70results about How to "Effective modeling" patented technology

The invention discloses a human body behavior recognition method and system based on a graph convolution network, and the method comprises the steps: extracting human body skeleton information from animage containing human body behaviors, obtaining a human body joint point position information sequence, and constructing a topological graph sequence with any length of a human body skeleton; performing feature extraction and topological structure adaptive evolution on the topological graph sequence through a topological learnable graph convolution-based space-time graph convolution network to obtain new node features fused with local space-time features and a topological graph sequence with a new topological structure; performing feature extraction through a graph convolution long-term andshort-term memory neural network; global spatio-temporal features are obtained through global pooling operation; and performing human body behavior recognition based on the global spatial-temporal features through a classifier. The features of a whole graph are directly learned, the weight matrix in graph convolution is expanded to the whole topological graph structure, the relation between any two nodes in the graph is learned, limitation of the topological structure is avoided, and the recognition accuracy is high.

The invention discloses a vertical graph recognition method for converting a building drawing into a three-dimensional BIM model, and the method comprises the following steps: a, obtaining a target graph layer of the CAD building drawing, and obtaining a wall graph layer, a door and window graph layer, an elevation graph layer, an axis symbol graph layer, and an axis network graph layer; b, performing direction identification, elevation symbol identification and story height acquisition on each vertical drawing of the CAD building drawings; c, performing building component recognition, visibility analysis and three-dimensional positioning on each layer of plane drawing of the CAD building drawing; d, carrying out bounding box construction on the elevation drawing paper in each direction ofthe CAD building drawing paper, and carrying out search and size measurement on the elevation drawing component; according to the method, when the CAD building drawing is converted into the three-dimensional BIM model, the components of the elevation map of the CAD building drawing are recognized, the size numerical value of the components is obtained, and the CAD building drawing recognition andthree-dimensional BIM model reconstruction efficiency is improved.

The invention discloses a text interaction matching method and device for financial knowledge questions and answers, belongs to the field of natural language processing, and provides a method for mapping user questions to standard questions in order to accurately judge the matching degree of the financial knowledge questions of a user and the standard financial knowledge questions, and the methodcomprises the following steps: 1, constructing a question pair knowledge base; s2, constructing a question pair matching model training data set; s3, constructing a question pair matching model, comprising the following steps: S301, constructing a character mapping conversion table; s302, constructing an input layer; s303, constructing a character vector mapping layer; s304, constructing a text coding model of an attention mechanism; s305, constructing a text interaction matching layer; and S4, training a question pair matching model and question selections with the intention. (2) The devicecomprises a question pair knowledge base construction unit, a question pair matching model training data set generation unit, a question pair matching model construction unit and a question pair matching model training unit.

The invention discloses a multi-visual-angle action recognition method. The method includes the two processes of action training and action recognizing. In the action training process, a two-dimensional conditional random field method is used for training a classifier. The action recognition process includes the following steps that space-time interest points are extracted; feature descriptors are calculated; dimension reducing is performed on the feature descriptors; the feature descriptors are clustered to obtain a preprocessing file; the preprocessing file is sent into the classifier obtained in the training process. The space-time relation among the space-time interest points is sufficiently utilized, and features among different actions are effectively described; K-means clustering is adopted for clustering different actions into different categories, and therefore the distinction degree of action recognition is increased; through introducing the two-dimensional conditional random field, the time action sequence of a single camera and a space action sequence among multiple cameras are effectively modeled, so that a training model is more accurate, and then actions of the human body are effectively recognized.

The invention discloses an aspect-level text sentiment classification method and system, and the method comprises the steps: extracting the long-distance dependence features of a sentence text according to the obtained local feature vectors of the sentence text, and obtaining the context feature representation of the sentence text; constructing a syntactic dependency relationship among words in the sentence text according to the context feature representation of the sentence text to obtain aspect-level feature representation of the sentence text; and constructing a dependency tree-based graphattention neural network, and obtaining aspect-level emotion categories of the text according to aspect-level feature representation of the sentence text. The method comprises the steps of extractinglocal feature information in a sentence by adopting a convolutional neural network, learning pooled features of the convolutional neural network by utilizing a bidirectional long-short-term memory network, obtaining context information of the sentence, constructing a dependency tree-based graph attention network model, and modeling a sentence dependency relationship by utilizing syntactic information of a dependency tree, thereby improving the performance of sentiment classification.

