172 results about "Feature transform" patented technology

The invention discloses a face recognition method based on deep transformation learning in an unconstrained scene. The method comprises the following steps: obtaining a face image and detecting face key points; carrying out transformation on the face image through face alignment, and in the alignment process, minimizing the distance between the detected key points and predefined key points; carrying out face attitude estimation and carrying out classification on the attitude estimation results; separating multiple sample face attitudes into different classes; carrying out attitude transformation, and converting non-front face features into front face features and calculating attitude transformation loss; and updating network parameters through a deep transformation learning method until meeting threshold requirements, and then, quitting. The method proposes feature transformation in a neural network and transform features of different attitudes into a shared linear feature space; by calculating attitude loss and learning attitude center and attitude transformation, simple class change is obtained; and the method can enhance feature transformation learning and improve robustness and differentiable deep function.

A computer implemented device is described that is adapted for improving memory efficiency for conducting machine learning on multi-dimensional vectors stored as specially configured data structures. The device includes network interfaces adapted for receiving data sets and communicating with data harvesting applications, as well as data storage and memory for storing the configured data structures. The multi-dimensional vectors and the system provide a data architecture that processes disparate data sets to programmatically extract features transformed from raw information, the extracted features stored in the form of data values suitable for conducting data approximations and neural network processing.

The present invention provides a biometrics-based cryptographic key generation system and method. A user-dependent distinguishable feature transform unit provides a feature transformation for each authentic user, which receives N-dimensional biometric features and performs a feature transformation to produce M-dimensional feature signals, such that the transformed feature signals of the authentic user are compact in the transformed feature space while those of other users presumed as imposters are either diverse or far away from those of the authentic user. A stable key generation unit receives the transformed feature signals to produce a cryptographic key based on bit information respectively provided by the M-dimensional feature signals, wherein the length of the bit information provided by the feature signal of each dimension is proportional to the degree of distinguishability in the dimension.

A “globally invariant Radon feature transform,” or “GIRFT,” generates feature descriptors that are both globally affine invariant and illumination invariant. These feature descriptors effectively handle intra-class variations resulting from geometric transformations and illumination changes to provide robust texture classification. In general, GIRFT considers images globally to extract global features that are less sensitive to large variations of material in local regions. Geometric affine transformation invariance and illumination invariance is achieved by converting original pixel represented images into Radon-pixel images by using a Radon Transform. Canonical projection of the Radon-pixel image into a quotient space is then performed using Radon-pixel pairs to produce affine invariant feature descriptors. Illumination invariance of the resulting feature descriptors is then achieved by defining an illumination invariant distance metric on the feature space of each feature descriptor.

The invention discloses a gesture recognition method of a small quantity of training samples based on an RGB-D data structure. The gesture recognition method is implemented by a feature extraction unit, a training unit and a recognition unit, wherein the feature extraction unit is used for extracting three-dimensional sparse SIFT (scale invariable feature transform) features in aligned RGB-D image sequences obtained by an RGB-D camera; the training unit is used for learning models by a small quantity of gesture training samples, and the recognition unit is used for recognizing input continuous gestures. The gesture recognition method can be applied to any camera or equipment, such as Kinect of Microsoft, Xtion PRO of ASUS or Leap Motion of the Leap company, which can provide RGB-D data; and a real-time recognition speed can be realized, so that the method can be used for man-machine interaction, sign language interpretation, smart home, game development and virtual reality.

The invention discloses a hand-written character recognition method which comprises: obtaining a rough-classification stencil-plate and a fine-classification stencil-plate; receiving a signal that hand-written characters of a user are input and collecting a discrete ordinates sequence of the signal, and processing pretreatment; processing feature extraction and calculating multidimensional feature vectors of the hand-written characters; matching the hand-written characters with the rough-classification stencil-plate, and selecting a plurality of candidate character types with the largest similarity; matching the hand-written characters after feature transformation with a sample center of the candidate character types selected from the fine-classification stencil-plate, thereby determining the most similar character type. The invention also discloses a hand-written character recognition system. The invention has higher speed of inputting and recognizing hand-written characters, and has higher recognition accuracy.

