163results about How to "Achieve dimensionality reduction" patented technology

The invention provides a face identification method based on texture feature fusion and an SVM, which belongs to the image processing field. The method comprises the following steps of firstly using a uniform LBP operator to extract NSCT transformation multi-scale multi-direction high-frequency sub-band texture features, and counting and further combining the uniform mode LBP feature information of each high-frequency sub-band to acquire a face texture feature ULNBH combining advantages of both the LBP operator and the NSCT. The ULNBH is lack of low frequency information. Therefore, the ULNBH feature and a Gabor feature are fused in a feature layer by using the characteristics of the Gabor feature to acquire a fusion feature with more complete face texture feature information. During a face identification stage, a principal components analysis (PCA) method is adopted to perform dimension reduction on high-dimension characteristic vectors. The SVM is further adopted to identify the fusion feature with the dimension reduced. The fusion feature has greater robustness relative to illumination and posture changes.

The invention discloses a speaker age identification method based on an SVM (Support Vector Machine) classifier. The method comprises the following steps that a voice library in which voice signals of speakers of different ages are stored is established; the voice signals in the voice library are preprocessed; voice feature parameters of the preprocessed voice signals are extracted; the SVM training is performed on the basis of the extracted voice feature parameters, and then an SVM model is obtained; and according to the SVM model, the voice feature parameters X of voice to be identified are predicted, after output of each SVM is logically judged in the process of prediction, the voice feature parameter with the most votes is used as the most probable age class, and then a final age identification result is obtained. By using the method provided by the invention, the blank of the prior art in related research on speaker age identification is filled to a certain degree, the speaker age can be judged better, and the method has a broad application prospect on occasions such as man-machine interaction, criminal search, games, entertainments and the like.

The invention discloses a convolutional neural network based ribbon edge burr defect detection method. A camera collects a ribbon picture, the edge is extracted from a ribbon, and a sample picture with a burr defect and a sample picture without the burr defect are obtained; and the collected sample pictures are detected in a classified way via the convolutional neural network with a multiscale parallel training structure, the convolutional neural network increase the depth and width of the neural network, a full connection layer is removed from a common convolutional neural network, common convolution is converted into sparse connection, and a quasi-optimal local sparse structure via dense components is used to maintain a high computing performance of the neural network. Thus, the blur defect detection method can be used to detect blur defects of the ribbon edge effectively, maintain or reduce the computational complexity of the convolutional neural network, and thus, improve the computing performance.

The invention discloses an audio classification method, device and a computer readable storage medium, and belongs to the technical field of electronics. The method comprises: collecting an audio signal; intercepting or supplementing the audio signal to adjust the duration of the audio signal to a preset duration; converting the audio signal to a target audio according to the frequency informationof the audio signal; extracting audio features of the target audio through a convolutional network contained in a preset classifier; extracting time-order features of the audio features through a threshold circulation network contained in the preset classifier; and determining a probability that a category of the target audio is a preset category identified by each of multiple preset category identifiers through a fully-connected network contained in the preset classifier according to the time-order features; and determining the preset category identified by a preset category identifier having the highest probability among the multiple preset category identifiers as the category of the target audio. With the adoption of the method, segmentation of the target audio is avoided, the integrity of the target audio is preserved, and the classification accuracy is relatively high.

The application discloses a switch cabinet fault feature selection method and apparatus. The method comprises: based on a sorting principle of importance degree of fault features, performing sorting on N fault features of a fault feature set to obtain to-be-selected fault feature subsets; based on mRMR criterions, performing screening on the to-be-selected fault feature subsets with an increment search method; and calculating the classification accuracy of each candidate fault feature subset in N candidate fault feature subsets obtained after screening, and determining the candidate fault feature subset with highest classification accuracy as an optimal candidate fault feature subset. According to the switch cabinet fault feature selection method and apparatus, the fault features in the fault feature set are sorted in advance, so that the dimension reduction of the fault features is realized; and then the screening is performed based on the mRNR criterions, so that noise components in a fault feature sample are effectively reduced, the feature dimensions are further reduced, the screening effect is improved, and the degree of matching between the subsequently obtained optimal fault feature subset and an actual fault reason is increased.