The invention relates to a cross-media sequencing method based on a multi-depth network structure. The method comprises the following steps of 1, building a cross-media data set including a plurality of media types, and extracting feature vectors of all media data; 2, training the multi-depth network structure by using the cross-media data set, and using the trained multi-depth network structure for unified expression of study of different media data; 3, using the trained multi-depth network structure to obtain the unified expression of different media data so as to calculate the similarity of different media type data; and 4, taking each datum of each media type to be used as an inquiry sample, retrieving data in another media, calculating the similarity of the inquiry sample and the inquiry sample, performing sequencing according to the sequence from high similarity to low similarity, and obtaining a result sequencing table of target media data. The method provided by the invention has the advantages that various network structures are used in a combined way; associated information between the media and inside the media can realize modeling at the same time; further, the unified expression study is performed by using two stages of networks; and the accuracy rate of the cross-media sequencing is improved.

The invention provides a vascular plaque composition recognition method (300) based on a multi-contrast magnetic resonance image. The method comprises the following steps: implementing composition labeling (S310) on the to-be-trained multi-contrast vascular plaque magnetic resonance image; inputting the labeled multi-contrast vascular plaque magnetic resonance image into a convolutional neural network and implementing network model training (S320); and inputting the to-be-recognized multi-contrast vascular plaque magnetic resonance image into a trained network model and predicting the image, so as to output a vascular plaque composition recognition result (S330). According to the vascular plaque composition recognition method, the multi-contrast vascular plaque magnetic resonance image undergoes learning and modeling via the convolutional neural network, so that a new sample can be effectively recognized to assist a doctor in a diagnosis process, and working efficiency of the doctor can be greatly improved. The technical scheme can be conveniently promoted to magnetic resonance image assisted diagnosis processes of other organs.

The invention discloses an information physical fusion system modeling method based on SysML / MARTE. A SysML subset and a MARTE subset are extracted and used for modeling continuous behaviors, random behaviors and nonfunctional properties of a system, and the purpose is to construct a model of the information physical fusion system in a multi-view modeling mode. The modeling method includes the following concrete implement steps that system requirements are analyzed; according to the system requirements, needed modeling elements are selected from the extracted SysML subset and the extracted MARTE subset; an expanded requirement diagram is used for defining property constraint needing to be met to achieve requirement modeling; the extracted SysML modeling elements are used for modeling an architecture of the system to achieve architecture modeling; the extracted MARTE modeling elements are used for modeling the continuous behaviors, the random behaviors and the nonfunctional properties of the system. The SysML / MARTE modeling element expanded subset is constructed, the information physical fusion system is modeled on the well-defined semantic basis, and an effective modeling mode is provided for designing and developing the information physical fusion system.

The invention discloses a question pair matching method and device based on a deep feature fusion neural network, belongs to the field of natural language processing, and aims to solve the technical problem of how to accurately judge the matching degree of a user question and a standard question and sort out a complete question pair matching model. The technical scheme is that the method comprisesthe following steps: S1, constructing a question pair knowledge base; S2, constructing a question pair matching model training data set; S3, constructing a question pair matching model, comprising the following steps: S301, constructing a character mapping conversion table; S302, constructing an input layer; S303, constructing a character vector mapping layer; S304, constructing a neural networkcoding layer based on depth feature fusion; S305, constructing a text similarity matching layer; and S4, training the question pair matching model and selecting a standard problem. The device comprises a question pair knowledge base construction unit, a question pair matching model training data set generation unit, a question pair matching model construction unit and a question pair matching model training unit.