The invention discloses an LDE (Linear Discriminant Analysis) algorithm-based WiFi (Wireless Fidelity) indoor locating method, relating to an indoor locating method and solving the problem of poor location instantaneity of the existing WiFi indoor locating method. The realizing process of the LDE algorithm-based WiFi indoor locating method comprises two stages of an offline stage and an online stage, wherein the offline stage comprises the steps of: constructing a WiFi network, measuring RSS (Received Signal Strength) and constructing a Radio Mao; estimating an intrinsic dimension of the Radio Map by adopting an intrinsic dimension estimating method; carrying out dimension reduction process on the Radio Map by adopting an LDE algorithm to obtain a Radio Map subjected to dimension reduction process and a feature transform matrix, wherein an optimal dimension reduction result and a corresponding feature transform matrix are used as a matching database and a corresponding RSS transform matrix in the online stage. The online stage comprises the steps of: carrying out feature transform on the RSS received in a testing point, matching by adopting a KNN (k-Nearest Neighbor) algorithm and the Radio Map subjected to dimension reduction process to obtain predicted coordinates of the testing point. The invention is suitable for indoor location.

The present invention provides a method for distinguishing nonlinear subspace of one-dimensional distance image which belongs to field of radar target recognition. Target one-dimensional distance image is transferred nonlinearly and mapped to high-dimensional characteristic space, nonlinear regular sub-image characteristics of each kind of target is obtained by characteristic transform matrix of high-dimensional characteristic space to form nonlinear sub-image space of each kinds of target, when the radar one-dimensional distance image of target is input, the sort of one-dimensional distance image is determined according to Euclidean distance between nonlinear regular sub-image and nonlinear sub-image space. Step: radar one-dimensional distance image training vector of target is determined; between-class scatter matrix sB after nonlinear transform and kernel function are determined; nonzero eigenvalue of C and corresponding eigenvector are determined; kind-in scatter matrix Q is determined; nonzero eigenvalue of Q and corresponding eigenvector are determined; all training one-dimensional distance image nonlinear regular sub-image of each kind target are determined; nonlinear sub-image space of each kind of target is determined; input one-dimensional distance image nonlinear regular sub-image of target is dertermined; Euclidean distance between nonlinear regular sub-image and nonlinear sub-image space is determined; sort number of input target one-dimensional distance image is determined.

Provided is a speech translation system for receiving an input of the original speech in a first language, translating an input content into a second language, and outputting a result of the translating as a speech, including: an input processing part for receiving the input of the original speech, and generating, from the original speech, an original language text and the prosodic information of the original speech; a translation part for generating a translated sentence by translating the first language into the second language; prosodic feature transform information including associated prosodic information between the first language and the second language; a prosodic feature transform part for transforming the prosodic information of the original speech into prosodic information of the speech to be output; and a speech synthesis part for outputting the translated sentence as a speech synthesized based on the prosodic information of the speech to be output.

The embodiment of the invention discloses a data processing method and related device and system. The invention relates to the field of artificial intelligence, in particular to the field of automaticfeature engineering. The method comprises the following steps that: an executing device performs multi-order feature transformation on a plurality of data features in an obtained first group of datasets, and selects an optimal data set from the data sets obtained by the plurality of feature transformation; wherein when the nth-order feature transformation is carried out, feature transformation is carried out on each data set in the nth group of data sets to obtain a plurality of candidate data sets; calculating a first evaluation value of each candidate data set in the plurality of candidatedata sets; and then, according to the first evaluation value of each candidate data set, determining an (n + 1) th group of data sets entering the next-order feature transformation, the number of thedata sets in the (n + 1) th group of data sets being smaller than the number of the plurality of candidate data sets.

The invention discloses an unsupervised cross-domain facial expression recognition method. The method comprises the following steps: mapping source domain and target domain samples to a common subspace by finding a feature transformation matrix; a parameter-free maximum mean value difference MMD is introduced to measure the distance between edge distribution and conditional distribution between source domain data and target domain data, and the edge distribution and the conditional distribution are jointly aligned in the subspace to minimize the distribution distance between domains. And thentraining the transformed features to obtain a domain adaptive classifier to predict the data label in the target domain. According to the cross-domain facial expression recognition method, the defectthat most of facial expression recognition needs to meet the requirement that the used test set and training set are from the same data set, namely, the training samples and the test samples meet theassumption that the training samples and the test samples are independently distributed in the same way is overcome, label samples do not need to be carried in a target domain, and the cross-domain facial expression recognition effect is improved.

