34 results about "Linear discrimination analysis" patented technology

The standard model creating apparatus which provides a high-precision standard model used for: pattern recognition such as speech recognition, character recognition, or image recognition using a probability model based on a hidden Markov model, Bayesian theory, or linear discrimination analysis; intention interpretation using a probability model such as a Bayesian net; data-mining performed using a probability model; and so forth, the apparatus comprising: a reference model preparing unit (102) operable to prepare at least one reference model; a reference model storing unit (103) operable to store the reference model (121) prepared by the reference model preparing unit (102); and a standard model creating unit (104) operable to create a standard model (122) by calculating statistics of the standard model so as to maximize or locally maximize the probability or likelihood with respect to the at least one reference model stored in the reference model storing unit (103).

A standard model creating apparatus which provides a high-precision standard model used for pattern recognition such as speech recognition, character recognition, or image recognition using a probability model based on a hidden Markov model, Bayesian theory, or linear discrimination analysis; intention interpretation using a probability model such as a Bayesian net; data-mining performed using a probability model; and so forth. The standard model creating apparatus includes a reference model preparing unit that prepares at least one reference model; a reference model storing unit that stores the reference model prepared by the reference model preparing unit; and a standard model creating unit that creates a standard model by calculating statistics of the standard model so as to maximize or locally maximize the probability or likelihood with respect to the reference model stored in the reference storing unit.

Provided is an image feature extraction method based on global and local structure fusion, characterized by comprising: 1) constructing a weight adjacent map; 2) determining laplacian matrix of similar matrix, degree matrix and images, 3) determining scatter matrix inside the kind and between the kind; 4) determining projection matrix, 5) identifying. The invention provides a feature extraction method of fusing the global structure information and the local structure information, wherein complex features fused of the global feature and the local feature are extracted, thereby the method has strong resolving power. The method not only has the characteristics of holding the reflection method locally, namely holding the characteristics of manifold structure of data; moreover has the characteristics of linear discrimination analysis method, namely assembling the date of the kind more compact to enlarge the distance between the kinds. The invention is applied in image recognition, thereby increasing identifying performance.

The invention relates to a DNN (Deep Neural Network)-HMM (Hidden Markov Model)-based civil aviation radiotelephony communication acoustic model construction method. The method includes the following steps that: a Chinese radiotelephony communication corpus is set up; civil aviation radiotelephony communication speech signals are pre-processed; Fbank features are extracted from the civil aviation radiotelephony communication speech signals and are adopted as civil aviation radiotelephony communication speech features; linear discrimination analysis, feature space maximum likelihood regression transformation and speaker adaptive training transformation processing are performed on the civil aviation radiotelephony communication speech features; and the processed speech features are utilized to build a DNN-HMM-based radiotelephony communication acoustic model. With the method of the invention adopted, the FBANK and MFCC features of radiotelephony communication speech are extracted to traina DNN network, so that the DNN-HMM acoustic model suitable for radiotelephony communication speech recognition can be obtained; and since a dictionary and a language model are combined, so that the feature enhanced DNN-HMM model can reduce the phoneme recognition error rate of the radiotelephony communication speech to 5.62% on the basis of constructed data.

The invention belongs to the field of mode recognition and a brain-machine interface and discloses a motor imagery electroencephalogram voting strategy sorting method based on extreme learning machines. The motor imagery electroencephalogram voting strategy sorting method comprises the following steps: dividing an original motor imagery electroencephalogram into S sections of sub-signals; carrying out dimensionality reduction on each section of sub-signal by a principal component analysis method; carrying out secondary dimensionality reduction on a feature vector subjected to the dimensionality reduction by a linear discrimination analysis method; carrying out the same processing on the S sections of sub-signals to finally obtain S (K-1)-dimensional feature vectors, and combining the S (K-1)-dimensional feature vectors to finally obtain an S*(K-1)-dimensional feature; and transmitting the S*(K-1)-dimensional feature into a plurality of ELM (Extreme Learning Machine) sorting devices so as to obtain a final sorting result by utilizing a voting sorting strategy. The invention provides a voting sorting strategy based on the ELMs; compared with a traditional multi-time ELM average accuracy scheme, the method provided by the invention has the advantages that the sorting accuracy is improved under the condition of not influencing the training sorting low consumption.