The invention discloses an intelligent electricity terminal plug and play method based on self recognition. The intelligent electricity terminal plug and play method based on the self recognition includes: connecting an intelligent electricity terminal into an intelligent electricity utilization automatic system; using a monitor master station of the intelligent electricity utilization automatic system to judge whether to follow an IEC61850 modeling standard, if yes, not acting, or if no, performing functional decomposition according to the IEC61850 modeling standard so as to obtain a logic device, a logic node and a data object according to function definition of the intelligent electricity terminal; building an IED information model; performing configuration through an IED configuration tool and a system configuration tool according to the IED information model so a to obtain ICD, CID and SCD configuration files, and then completing the configuration for IED; using wed service for mapping, and using an HTTP (hyper text transport protocol) to achieve communications. Accordingly, the intelligent electricity utilization automatic system can obtain login information of the intelligent electricity terminal, and a plug and play function of the intelligent electricity terminal and information interaction between the intelligent electricity terminal and the intelligent electricity utilization automatic system are achieved.

The invention relates to a multi-modal emotion recognition method. The method comprises the steps of extracting frame-level audio features, frame-level video features and word-level text features respectively; respectively inputting the extracted features into a feature encoder for modeling to obtain encoded audio encoding, video encoding and text encoding features; modeling an interaction relationship in a modal by using coded features through respective self-attention modules, sorting and combining the interaction relationships in pairs, and inputting the sorted and combined interaction relationships into a cross-modal attention module to model the interaction relationship between every two modals; performing time sequence pooling on output of the self-attention module and the cross-modal attention module to obtain global interaction features in all modals and global interaction features between every two modals; and respectively carrying out weighted fusion on the global interactioncharacteristics in the modals and between the modals by utilizing an attention mechanism to obtain characteristic representations in the modals and between the modals of the whole sample to be detected, and splicing the two to be detected to obtain a final emotion classification result through a full connection network.

The invention discloses a cross view angle face recognition method based on structuralized dictionary domain transfer. The method comprises the steps that S1, trained sub-dictionaries having discrimination performance on sample categories are connected in series to form a structuralized source domain dictionary; S2, a target domain and a plurality of intermediate domain dictionaries are learned; S3, image face codes of the source domain and the target domain, the source domain dictionary, the target domain dictionary and the intermediate domain dictionaries are calculated, source domain reconstruction images of face images of the source domain and the target domain, target domain reconstruction images and intermediate domain reconstruction images are obtained and connected in series to form the domain sharing characteristic of the face images of the source domain and the domain sharing characteristic of the face images of the target domain; S4, according to the domain sharing characteristic of the face images of the source domain, a support vector machine model is trained for each sample category in a face set of the source domain, the domain sharing characteristic of the face images of the target domain is input into the support vector machine models of all categories, and the category corresponding to the support vector machine model with the highest score is defined as the category of the face images of the target domain.

The invention discloses an automatic problem solving method for an application problem based on a graph neural network, and the method comprises the steps: firstly employing a cyclic neural network tocode an inputted application problem text, constructing a numerical value unit graph and a numerical value comparison graph, and enabling the output (word-level representation) of the cyclic neural network to serve as a node feature; inputting node features and two constructed graphs into a graph neural network-based encoder together to learn graph representation features of questions, so that the final graph features can contain text relationships and size information of numerical values; using one pooling item for aggregating the graph features of different groups into one, so that the output of the graph converter is obtained; finally, using the output graph features as inputs to a tree structure-based decoder to generate a final solution expression tree. According to the method, the task performance is improved through numerical representation in rich problems, and a better problem solving effect can be achieved.

The invention discloses an unmanned aerial vehicle path planning method based on a parallel heuristic algorithm. The unmanned aerial vehicle path planning method based on the parallel heuristic algorithm uses a mechanism combining global search and local optimization to push the algorithm to make a cost function generate a step-down tendency as the iteration progresses; so that the algorithm can search for a better solution of a cell in a global environment, and can continue local mining around the searched better solution; therefore, the diversity and accuracy of the algorithm is improved.