The invention discloses a moving vehicle tracking method based on corner matching. The moving vehicle tracking method comprises the steps of extracting corners of a detected region through a Harris algorithm according to a video image captured by a surveillance camera with a fixed field of view, calculating scale-invariant rotation feature transform descriptors of corresponding corners, matching the corners according to the scale-invariant rotation feature transform descriptors to form a track, performing calculation to obtain information such as a speed and a direction of the track, dynamically grouping the track, and performing reasonable judgment on a result obtained by initial grouping to determine a vehicle. By means of the moving vehicle tracking method based on corner matching, tracking of the vehicle is completed by performing matching tracking on the track and grouping the track. The moving vehicle tracking method based on corner matching can deal with effect of illumination variation of scenes and partial coverage of the vehicle on vehicle detection and tracking well.

The invention provides a text sentiment analysis method based on deep learning. The method comprises the following steps: (1) inputting text data, removing stop words, and extracting keywords to forma keyword set; (2) forming a dense sub-graph by constructing a keyword co-occurrence graph; vector representations of sentences in the sub-graphs and the document are obtained, and then the sentencesare distributed to the sub-graphs; designing edge connection and edge weight between the sub-graphs to form topological interaction graph expression of the document; and (3) taking the topological interaction diagram as the input of an Emo-GCN model, carrying out node feature extraction transformation, and then fusing local structure information to obtain a node aggregation matrix. The nonlinear transformation is carried out on the aggregated information. The Emo-GCN model adopts a hierarchical structure, and the features are extracted layer by layer. According to the method, the novel topological interaction graph is adopted to express the text information, then the graph convolutional neural network is used for text sentiment analysis, and the method still has strong adaptability. The method is applied to product recommendation, market prediction and decision adjustment, and has extremely high commercial value.

The invention relates to a system and a method for producing an image of a physical object and to a computer program element and a computer readable medium. In order to provide improved stent boost subtract also showing wire state information, a system and a method are provided, the method comprising the following steps: a) tracking a predetermined first feature (126) and a predetermined second feature (128) in a first plurality (114) of first images (116), which images reveal a first criterion (118); and determining a first feature transform; and determining second feature distortion vector fields relative to the first feature transform; b) associating and recording second feature distortion vector fields corresponding to at least two phase attributes (120); c) tracking the predetermined first feature (126) in at least one secondary image (142) which image reveals a second criterion; d) determining a first-feature-based inter-criterion the first-feature-based inter-phase transform and the second feature distortion vector fields corresponding to a matching phase attribute (120); and f) generating a combined inter-criterion image (162) based on the restored physical distortion.

The invention discloses an article recommendation method based on hierarchical multi-granularity matrix decomposition. In a recommendation system, a matrix decomposition algorithm is a recommendationalgorithm for decomposing a scoring matrix into two low-dimensional matrixes, and user preferences and article features can be learned. However, an existing matrix decomposition algorithm and an improved algorithm of the matrix decomposition algorithm only utilize a single feature vector to represent a user and an object, and therefore the problem of low prediction precision exists. In order to solve the technical problem, the invention provides a hierarchical multi-granularity matrix decomposition recommendation method based on deep learning, which can be used for recommending purchased articles with user scores. According to the method, the advantages of feature extraction by deep learning are combined, and the same user or article is represented by utilizing a plurality of different feature vectors, so that the preference representation of the user is more accurate. In addition, the technical problem that an existing recommendation algorithm based on deep learning only uses the lastlayer for prediction, but neglects information loss caused by feature transformation of each layer of the neural network is also solved.

A local connectivity feature transform (LCFT) is applied to binary document images containing text characters, to generate transformed document images which are then input into a bi-directional Long Short Term Memory (LSTM) neural network to perform character / word recognition. The LCFT transformed image is a gray scale image where the pixel values encode local pixel connectivity information of corresponding pixels in the original binary image. The transform is one that provides a unique transform score for every possible shape represented as a 3×3 block. In one example, the transform is computed using a 3×3 weight matrix that combines bit coding with a zigzag pattern to assign weights to each element of the 3×3 block, and by summing up the weights for the non-zero elements of the 3×3 block shape.

The invention provides a multi-feature and multi-model living body face recognition method. The method comprises the steps of obtaining a to-be-recognized RGB face image; decomposing the whole area ofthe RGB face image into a plurality of local areas, and segmenting to obtain an RGB image associated with each local area; performing feature transformation on the RGB image of the local area to obtain a corresponding HSV image, combining the RGB image and the HSV image to form input image information, and outputting the input image information; respectively inputting the input image informationinto each neural network model in the corresponding classification network model for identification so as to respectively obtain model features output by each neural network model corresponding to thelocal area; and inputting model features output by all neural network models in all classification network models into a feature output layer in a unified manner to form a feature output matrix, andinputting the feature output matrix into the fusion feature network model for recognition to output a living body face recognition result of the RGB face image. The method has the beneficial effect ofimproving the accuracy of living body face recognition.