The invention provides a feature extraction method of a radiated noise signal of an underwater target, comprising the steps of: firstly researching a visualization method of an acoustic signal to obtain a time-frequency visual image; then carrying out visual feature extraction method research including feature extraction and image centre torque feature extraction based on image texture by regarding a spectrogram in the time-frequency visual image as a searched object; on the basis, respectively utilizing principal component analysis (PCA) and linear discrimination analysis (LDA) algorithms to carry out optimal selection and dimension reduction on the visual feature obtained by the method, and finally applying the visual feature to simulation of underwater target identification. The experimental results show that the average recognition rates of three types of underwater noise signals under the mixed feature can be up to 91.97%, and the recognition rates of the selected visual feature parameters are obviously improved compared with the feature parameters extracted by the conventional method.

The invention discloses a manual alphabet identification method based on an RGB-D image, and belongs to the technical field of behavior identification of computer vision. According to different types of hand texture information of different types of manual alphabets, histogram-of -oriented-gradient characteristics and super-normal-vector characteristics of an RGB-D image video frame obtained through an RGB-D camera are extracted; a principal component analysis and a linear discrimination analysis are combined for characteristic attribute optimizing processing, so that a characteristic attribute with saliency is obtained, and BoW sparse representation is conducted on the characteristics; a data contrast mining technology is used for obtaining representative exemplars in all the types of manual alphabet; finally, a hierarchy judgment policy based on a greedy thought is adopted, the manual alphabets easy to classify are quickly classified through a nonparametric k-Nearest Neighbor Algorithm classifier firstly, and the manual alphabets difficult to classify are judged through a trained support vector machine model based on templates. Compared with the prior art, the manual alphabet identification method is high in both identification accuracy and identification efficiency.

The invention discloses a method for dynamically monitoring abnormal states of users of a treadmill. The method includes firstly, acquiring training data of the users of the treadmill; secondly, preprocessing the data; thirdly, extracting semantic features of the processed data by means of LDA (linear discrimination analysis); fourthly, accumulating knowledge of normal states by using the semantic features as observed quantities, in other words, training an HMM (hidden Markov model); fifthly, judging the abnormal states by the aid of the acquired model and the semantic features. The method has the advantages that the method is based on an algorithm for dynamically monitoring the abnormal states of the users of the treadmill, so that the abnormal states can be quickly and accurately monitored, purposes of training, monitoring and accumulating the knowledge of the normal states can be achieved, and the abnormal states of the users can be effectively monitored under the condition that health states of the users are unknown, and can be monitored intelligently.

The invention discloses a face recognition method comprising the following steps of: establishing a face training database including a plurality of face training images, structuring a projection matrix based on all the face training images of the face training database, and projecting all the face training images in the face training database onto the projection matrix to obtain the characteristic matrixes of all the training images, respectively; projecting an image to be recognized to the projection matrix to obtain the characteristic matrix of the image to be recognized, calculating a distance between the characteristic matrix of the image to be recognized and the characteristic matrix of each training image, and selecting the type of the characteristic matrix of the training image having the shortest distance away from the characteristic matrix of the image to be recognized as the type of the image to be recognized. Therefore, no matrix inversion operation is needed in the method, and the restraint of the inversion operation of a singular matrix is avoided; and the method is higher in robustness and wider in application range, compared with Two-Dimensional Linear Discrimination Analysis (2DLDA).

The invention discloses a big data image classification method. The big data image classification method comprises a first step of enabling image samples to be collected to serve as a training set, a second step of searching a projection matrix optimal in big data image classification, a third step of performing projection on data without marks and a fourth step of adopting a minimum distance classifier to classify the samples after projection. According to the method, local geometric information of sample distribution can be effectively utilized, classified discrimination information is extracted, dependence of big data image classification on manually marked samples is reduced, storage cost in the training process is effectively reduced, and the big data image classification method has higher classification accuracy than a representative image classification method based on linear discrimination analysis.

The invention provides a hyperspectral image target detection method based on tension linear discrimination analysis dimension reduction. The objective of the invention is to solve problems that in the current hyperspectral image target detection method, characteristics of spatial constraint enhancement under the condition of high scores are not fully considered, information excavation cannot be performed on the whole three-dimensional information and the detection precision is quite low. The method comprises steps of 1, acquiring three-order target tension blocks, three-order background tension blocks and three-order to-be-detected test sample tension blocks; 2, allowing sub-space after the projection of the target tension blocks, the background tension blocks and the to-be-detected test sample tension blocks to have the biggest separability; 3, projecting the target tension blocks, the background tension blocks and the to-be-detected test sample tension blocks to a tension sub-space with the biggest separability; 4, calculating the total distance from each to-be-detected test sample to the background and the target; and 5, setting a threshold value and if the gray scale value is larger than the threshold value, determining the pixel of the central point as the target, or else, determining pixel of the central point as the background. According to the invention, the method is applicable to image processing field.