The invention discloses an acoustic scene identification method based on data enhancement. The method comprises the following steps: firstly, collecting and marking audio samples of different sound scenes; then preprocessing is carried out, and pre-emphasis, framing and windowing processing are carried out on the audio samples; data enhancement is then performed, extracting a harmonic source and an impact source of each audio sample to obtain more sufficient audio samples, extracting logarithmic Mel filter bank characteristics from the audio samples and the harmonic sources and the impact sources of the audio samples, stacking the three characteristics into a three-channel high-dimensional characteristic, and constructing more abundant training samples by adopting a hybrid enhancement technology; and finally, inputting the three-channel high-dimensional features into an Xception network for judgment, and identifying the sound scene corresponding to each audio sample. According to the data enhancement method, the generalization capability of the Xception network classifier can be effectively improved, and the training process of the network is stabilized. When the acoustic scene isidentified, the method can obtain a better identification effect.

The invention discloses a layered structure-based dynamic plasma sheath electron density modeling method. The method comprises the steps of dividing a plasma sheath into multiple layers, wherein each layer can be approximately considered to be uniform; performing modeling on time-varying electron density of each layer; and finally synthesizing a space-time electron density matrix. According to the method, a dynamic random process of electron density change is established by utilizing a classical theory, so that the difficulty in the condition that data cannot be obtained from a flight experiment is overcome; plasmas are equivalent to a layered structure, and the electron density of each layer is modeled, so that high difficulty and high complexity in direct modeling of space-time two-dimensional electron density are avoided; for dynamic plasmas in different conditions, the space-time electronic densities of the dynamic plasmas can be modeled by using a unified modeling method; and theoretical and data support is provided for research on electromagnetic wave propagation in the dynamic plasmas, influence of the dynamic plasmas on signals of different systems and the like.

The invention provides an SAR object identification method based on range profile time-frequency diagram non-negative sparse coding. The method utilizes non-negative sparse coding, and in the whole identification process, SAR image objects do not need azimuth angle estimation, thereby reducing the identification complex degree, avoiding the dependence of identification accuracy on object azimuth angle estimation, and helping to improve the object identification rate; and meanwhile, radar object identification is carried out based on the range profile time-frequency diagram non-negative sparse coding technique, and good identification performance is also achieved in the noise environment, thereby not influencing identification effect when not-high image quality, due to factors of defocusing or signal to noise ratio and the like during the object moving, is caused, and helping to improve the robustness performance of the radar object identification.

The invention provides a gas supply and demand dynamic prediction system for steel enterprises and a method thereof, belonging to the technical field of gas prediction in steel industry. The system comprises a data acquisition subsystem, a data processing subsystem, a data modeling subsystem, a model verification subsystem, a model application subsystem and a computer network connecting each subsystem. The gas supply and demand dynamic prediction method comprises the following steps: abstracting generation and consumption characteristics of gas; qualitatively analyzing the generation and consumption characteristics of the gas; dynamically predicting and modeling in a sectional short-term method; and dynamically predicting the supply and demand of the gas. The system and the method of the invention have the advantages of proposing the sectional short-term dynamic prediction and modeling method which is based on historic statistic data, combines the production state information and comprehensively applies real-time data, the production state data and production process technical data. The invention provides decision-making support for distributing staff in short-term gas distribution, and simultaneously provides basic data for the implementation of various intelligent distribution schemes.

The invention discloses an advertisement click rate estimation method based on an improved Transformer, and the method is characterized in that the method comprises the steps: obtaining a historical behavior record of a user, constructing a historical click sequence of the user, and obtaining a target advertisement feature vector, a context feature vector and a user portrait feature vector; inputting into an embedded layer, and obtaining a corresponding embedded vector through an Embedded technology of the embedded layer; and inputting the embedded vector of the historical click sequence of the user into an improved Transformer network, carrying out improved coding on the article position of the click sequence of the user, extracting the historical interest of the user, and extracting theembedded vector of the historical interest of the user and the embedded vector of the target advertisement through an attention mechanism by adopting Sampleoff supervision interest; otaining user historical interests after target advertisement relevancy weighting; and splicing the weighted embedded vectors of the user historical interests, the target advertisement features, the context features and the user portrait features, then inputting the spliced embedded vectors into a subsequent multilayer perceptron, and obtaining an estimated advertisement click probability through a softmax activation function.