The invention discloses a customer segmentation method and device based on cluster analysis. The method comprises the following steps: acquiring a customer information original data set to perform numerical preprocessing so as to obtain a data sample; performing dimension-reduction and feature extraction on the data sample through an automatic encoder, performing the variation coefficient method on the data sample processed by the automatic encoder so as to compute the weight of the attribute feature, and computing the distance between the sample points by adopting the weighted Euclidean distance formula; computing the average distance among all data samples, traversing an adjacent point, wherein the distance between each sample point and the adjacent point is less than the average distance, and counting all sample adjacent point amount and sorting according to descending order to determine k initial cluster center points; and clustering the remaining data according to the weighted Euclidean distance formula to accomplish the customer segmentation method.

The invention discloses a visualization method and device for a random forest model and a storage medium, and relates to the technical field of machine learning, and the method comprises the steps ofscreening a target training sample meeting a preset condition from a training sample set corresponding to each decision tree of the random forest model, so as to form a target training sample set forconstructing a classification tree; obtaining the variable importance degree of each characteristic variable in each decision tree, and carrying out descending sorting on all the characteristic variables according to the variable importance degrees; according to the target training sample set and all the feature variables after descending sorting, starting from a root node of the classification tree, optimal feature variables and optimal segmentation values corresponding to all the nodes in the classification tree are sequentially determined by taking the Gini coefficient as a splitting rule,so that the classification tree is constructed; and generating a tree-shaped visual graph corresponding to the classification tree and outputting the tree-shaped visual graph. According to the invention, the decision process of the random forest model can be visually displayed, and the interpretability of the model is improved.

The invention discloses a deep learning-based fraud transaction identification method, system and a storage medium. The method comprises the following steps: acquiring a training sample, wherein the training sample is transaction data for establishing a fraud transaction detection model; constructing a stacked RBM neural network structure; training the stacked RBM neural network structure based onthe training sample, and generating a dimension reducer; performing dimension reduction on the training sample via the dimension reducer, and clustering the binary state vector obtained by dimensionreduction so as to establish a fraud transaction detection model; obtaining transaction data to be detected, and analyzing the transaction data to be detected according to the fraud transaction detection model so as to identify fraudulent transactions. The deep learning-based fraud transaction identification method, system and the storage medium in the invention can improve the accuracy of fraudulent transaction identification, and does not need to define a similarity measurement method in advance, thereby reducing difficulty and cost, and the high tolerance to sample data is achieved.

The present invention provides a liquid crystal hyperspectral calculation imaging measurement device of three-dimensional encoding. The device comprises a front-end lens 2, a wave band selection and splitting module 3, a space encoding module 4, a collimating lens 5, an area-array detector 6, a data storage module 7 and a calculation reconfiguration module 8; and based on the three-dimensional encoding, the measurement device performs projection measurement of the three-dimensional spectral data of an object consisting of two-dimensional space information and one-dimensional spectral information in the random encoding information, and performs dimensionality reduction of the hyperspectral data at the data collection phase to obtain the compressed hyperspectral data with the selected central wavelength. Compared to the traditional hyperspectral imaging system, the liquid crystal hyperspectral calculation imaging measurement device and method of three-dimensional encoding can realize the compressing sampling on the space, and can perform spectrum selection at the data collection phase so as to realize the data dimensionality reduction, avoid data redundancy, reduce the data volume, improve the information utilization and facilitate rear-end transmission and storage.

The invention belongs to the technical field of a heat engineering technology and an artificial intelligence crossing technology, and relates to a variable dimensionality reduction modeling method for boiler combustion optimization. According to the method, DVs (disturbance variables) and MVs (manipulated variables) are selected as auxiliary variables of a model, CVs (controlled variables) to be predicated are used as output of the model, historical operation data is selected as initial training samples, principal component analysis is utilized for carrying out feature extraction on the DVs of the model, the dimensionality reduction of input variables is realized, the extracted feature variables and the MVs are simultaneously used as the input of the model, and an LSSVM (least square support vector machine) is used for building a CV model of a boiler. The variable dimensionality reduction modeling method has the advantages that through the dimensionality reduction on the input variables, the predication precision and the generalization capability of the model can be effectively improved, the precise prediction on the CV can be realized, and the important significance is realized on the combustion optimization of a power station boiler.

The invention discloses a parallel network flow classification method. Based on a MapReduce parallel frame provided by a Hadoop cluster platform, a data set is firstly pre-processed, and high-dimensional network flow data dimension reduction is performed by means of a feature selection approach to remove uncorrelated and redundant characteristics; multiple base classifiers are trained through selective ensemble learning, the high-accuracy and large-difference base classifier ensembles are selected; finally a final classifying result is obtained in a majority voting mode. By means of the parallel network flow classification method, the mass data dimension reduction and classification problems can be effectively solved, and data processing efficiency is improved to the great extent.