The invention discloses a method and apparatus for training voiceprint recognition system, belonging to the technical field of voiceprint recognition. The method comprises a step of determining the identity vector of each section of speech in a speed training set, and dividing the identity vectors of the speech of a same user in the determined identity vectors into a category, a step of establishing a first function for calculating the first value corresponding to each category, wherein the first value is the sum of the similarities of a first identity vector normalized by using a normalization matrix in a corresponding category and other identity vectors normalized by using the normalization matrix in the corresponding category, a step of determining the normalization matrix which allows the sum of the first value of each category to be maximum, and a step of normalizing the identity vector of the speech obtained in the voiceprint recognition system by using the determined normalization matrix. The problem of the low improvement of the accuracy of the voiceprint recognition by using an identity vector which is subjected to linear discrimination analysis processing in the related technology is solved, and the accuracy of the voiceprint recognition is improved.

The invention discloses a method for identifying rotation of the neck. The method for identifying rotation of the neck includes the steps that a wireless electrode is used for collecting an epidermis electromyographic signal of the neck of the human body and transmitting the epidermis electromyographic signal to an electric signal receiving device; after the electromyographic signal is pretreated, an analog signal is converted into an electric signal through an analog-to-electric conversion device; the electric signal is transmitted into a chip to calculate the characteristics of electromyographic data, the linear discrimination analysis LAD method is used for processing the electromyographic data; an LDA classifier is trained, accurate classification of actions is achieved by the LDA classifier; channel change of a television is achieved, and operation of adjusting the volume is achieved. With the method for identifying rotation of the neck, interaction between the neck and the television is achieved through rotation of the neck.

The standard model creating apparatus which provides a high-precision standard model used for: pattern recognition such as speech recognition, character recognition, or image recognition using a probability model based on a hidden Markov model, Bayesian theory, or linear discrimination analysis; intention interpretation using a probability model such as a Bayesian net; data-mining performed using a probability model; and so forth, the apparatus comprising: a reference model preparing unit (102) operable to prepare at least one reference model; a reference model storing unit (103) operable to store the reference model (121) prepared by the reference model preparing unit (102); and a standard model creating unit (104) operable to create a standard model (122) by calculating statistics of the standard model so as to maximize or locally maximize the probability or likelihood with respect to the at least one reference model stored in the reference model storing unit (103).

The invention discloses a graph model based on multi-view dictionary learning. The method comprises the steps: carrying out the data dimension reduction and preprocessing of original graph data through principal component analysis and linear discrimination analysis, removing redundant features in the data, and maintaining the high discrimination of the data; learning essential characteristics contained in the data by using a multi-view dictionary learning method, and training to obtain a comprehensive dictionary, an analysis dictionary, sparse codes corresponding to the samples and an SVM linear classifier; inputting the sparse codes of the samples into an SVM classifier, and generating a plurality of prediction labels under different visual angles according to a multi-visual-angle principle after the sparse codes are processed by the classifier; the predicted multi-view labels are integrated through a voting mechanism, and final sample labels are generated and used for calculating model accuracy; the method has the advantages of high information utilization rate, more efficient decision making, high specificity and the like.

The invention discloses a kernel image differential filter designing method based on learning and characteristic discrimination. The method includes the steps of stretching an original image block vector corresponding to any point p in a linear discrimination filter in rows about the partial region with p as the center to form an image block vector of neighboring domain pixel information only containing p, conducting dimension expansion on the basis of the image block vector, serially connecting the first and second order differential information at the pixel in the original image block vector, constructing an image block vector matrix to form an intra-class dispersion matrix and inter-class dispersion matrix, and introducing kernel operation for the intra-class dispersion matrix and inter-class dispersion matrix. Filter learning is conducted in a high dimension space in combination with a kernel method, and linear discrimination analysis concept is merged into the learning process, so that detail and non-linear information in an image can be better used to obtain an image filter with better discrimination characteristic description.

Provided is an image feature extraction method based on global and local structure fusion, characterized by comprising: 1) constructing a weight adjacent map; 2) determining laplacian matrix of similar matrix, degree matrix and images, 3) determining scatter matrix inside the kind and between the kind; 4) determining projection matrix, 5) identifying. The invention provides a feature extraction method of fusing the global structure information and the local structure information, wherein complex features fused of the global feature and the local feature are extracted, thereby the method has strong resolving power. The method not only has the characteristics of holding the reflection method locally, namely holding the characteristics of manifold structure of data; moreover has the characteristics of linear discrimination analysis method, namely assembling the date of the kind more compact to enlarge the distance between the kinds. The invention is applied in image recognition, therebyincreasing identifying performance.