The invention provides a loose point cloud data segmentation method and device, discloses a point cloud data processing method and device, relates to radar point cloud data segmentation and especially relates to loose point cloud data segmentation. The segmentation method comprises the following steps: S101, inputting point cloud data A and an initial subclass B of inner points, wherein the subclass B of the inner points is also a subclass of the point cloud data A, and the point cloud data A is a set of points with space correlation, formed by radar scanning; S102, calculating a model b most conforming to the subclass B through a random sample consistency algorithm; and S103, taking all points, which do not belong to the subclass B, in the point cloud data A as test points a, testing the model b by use of the test pints a, and dividing the test points into the inner points, outer points and unknown points according to a test result. The segment method provided by the invention can be applied to processing point cloud data established on standard Cartesian coordinates, data compression operation can be selected, and the operation efficiency is improved.

The invention relates to the technical field of artificial intelligence, and discloses a face recognition model training method, device and equipment and a storage medium, and the method comprises the steps: employing a plurality of first training samples to train an initial three-dimensional face texture model generation module, and obtaining a target three-dimensional face texture model generation module; based on a target three-dimensional face texture model generation module, obtaining a to-be-disturbed three-dimensional face texture model according to a target second training sample obtained from the plurality of second training samples; performing illumination and texture disturbance and image generation according to the to-be-disturbed three-dimensional face texture model, a preset image projection and texture mapping model and a preset projection angle set to obtain a to-be-trained two-dimensional face image set; and training the initial face recognition model according to the to-be-trained two-dimensional face image set and the face image calibration value to obtain a target face recognition model. Effective modeling aiming at the change of illumination and expression textures is realized, and the robustness of the model to disturbance is improved.

The invention discloses a triple non-stationary wireless communication channel modeling method under space-time consistency, which organically integrates a low-complexity parametric method and a high-precision geometric method by capturing smooth and consistent evolution of clusters in an environment in a space-time domain and path gains related to modeling frequency. Smooth and consistent cluster evolution, soft switching of cluster power and effective modeling of frequency-related path gains are realized, the problem that an existing channel model cannot model space-time-frequency non-stationary characteristics of a channel under space-time consistency is solved, and the accuracy and universality of constructing the channel model are improved. The statistical channel model under the millimeter-wave large-scale MIMO high-dynamic scene, which is constructed by adopting the method disclosed by the invention, has the characteristics of high accuracy, simplicity and feasibility, can support reasonable design and performance analysis of a communication and inductance integrated system, and meanwhile, provides a real and reliable simulation verification platform for a wireless communication network system-level algorithm.

The invention discloses a social group recommendation method, system and device, and a storage medium, which can directly and automatically calculate the social influence of a user from a user social network, and enhance the accuracy and propagation ability of group recommendation by using the social influence of the user. The deeper feature representation of the user can be obtained by using a user influence diffusion method, and meanwhile, the influence of each user in the decision process is learned by using an attention mechanism when the preference of the group is modeled, so that the group feature representation is more effectively obtained; and finally, single user recommendation and group user recommendation tasks are optimized at the same time through a combined learning mode, the performance of a social group recommendation model is improved, the accuracy of group recommendation is improved, and the propagation capability of group recommendation is improved.

The invention discloses a breast cancer sentinel lymph node metastasis prediction method and system and a storage medium, and the method comprises the steps: obtaining a WSI with a label as a training data set, and carrying out the preprocessing, and obtaining an image block set; constructing a WSI classification model; pre-training a feature extractor by using the image block set to obtain a feature vector set; inputting the feature vector set into a prototype clustering module, and extracting a plurality of prototypes through clustering; the breast cancer sentinel lymph node WSI is divided into image blocks, and then the image blocks are input into a feature extractor to extract image block features; matching the image block features with a prototype input feature fusion module, generating a soft distribution histogram, and constructing a feature vector of breast cancer sentinel node WSI; and sending the feature vector of the breast cancer sentinel lymph node WSI into a full connection layer to obtain a WSI classification score, and performing transfer judgment. The method can better solve the problem of micro-metastasis identification while maintaining accurate identification of macro-metastasis, so that breast cancer sentinel lymph node metastasis can be accurately diagnosed.