The invention provides a webpage classification method. When a user needs to access a webpage, the accessed webpage is classified, webpage features are extracted from webpage elements of the webpage,and each extracted webpage feature is sent to an initial classifier corresponding to the extracted webpage feature. A plurality of different initial classifiers are applied, each initial classifier can identify malicious webpages, and the types of the malicious webpages identified by the initial classifiers are different; each initial classifier processes each received webpage feature to obtain aprimary category of the webpage, and then each primary category is analyzed in an integrated classifier, so that the webpage category of the webpage can be determined finally, and thus the user determines whether to access the webpage or not according to the webpage category of the webpage to be accessed, and the safety of accessing the webpage by the user is improved.

The invention discloses a rehospitalization risk predicting method based on a cost-sensitive integrated learning model. The method comprises the following specific steps of: 1), acquiring medical andexternal environment data information, and constructing a multi-source high-dimension characteristic matrix; 2), performing high-dimension characteristic matrix nonlinear compression expression basedon an automatic encoder; 3), constructing an integrated learning model in which a cost-sensitive support vector machine is used as a weak learner; and 4), through characteristic processing of the step1 and the step 2, inputting a predicting set into a training model, and obtaining a rehospitalization risk predicting result. The method aims at patient demography information, previous hospitalization history, family history and an external environment characteristic and constructs the multi-source high-dimension characteristic matrix, thereby extracting more characteristic information which fully reflects the health condition of the patient. Based on high-dimension characteristic matrix nonlinear compression expression of the automatic encoder, dimension reduction on a sparse characteristicis realized. For aiming at a sample disproportion problem, the integrated learning model in which the cost-sensitive support vector machine is used as the weak learner is constructed, thereby improving rehospitalization risk identification precision.

The invention belongs to the field of multimedia information security protection, in particular to an image security retrieval method based on the combination of a word bag model and a minimum hash principle, which can be used for the security retrieval of large-scale images. A content owner combines a sack model with the minimum hash principle to construct a secure index of the image features. Inthe safe index data set of image features, the noise index vector is introduced, and the index vector corresponding to some visual words is randomly extracted to construct the safe index table. The image security index table and the encrypted image are uploaded to the cloud server. When the user requests retrieval, the cloud service only searches the index table according to the query image indexinformation, and the user obtains the image to be retrieved according to the similarity between the index vectors. This retrieval method has higher efficiency and is more suitable for large-scale dataset retrieval. The feature vector based on SIFT descriptor and binary signature can achieve high precision matching, and has high retrieval accuracy.

The invention relates to a text classification method and system based on deep learning of a hybrid automatic encoder. The method is a method for combining a sparse restricted Boltzmann machine (SRBM)with a contraction automatic encoder (CAE) to form a hybrid automatic encoder training model, the advantage of feature extraction of the robustness of the contraction automatic encoder (CAE) and theadvantages of feature representation and contrast divergence fast learning of the sparsity of the sparse restricted Boltzmann machine (SRBM) are combined, the learning capability of the hybrid automatic encoder is improved, and the dimension of the feature space is reduced; an unsupervised layer-by-layer greedy learning algorithm is used for training the model, the parameter convergence rate is increased by adding Polyak Averaging when parameters are updated, and fine tuning is conducted on the model by means of a back propagation (BP) algorithm; finally, a support vector machine (SVM) is adopted for classification, the text feature dimension is reduced, and the accuracy of the text classification is improved.

The invention relates to a dense medium suspension liquid coal slurry content determination method by using principal component analysis and a support vector machine, which belongs to coal slurry content industrial detection methods. The method is characterized in that a plurality of principal components are constructed by utilizing the principal component analysis method, so as to facilitate the dimensionality reduction treatment of high-dimensional pressure difference data in the static settling process of dense medium suspension liquid, and a training sample is formed after extracting the principal components; an epsilon type support vector regression machine is further used for establishing a coal slurry content measurement module, and the basic genetic algorithm is used for realizing the selection of optimal parameters of the support vector machine during the modeling process. The error of the coal slurry content measured by the method is +/-4%.