The present invention discloses a multi-attribute decision tree power grid stability margin assessment method based on linear discrimination analysis. a key variable discovery model is established based on the offline simulation data and the real-time monitoring data of a power grid to perform effective screening of historical sample data to reduce the data dimensions, a combination relation model among key variables is established to discover the association relation among the variables, extract combination features capable of reflecting important degree contrast of each variable, establish the association relation between the power grid operation state and the transient stability margin, determine the main reasons of system stability level changing, form a concise and accurate knowledge rule base and regulate the decision reference so as to rapidly assess the current stability level according to the system operation state, provide quantification information support for operators' auxiliary decisions and improve the standardization, the rapidity and the adaptive capability of the power grid stability assessment, and therefore the multi-attribute decision tree power grid stability margin assessment method based on the linear discrimination analysis has wide application prospects.

The invention provides a printed font character identification method based on Arabic character set which comprises, extracting region information, character font information, and constituent part information unique for Arabic character set, carrying presorting, determining character class subset of the input character, extracting direction characteristic for reflecting the character stroke composition information, finally employing two steps of characteristic optimization processing including, characteristic dressing, eigen transformation by integrating linear discrimination analysis (LDA) and K-L transformation, finally proceeding categorization judgment through modified quadratic discriminating function (MQDF) statistical classification device.

The invention discloses an LDoS attack detection method based on frequency domain feature fusion, which belongs to the field of computer network security. The method includes: firstly, obtaining the network data message in the router to obtain the sample sequence; then, based on discrete Fourier transform and discrete wavelet transform, transforming the sample sequence from the time domain to the frequency domain, fully extracting the sample sequence Frequency-domain features; then, linear discriminant analysis is used to fuse the extracted frequency-domain features to obtain decision features, which can significantly improve the classification performance of features; finally, input the decision features into the pre-trained single-class classification anomaly detection model, And according to the output of the anomaly detection model, the network data packets in the unit time are judged and detected. If the output of the anomaly detection model is -1, it is determined that an LDoS attack has occurred in the network in the unit time. The detection method based on frequency domain feature fusion proposed by the invention can efficiently, quickly and accurately detect LDoS attacks.

The invention discloses an unsupervised object recognition method combining multi-source feature learning and group sparsity constraints, which includes the following steps: Step 1, obtaining V types of views from an image set to be processed containing c categories, and forming them into a data set X =[x 1 , x 2 ,...,x n ]∈R d×n , where d represents the feature dimension of the data, and n represents the number of samples in the data set; step 2, extract the total scatter matrix S of the data set X t ; Step 3, building a KM clustering model based on linear discriminant analysis on the basis of step 2; Step 4, building a multi-source data joint clustering model based on group sparse constraints and feature selection on the basis of step 3; Step 5, Solve the objective function of the multi-source data joint clustering model obtained in step 4, and optimize it. This method can improve the accuracy of the clustering method, quickly locate the optimal feature subset, and effectively suppress the noise interference in the data set, and finally provide effective support for machine learning and computer vision related applications.

The invention belongs to the field of food analysis and detection and particularly relates to a steamed Lipu taro quality discrimination method based on sensory evaluation and electric nose analysis.The method comprises the steps that (1) the flavor characteristics of steamed Lipu taros are studied, and a sensory evaluation method is established; (2) electric nose detection on the steamed Lipu taros is performed; and (3) quality discrimination on the to-be-detected steamed Lipu taros is performed, wherein data collected by an electric nose sensor array during the electric nose detection in the step (2) is obtained, a corresponding relation between the data and sensory scores obtained in the step (1) is constructed, and according to the characteristic that electric nose sensors can quicklyrespond to different gases, a principal component analysis method, a linear discrimination analysis method and a load analysis method are adopted to analyze sensor characteristic values so as to quickly discriminate the quality grade of the to-be-detected steamed Lipu taros. Through the method, the quality of the Lipu taros treated at different steaming time can be discriminated and distinguishedrapidly and accurately; and the method has the advantages of being objective in result, good in repeatability, high in sensitivity, high in analysis speed, easy to operate, economical, efficient andthe like.