The invention discloses a speech enhancement method and system fusing ultrasonic signal features. The method comprises the steps: firstly predefining an ultrasonic signal, and then actively transmitting and receiving the ultrasonic signal through a loudspeaker and a microphone of a device; and performing channel estimation to obtain a channel impact response, and inputting the channel impact response to a neural network to realize speech enhancement, wherein the channel impact response reflects motion features of facial vocal organs when a user speaks, and the motion features serve as supplementary modal information of speech. According to the invention, the speech enhancement task is assisted by making full use of the user voice action characteristics, the speech enhancement effect is improved, and the method has wide application prospects.

The invention discloses a method for realizing fashion suit recommendation through a graph neural network, and the method comprises the steps: constructing a network structure comprising a user node,a suit node and a single-item node, initializing the vector representation of each node, and constructing the relation between different nodes through a connection edge; realizing information transmission among the single items by utilizing classification of the single items, so that each single item contains matching information with other single items, and updating of node vector representationof the single items is further realized; updating the suite node vector representation by utilizing the updated plurality of single-item node vector representations; updating the user node vector representation by using the updated suit node vector representation, and calculating the preference score of the user for each suit by using the updated user node vector representation and the suit node vector representation; and sorting the suites according to the preference scores so as to recommend the suites to the corresponding users. According to the method, the complex interaction information among the user, the suits and the single items can be effectively modeled, and the recommendation performance is improved.

The invention discloses a generation type-based auxiliary template enhanced garment matching scheme generation method and a system. The method comprises the steps of constructing a generation type-based auxiliary template enhanced garment matching model; constructing a training set; inputting the training set into the constructed auxiliary template enhanced garment matching model based on the generative mode for training to obtain a trained auxiliary template enhanced garment matching model based on the generative mode; inputting an upper garment to be matched into the trained auxiliary template enhanced garment matching model based on the generative mode, and outputting the most matched lower garment; and outputting the upper garment to be matched and the most matched lower garment as a final garment matching scheme.

The invention discloses a false news detection method based on a knowledge perception attention network, and belongs to the technical field of artificial intelligence. The method comprises the steps that related knowledge in a knowledge graph is extracted based on news texts, the news texts and the extracted related knowledge serve as input data, a false news detection model based on knowledge perception is constructed, and news samples are classified. Firstly, entity mentions in news are recognized through entity links and aligned with corresponding entities in a knowledge graph, and an entity sequence is obtained; secondly, for each entity in the entity sequence, obtaining a neighbor entity of the entity in the knowledge graph as an entity context of the entity; and finally, through a false news detection model, fusing the news text with the entity and the entity context features to complete false news detection. According to the method, the ambiguity problem caused by entity mentionin news texts can be solved, and meanwhile, news representations of supplementary information, learning semantic level and knowledge level can be provided for entities in news.

The invention discloses a video statement positioning method based on a multi-stage aggregation Transformer model, and in the video statement Transformer model, each video slice or word can adaptivelyaggregate and align information from all other video slices or words in two modes according to the semantics of the video slice or word. Through multi-layer superposition, finally obtained video statement joint representation has rich visual language clue capture capability, and finer matching can be realized. In the multi-stage aggregation module, the stage characteristics of the starting stage,the stage characteristics of the intermediate stage and the stage characteristics of the ending stage are connected in series to form the characteristic representation of the candidate segment. Because the obtained characteristics representation captures the specific information of different stages, the method is very suitable for accurately positioning the starting position and the ending position of the video clip. The two modules are integrated together to form an effective and efficient network, and the accuracy of video statement positioning is improved.