The invention provides an atomizing inhalation system. The system comprises an integrated distribution box, a distribution box shell, an atomizing box, an oxygen box, supporting pillars, a first oxygen tube, a second oxygen tube, an oxygen supply source distribution chamber, a valve, a ventilating device, a ventilating device shell, an electric telescopic rod, a ventilating baffle, a limiting plate, an air screen, a treatment device, an adjustor, an allowance displayer and an oxygen supply source. An electronic medical treatment information end, an emotion sensing module, a brain wave sensing module, a body temperature sensing module, a remote diagnosis unit and a monitoring system are arranged inside the integrated distribution box. Through the integrated distribution box, oxygen is directly atomized through the atomizing box, atomizing time is shortened, and operation is convenient for medical workers; through the emotion sensing module and the brain wave sensing module, the mental condition of a patient is monitored in real time; treatment conditions are monitored in real time through the body temperature sensing module, electronic medical treatment information end, a remote diagnosis unit and the monitoring system, and the accuracy and safety of treatment are guaranteed.

The invention belongs to the technical field of mechanical equipment key part service life, and discloses a mechanical equipment key part residual life prediction method combining AE and bi-LSTM, andthe method comprises the following steps: carrying out the feature extraction of input data through an auto-encoder; dividing the data after feature extraction to obtain a training set and a test set;constructing a bidirectional LSTM prediction model, wherein in the bidirectional LSTM prediction model, an LSTM network hidden layer comprises a forward layer and a backward layer; carrying out training through training set data and test set data until an evaluation index is close to the optimal, and storing the bidirectional LSTM prediction model and parameters thereof; and inputting the to-be-predicted data into the bidirectional LSTM prediction model, and outputting the predicted life. The method improves the prediction result, and can be applied to the field of mechanical part life prediction.

The invention discloses an information hiding method based on the combination of image normalization and principal component analysis (PCA). The information hiding method belongs to the technical area of a computer image processing and information security. Based on the theory of image normalization and invariant centroid, the information hiding method has a good resistance to the geometric attack. Before implanting the hidden information, the chaotic technology is utilized to carry out the encryption processing on the hidden images so that the confidentiality and security of the hidden information are effectively improved. According to the features of the visual system---illuminance masking and texture masking, a perceptual masking template is designed to determine the intensity factor embedded by each sub-block hidden information and then the hidden information is embedded into multiple subspaces of the images so that the information embedding capacity of the cover images can be effectively improved.

The embodiment of the invention discloses a item recommendation method and device and a storage medium, and the method comprises the steps: extracting item characteristics from a item which a currentuser is interested in; expanding the extracted item features to obtain target features; pulling an item according to the target feature to obtain a candidate item; extracting a user tag from the userportrait of the current user; calculating the similarity between the current user and the candidate item according to the extracted user tag and the item characteristics of the candidate item; and selecting a recommended target item from the candidate items according to the similarity between the current user and the candidate items. According to the embodiment of the invention, the algorithm complexity can be reduced, and the recommendation accuracy is improved.

The invention relates to a random forest visualized data analysis method based on LargeVis. The random forest visualized data analysis method comprises the steps of preprocessing a training dataset; extracting important characteristics of the training dataset through a random forest; adopting LargeVis to reduce the dimension; based on the random forest of LargeVis, conducting visualized processing. By means of the random forest visualized data analysis method based on LargeVis, aiming at high-dimension data, through the characteristic importance trained by the random forest, new sub-high-dimension data is formed, and then through the data subjected to LargeVis dimension reduction, the sub-high-dimension data is sent into the random forest to be predicted and analyzed to form visualization,the classifying precision can be improved, the visualization time can also be prolonged, and meanwhile, the random forest visualized data analysis method adapts to different pieces of data.

The invention discloses a high-dimensional damaged data wireless transmission method based on a noise reduction auto-encoder. The method comprises model training and end-to-end transmission. In the model training, firstly, data preprocessing is performed on a historical perception data set, and the historical perception data set is divided based on a K-fold cross validation method; and then constructing a noise reduction auto-encoder model, and training the noise reduction auto-encoder model based on a proposed novel noise adding mode of introducing random Gaussian noise in batches. According to end-to-end transmission, firstly, a noise reduction auto-encoder obtained through training is divided into two parts to be deployed at a sending end and a receiving end, then, perception data of unknown type of noise interference is subjected to preprocessing and dimension reduction operation at the sending end, the data subjected to dimension reduction is transmitted to the receiving end, finally, reconstruction operation is executed at the receiving end, and reconstruction data of the undamaged perception data is obtained. According to the method, dimension reduction transmission, noise reduction processing and reconstruction of high-dimensional damaged sensing data can be effectively carried out, and noise interference is filtered and dimension reduction transmission is carried out when data collection is carried out in a severe environment.