deep potential features of a current user can be mined so that information recommending efficiency and accuracy are improved.

The disclosure provides a real-time traffic situation awareness system. In one embodiment, the real-time traffic situation awareness system receives driving data from a car, wherein the driving data comprises an image, GPS data, and gyroscope sensor data. The real-time traffic situation awareness system comprises an image processing unit, a feature extraction unit, a feature matrix database, a data grouping unit, and a situation awareness unit. The image processing unit processes the image to generate a processed image. The feature extraction unit generates a data point according to the processed image, the GPS data, and the gyroscope sensor data. The data grouping unit searches the feature matrix database for a plurality of feature groups of an optimal feature matrix corresponding to a geographic area according to the GPS data of the data point. The situation awareness unit analyzes the feature groups to generate traffic information.

The embodiment of the invention provides a multi-tag video classification method and system, and a system training method and device. The multi-tag video classification method comprises the followingsteps: obtaining a video to be processed and extracting an initial feature of the video to be processed; transforming the extracted initial video feature matrix and the extracted initial audio featurematrix to generate a new video feature matrix and a new audio feature matrix; aggregating the new video feature matrix and the new audio feature matrix to generate aggregated feature vectors; generating a plurality of classification tags of a video to be processed and a confidence level corresponding to each classification tag by using an aggregated feature vector. The multi-tag video classification method provided by the embodiment of the invention can improve the accuracy of the multi-tag video classification.

Multi-Task Multi-View Tracking (MTMVT) is used to visually identify and track an object. The MTMVT employs visual cues such as color, edge, and texture as complementary features to intensity in the target appearance representation, and combines a multi-view representation with a robust multi-task learning to solve feature fusion tracking problems. To reduce computational demands, feature matrices are sparsely represented in a single matrix and then decomposed into a pair of matrices to improve robustness to outliers. Views and particles are further combined based on interdependency and commonality single computational task. Probabilities are computed for each particle across all features and the particle with the greatest probability is selected as the target tracking result.

The invention discloses a paper currency number identification method based on a currency detector. A COMS (Complementary Metal Oxide Semiconductor) sensor is firstly installed on the currency detector and connected with a control system. The identification method comprises the following steps of: firstly, preliminarily locating a number region on paper currency; secondly, dividing the paper currency numbers which comprises the steps of: preliminarily dividing ten characters, precisely dividing the ten characters and scaling the characters; and thirdly, character identification which comprises the steps of calculating a gridding feature matrix of the divided characters, making a template matrix and identifying the characters by utilizing a matrix template matching identification method. The method mainly aims to Renminbi of 2005 version, adopts vision inspection of a computer as important technical means, utilizes a digital image processing technology, integrates image acquistion, number locating, number dividing and character identification into a whole, finally identifies 26 letters from A to Z and ten numbers from 0 to 9 and realizes the online automatic identification of the paper currency numbers.

The invention relates to a clustering method and system of a parallelized self-organizing mapping neural network based on a graphic processing unit. Compared with the traditional serialized clustering method, the invention can realize large-scale data clustering in a faster manner by parallelization of an algorithm and a parallel processing system of the graphic processing unit. The invention mainly relates to two aspects of contents: (1) firstly, designing the clustering method of the parallelized self-organizing mapping neural network according to the characteristic of high parallelized calculating capability of the graphic processing unit, wherein the method comprises the following steps of obtaining a word-frequency matrix by carrying out parallelized statistics on the word frequency of keywords in a document, calculating feature vectors of a text by parallelization to generate a feature matrix of data sets, and obtaining a cluster structure of massive data objects by the parallelized self-organizing mapping neural network; and (2) secondly, designing a parallelized text clustering system based on a CPU / GPU cooperation framework by utilizing the complementarity of the calculating capability between the graphic processing unit (GPU) and the central processing unit (CPU).

The invention discloses a high-resolution SAR terrain classification method based on multiscale convolution and feature fusion, and mainly aims at solving the problem in the prior art that the classification precision is low and overfitting easily occurs. The high-resolution SAR terrain classification method comprises the steps of 1, extracting textural features and wavelet features of to-be-classified images; 2, fusing the to-be-classified images, the textural features and the wavelet features to constitute a fusion feature matrix; 3, according to the fusion feature matrix, constructing a training dataset and a testing dataset; 4, adding a multiscale convolution layer and a shuffle layer to an existing CNN network, changing a full-joint layer into a convolution layer, and constructing a multiscale convolution fusion network; 5, using the training dataset to train the multiscale convolution fusion network to obtain model parameters; 6, using the model parameters to initialize the multiscale fusion network to classify a test set. By means of the high-resolution SAR terrain classification method based on the multiscale convolution and the feature fusion, the parameters of the networkare reduced, the overfitting phenomenon of a small sample problem is solved, the classification precision is improved, and the high-resolution SAR terrain classification method can be applied to high-resolution SAR image terrain classification.

The invention discloses a short video recommendation method, a device and a readable medium, belonging to the technical field of video recommendation. In the method and the device provided by the invention, after receiving a short video pull request, a short video sequence composed of a short video list viewed by a user history and a short video list not viewed is obtained. Determining a sequencevector for characterizing a short video feature in the short video sequence according to the short video sequence and the trained short video feature matrix for characterizing all the short video features; according to the sequence vector and the training short video recommendation model, the probability of each short video in the unwatched short video list is determined. According to the probability of each short video, the short video of interest is recommended to the user. By implementing the method, the short video of interest to the user is determined from the mass short video, and the short video is recommended to the user, which not only meets the viewing requirements of the user for the short video, but also improves the utilization rate of the short video application program by the user.