The invention provides a violence and horror video detection method based on MoSIFT (Motion ScaleInvariant Feature Transform) and CSD (Color Structure Descriptor) features, and the method comprises the steps: step1, calculating CSD of a video; step2, calculating a maximum dynamic density point of CSD features; step3, calculating a CSD fraction; step4, calculating MoSIFT of the video; step5, performing vertical dimensionality reduction for MoSIFT features; step6, performing horizontal dimensionality reduction for the video; step7, performing clustering for the MoSIFT features; step8, training SVM-1; step9, calculating a MoSIFT fraction; step10, training SVM-2; and step11, obtaining a classification result. According to the invention, the MoSIFT features and the CSD features are well-used, the complexity of the algorithm is reduced, and a relatively good detection effect can be obtained.

The invention discloses a brain-computer interface transfer learning method based on manifold embedding distribution alignment. The brain-computer interface transfer learning method comprises the following steps of acquiring EEG data of a source subject and EEG data of a target subject respectively; carrying out preprocessing and feature extraction on the EEG data; constructing a transfer learningmodel based on manifold embedding distribution alignment, and training the transfer learning model by using data to obtain a training model; and utilizing the classifier obtained by training to classify the label-free EEG data of the target subject. On the basis of Riemannian tangent plane mapping and manifold feature transformation, feature distribution alignment is integrated into classifier training, and an effective classifier is obtained through training. According to the method, the performance of the brain-computer interface system used by the target user can be effectively improved, and the training burden of the user is reduced.

Systems and methods are provided for producing a road layout model. The method includes capturing digital images having a perspective view, converting each of the digital images into top-down images, and conveying a top-down image of time t to a neural network that performs a feature transform to form a feature map of time t. The method also includes transferring the feature map of the top-down image of time t to a feature transform module to warp the feature map to a time t+1, and conveying a top-down image of time t+1 to form a feature map of time t+1. The method also includes combining the warped feature map of time t with the feature map of time t+1 to form a combined feature map, transferring the combined feature map to a long short-term memory (LSTM) module to generate the road layout model, and displaying the road layout model.

The invention provides a new One-Shot learning method based on Simpleshot. A feature extraction network Effect Nets with optimal performance in the field of image classification at present is used asa feature extraction network; training is carried out on the base class to obtain a feature space, then small samples of a new class are input, feature transformation is carried out on the image in the feature space, namely, the regularization technology is used for relieving the overfitting problem, and finally nearest neighbor classification is carried out.

The invention discloses a freezing gait detection method and a system based on staged feature extraction, and belongs to the field of machine learning. The freezing gait detection method of the present invention comprises the steps, constructing a sample set with a label based on original acceleration data collected in the walking process of a user, wherein the sample is an acceleration data sequence after windowed processing, and the label represents whether the sample belongs to a normal gait or a freezing gait; carrying out staged feature extraction on each sample; performing feature transformation on a gait feature set by using PCA to obtain a low-dimensional new gait feature set, performing feature selection on the low-dimensional new gait feature set to obtain an optimal gait featuresubset; training a freezing gait detection model based on machine learning by using the optimal gait feature subset; extracting a staged feature of the sample to be detected, and inputting the stagedfeature into the trained freezing gait detection model to obtain a freezing gait detection result. According to the method, a motion component and a freezing zone in an acceleration signal are extracted, and a motion signal is synthesized and decomposed, so that the potential features of the original data are brought into full play.

The invention relates to a multichannel convolutional neural network face expression recognition method based on attention mechanism fusion. The method comprises the steps of firstly detecting a faceregion from an input gray-scale image through a Viola-Jones face detector and rotation correction, and reducing the impact on the face expression recognition accuracy from an unrelated region as muchas possible, secondly, applying the detected face region to the depth image and the local binary pattern image to obtain three kinds of complementarity face region data, then, adopting a single-channel feature extraction network to automatically extract features related to expressions from the three types of face region data, sending the extracted features to an interactive attention fusion moduleto be fused, and enabling the module to extract spatial correlation of any two kinds of face region features based on an interactive attention mechanism, thereby realizing effective feature fusion ofdifferent types of face regions, and finally, after the features output by the interactive attention fusion module are spliced and fused again, conducting feature transformation through a full connection layer, and finally acquiring an expression recognition result through softmax operation.