The invention belongs to the field of mode recognition and a brain-machine interface and discloses a motor imagery electroencephalogram voting strategy sorting method based on extreme learning machines. The motor imagery electroencephalogram voting strategy sorting method comprises the following steps: dividing an original motor imagery electroencephalogram into S sections of sub-signals; carrying out dimensionality reduction on each section of sub-signal by a principal component analysis method; carrying out secondary dimensionality reduction on a feature vector subjected to the dimensionality reduction by a linear discrimination analysis method; carrying out the same processing on the S sections of sub-signals to finally obtain S (K-1)-dimensional feature vectors, and combining the S (K-1)-dimensional feature vectors to finally obtain an S*(K-1)-dimensional feature; and transmitting the S*(K-1)-dimensional feature into a plurality of ELM (Extreme Learning Machine) sorting devices so as to obtain a final sorting result by utilizing a voting sorting strategy. The invention provides a voting sorting strategy based on the ELMs; compared with a traditional multi-time ELM average accuracy scheme, the method provided by the invention has the advantages that the sorting accuracy is improved under the condition of not influencing the training sorting low consumption.

The invention discloses a FY-3D infrared hyperspectral cloud detection method based on linear discriminant analysis, and relates to the technical field of satellite remote sensing. The present invention uses the traditional imager-based cloud detection method to match the cloud detection segment products of MERSI with HIRAS pixels to obtain HIRAS data with cloud labels to form a training data set, and then use the linear discriminant analysis algorithm to train the data set , in the training process, the grid search method is used to select the appropriate linear discriminant analysis algorithm hyperconstant, and the model parameters with the best generalization performance are obtained. Finally, the linear discriminant analysis cloud detection model obtained by training is used for the infrared hyperspectral of YF-3D Data cloud detection processing. The linear discriminant analysis cloud detection model of the invention can achieve a classification accuracy of 0.95, and the prediction speed is about 58 times that of the traditional method, showing good generalization performance.

The disclosure provides a voiceprint feature fusion method, including: extracting voice spectrum features, using the voice spectrum features as input, using a general background model and a global difference space matrix to extract i-vector voiceprint features; using deep neural network Network, extracting x-vector voiceprint features and d-vector voiceprint features; using the i-vector voiceprint features, the x-vector voiceprint features and the d-vector voiceprint features as samples, based on linear discrimination The fusion of the voiceprint features is completed through analysis. By introducing a method based on linear discriminant analysis to fuse multiple voiceprint features, the complementarity of multiple voiceprint features and the discrimination of fusion features are improved, which can ensure that the target speaker can pass through voiceprint authentication. rate, reduce the misrecognition rate of non-target speakers, and improve personalized user experience.

The present invention provides a method of identifying origin of a metastasis of unknown origin by obtaining a sample containing metastatic cells; measuring Biomarkers associated with at least two different carcinomas; combining the data from the Biomarkers into a linear discrimination analysis where the linear discrimination analysis normalizes the Biomarkers against a reference; and imposes a cut-off which optimizes sensitivity and specificity of each Biomarker, weights the prevalence of the carcinomas and selects a tissue of origin determining origin based on highest probability determined by the linear discrimination analysis or determining that the carcinoma is not derived from a particular set of carcinomas; and optionally measuring Biomarkers specific for one or more additional different carcinoma, and repeating the steps for additional Biomarkers.

The invention discloses a voiceprint rapid verification method based on application scenes and application thereof. The method comprises the steps: determining a voiceprint judgment threshold based on an application scene after obtaining the application scene; acquiring a to-be-verified voice, inputting the to-be-verified voice into a probability linear discrimination analysis classifier, and outputting a voiceprint score by the probability linear discrimination analysis classifier; and judging whether the voiceprint score is greater than or equal to a voiceprint judgment threshold, if so, passing voiceprint verification, and otherwise, not passing voiceprint verification. The training process of the probability linear discrimination analysis classifier is a process of continuously adjusting parameters by taking the voice with known identity as input and taking the voiceprint score of the voice as theoretical output. According to the method, different threshold values are set for different application scenes, the problem that the data processing amount is too large is solved, verification data are improved, user experience is improved, and the method has great application prospects.

The present invention relates to a face verification method and system. The method includes: using principal component analysis and linear discriminant analysis to respectively preprocess high-dimensional face feature data, including setting the data dimension after principal component analysis dimension reduction; establishing a discriminant High-order Boltzmann machine, set the number of nodes in the hidden layer; use tensor diagonalization strategy to reduce the model parameters of the discriminative high-order Boltzmann machine; input the paired face data into In the discriminative high-order Boltzmann machine, the stochastic gradient descent algorithm is used to maximize the conditional probability of the relationship category, thereby iteratively optimizing the weight of the Boltzmann machine, so as to obtain the final discriminative high-order Boltzmann machine; input the paired face data to be verified into the discriminant high-order Boltzmann machine model, and obtain the corresponding verification result data. The invention enhances the discriminative power of the model by introducing the data relationship category information into the unsupervised Boltzmann machine model, and is more suitable for face verification with precision requirements.