The invention discloses a multi-dimensional geographic scene identification method fusing geographic region knowledge. The method comprises the steps of preprocessing images in a database to obtain satisfied geographic scene images; obtaining object region image blocks by utilizing a method for quickly searching for object regions in the images; pre-training the obtained object region image blocks of the geographic images by using a deep convolutional neural network, performing an accurate adjustment process until the performance of the deep convolutional neural network of the scene images is no longer improved, and fusing feature matrixes into output eigenvectors; pre-establishing a geographic entity noun keyword dictionary by acquired entity noun data in geographic scene classification, performing word segmentation on target identification result data to obtain key words in a target identification result, and establishing text features; and fusing the text features and multi-dimensional image features into eigenvectors as inputs, realizing cross-media-data identification classification, and realizing scene classification fusing geographic entity information.

The invention relates to a resonant earthed system fault line selection method utilizing fuzzy K-means clustering. The resonant earthed system fault line selection method comprises the following steps of 1 performing stretching transformation treatment on transient-state zero-sequence current of each line to improve similarities of transient-state zero-sequence currents of non-fault lines; 2 dividing the transient-state zero-sequence currents of all of lines according to a certain time period, performing subsection phase plane transformation to obtain Euclidean distances from all of phase points of the transient-state zero-sequence current of each section to determinacy points x and y on a phase plane so as to extract local features of the transient-state zero-sequence currents of all of subsections and obtain feature matrixes of global features of all of lines; 3 performing normalization processing on elements in the feature matrixes to improve comparability; 4 utilizing a fuzzy K-means clustering method to perform clustering on the normalized feature matrixes, dividing the transient-state zero-sequence currents of all of lines into two categories and the lines independently included in one category are fault line. The method improves the automation degree and line selection margin.

The invention discloses a low-resolution multi-spectral palm print and palm vein real-time identity recognition method and system. Palm images are collected by the system under the condition of five spectrums, and complementarity of multi-spectrum image information is fully utilized to improve the system recognition rate; meanwhile, palm vein information is collected under the condition of near infrared spectrums so that the system can have the living body detection ability and the counterfeit attack preventing ability of the system can be improved; characteristic extraction speed and other postprocessing speed are improved through the down sampling technology based on bicubic interpolation, and storage space of a characteristic template is saved; characteristic extraction is carried out through a multi-scale multi-directional filter, the influence of lighting changes on characteristic extraction is reduced, and the robustness of the system is improved; a characteristic matrix is coded through a hash table, and system matching speed is further improved; the recognition rate of the system is further improved through the unique fraction-level multi-spectral characteristic fusion method. The system has the advantages of being high in resolution ratio, high in recognition speed, good in stability and expansibility, resistant to counterfeit attack and the like.

The invention discloses a multi-feature fusion multi-target tracking method based on Kalman filtering assistance, and the method comprises the steps: firstly, reading any two frames of images in a video frame, inputting a preprocessed image into a multi-target detector, and obtaining a detection result of each frame in a video; introducing a target occlusion mechanism, enabling the judgment mechanism to carry out judgment according to coordinates of a target center point and the size of a target, and if the occluded part is small or no occlusion exists, enabling the detector to input the masscenter coordinates of the detection frame and the preprocessed video frame into a pre-trained convolutional neural network, extract the semantic information of the shallow layer and the deep layer ofthe target, perform cascading to form a feature matrix, and then perform similarity estimation on the feature matrixes of the two frames to obtain an optimal track; and if the detected target occlusion condition is serious, inputting the mass center coordinates of the detection frame into a Kalman filter, estimating the position information of the target in the next frame according to the previousmotion state of the target, and comparing the estimated coordinate information with an actual detection result to obtain an optimal track.

The invention discloses a text classification method and system based on a graph convolutional neural network. The method comprises the following steps: 1) for each classified annotated text in a texttraining set in a target field, generating a text feature vector of the text according to the word frequency and inverse document rate of words in the text; combining the text feature vectors to generate a text feature matrix, namely a TF-IDF matrix, and constructing a graph structure of the text training set according to the word vector similarity of the words; 2) training a graph convolutionalneural network by using the graph structure and the text feature matrix; and 3) for a to-be-classified text a in the target field, inputting the text feature vector of the text a into the trained graph convolutional neural network to obtain the category of the text a. According to the method, the semantic structure information of the text is considered, the hidden features of the text are capturedfrom another perspective, and the classification accuracy is high.

Embodiments of the present invention provide a three-dimensional point cloud marking method and apparatus based on a fusion voxel. The method comprises the following steps: the data set of the three-dimensional point cloud is voxelized and voxel features in the voxels are extracted based on the processing results to form a first voxel feature matrix; the first voxel feature matrix is used as the input of the three-dimensional convolution neural network to calculate the multi-scale feature of the voxel, and the multi-scale feature is fused in series to obtain the second voxel feature matrix. The first voxel feature matrix is used as the input of the three-dimensional convolution neural network to calculate the multi-scale feature of the voxel. Based on the feature interpolation algorithm, the voxel features in the second voxel feature matrix are extended to the points in the three-dimensional point cloud data set to obtain the point cloud feature matrix. The feature matrix of point cloud is inputted into the multilayer perceptron to mark the attributes of three-dimensional point cloud. The invention can realize fine classification and recognition point by point, so as to further improve the performance of point cloud marking.