The invention discloses an remote sensing image characteristic dimension reduction method based on mRMR and KPCA, belonging to the remote sensing image processing technology field. The remote sensing image characteristic dimension reduction method firstly uses an mRMR method to perform characteristic selection on an original characteristic set of the remote sensing image to obtain an original set of the remote sensing image characteristics and then uses a KPCA method to perform further dimension reduction on the original subset of the remote sensing image characteristics to obtain an optimized subset of the remote sensing image characteristics. The invention also discloses a remote sensing image characteristic dimension reduction device based on the mRMR and KPCA and a remote sensing image classification method and device. The remote sensing image characteristic dimension reduction method organically combines a characteristic selection method and a characteristic conversion method to perform remote sensing image characteristic dimension reduction, and can effectively improve classification accuracy of the remote sensing image while effectively solving a problem of dimension disaster of the remote sensing image.

The invention discloses a multi-block fault monitoring method based on fault sensitive slow features, and the method comprises the steps: collecting data in an industrial production system, and dividing the obtained data into a training set and a test set; performing slow feature analysis on the training set, and calculating a slow feature transformation matrix; defining a fault sensitivity degreecoefficient, and obtaining the sensitivity degree of each slow feature to the fault according to the coefficient in the slow feature transformation matrix; defining a fault sensitivity degree threshold value, selecting fault sensitivity slow characteristics of each dimensional variable of the training set, and taking the fault sensitivity slow characteristics as training sub-blocks. According tothe invention, through the analysis of the statistical magnitude calculation formula, the fault sensitivity degree coefficient is defined, the slow features are reordered, the slow feature most sensitive to the fault in the variable direction is selected, the sensitive slow feature of each dimension variable is selected, parallel monitoring is carried out, and effective monitoring in the multi-dimension dynamic and large-scale process can be effectively achieved.

The invention provides a ContenseNet algorithm fused with an attention selection mechanism, and the algorithm comprises the steps: carrying out the feature extraction on data in a convolutional neuralnetwork through m network structure blocks, wherein each structure block comprises n groups of complete feature transformation layers, enabling the data to obtain a corresponding feature matrix through each feature transformation layer, cascading the m network structure blocks, performing feature extraction on data of the network structure blocks to obtain a final feature matrix, calculating a loss value of network training according to the obtained feature matrix, calculating an error term and a weight gradient of each layer, and judging whether the network is converged or not according to the loss value, if not, adjusting the convolutional neural network initialization parameter according to the weight gradient to perform training again, and if so, outputting a network training result.According to the CondenseNet algorithm fused with the attention selection mechanism, multi-dimensional feature information is efficiently utilized, the learning and expression capacity of a deep network is enhanced, and the classification accuracy is improved.

Methods and apparatuses for target radial extent determination using deconvolution are presented. One embodiment features a method for determining a radial extent of an target using an active sensor performing noise reduction on a received signal, deconvolving the noise reduced received signal by a transmit signal to produce a target profile, where both the noise reduced received signal and the transmit signal are over-sampled, and determining the target radial extent using the target profile. Another embodiment further features transforming the digitized transmit signal, transforming the digitized receive signal, thresholding the transformed digitized receive signal, dividing the noise reduced, transformed digitized receive signal by the transformed digitized transmit signal, and inverse transforming the divided signal.

The invention discloses a lap joint welding quality defect prediction method and system and a computer readable storage medium. The method comprises the steps that firstly, welding electrical parameters, quality defect labels and labels, deviating from the center of a welding wire, of the center of a welding seam in the welding production process are collected as a training library; effective window extraction is carried out on the collected electrical parameter time series data, Hilbert transform is carried out on the extracted data, and feature extraction is carried out respectively on an electrical parameter real value and a complex value; whether the sample weld joint center deviates from the welding wire center or not is learnt by using a gradient boosting tree; whether the predictedweld joint center deviates from a welding wire center label or not serves as a feature, the maximum and minimum standardization and recursive feature elimination method in the packaging method is usedfor selecting the feature, and a decision tree model is used for conducting welding quality defect classification; and welding electrical parameters are collected in real time, window extraction andcorresponding feature transformation and extraction are performed, the welding electrical parameters are substituted into the welding line center offset welding wire center model for prediction, and aprediction result as a feature is substituted into the welding quality prediction model to determine whether the welding quality has defects or not.