114 results about "Graph encoding" patented technology

A method for presenting a perspective view of a real urban environment, augmented with associated geo-coded content, and presented on a display of a terminal device. The method comprises the steps of: connecting the terminal device to a server via a network; communicating user identification, user present-position information and at least one user command, from the terminal device to the server; processing a high-fidelity, large-scale, three-dimensional (3D) model of an urban environment, and associated geo-coded content by the server; communicating the 3D model and associated geo-coded content from said server to said terminal device, and processing said data layers and said associated geo-coded content, in the terminal device to form a perspective view of the real urban environment augmented with the associated geo-coded content. The 3D model comprises a data layer of 3D building models; a data layer of terrain skin model; and a data layer of 3D street-level-culture models. The processed data layers and the associated geo-coded content correspond to the user present-position, the user identification information, and the user command.

Techniques for encapsulating video streams containing multiple coded views in a media file are described herein. In one example, a method includes parsing a track of video data, wherein the track includes one or more views. The method further includes parsing information to determine whether the track includes only texture views, only depth views, or both texture and depth views. Another example method includes composing a track of video data, wherein the track includes one or more views and composing information that indicates whether the track includes only texture views, only depth views, or both texture and depth views.

The invention discloses a 3D point cloud semantic segmentation method under a bird's-eye view coding view angle. The method is applied to an input 3D point cloud. The method comprises: converting a voxel-based coding mode into a view angle of a bird's-eye view; extracting a feature of each voxel through a simplified Point Net network; converting the feature map into a feature map which can be directly processed by utilizing a 2D convolutional network; and processing the encoded feature map by using a full convolutional network structure composed of residual modules reconstructed through decomposition convolution and hole convolution, so that an end-to-end pixel-level semantic segmentation result is obtained, point cloud network semantic segmentation can be accelerated, and a point cloud segmentation task in a high-precision real-time large scene can be achieved under the condition that hardware is limited. The method can be directly used for tasks of robots, unmanned driving, disordered grabbing and the like, and due to the design of the method on a coding mode and a network structure, the system overhead is lower while high-precision point cloud semantic segmentation is achieved,and the method is more suitable for hardware-limited scenes of robots, unmanned driving and the like.

Techniques for encapsulating video streams containing multiple coded views in a media file are described herein. In one example, a method includes parsing a track of multiview video data, wherein the track includes at least one depth view. The method further includes parsing information to determine a spatial resolution associated with the depth view, wherein decoding the spatial resolution does not require parsing of a sequence parameter set of the depth view. Another example method includes composing a track of multiview video data, wherein the track includes the one or more views. The example method further includes composing information to indicate a spatial resolution associated with the depth view, wherein decoding the spatial resolution does not require parsing of a sequence parameter set of the depth view.

Provided are a multi-view video decoding apparatus and method and a multi-view encoding apparatus and method. The decoding method includes: determining whether a prediction mode of a current block being decoded is a merge mode; when the prediction mode is determined to be the merge mode, forming a merge candidate list including at least one of an inter-view candidate, a spatial candidate, a disparity candidate, a view synthesis prediction candidate, and a temporal candidate; and predicting the current block by selecting a merge candidate for predicting the current block from the merge candidate list, wherein whether to include, in the merge candidate list, at least one of a view synthesis prediction candidate for an adjacent block of the current block and a view synthesis prediction candidate for the current block is determined based on whether view synthesis prediction is performed on the adjacent block and the current block.

Techniques for encapsulating video streams containing multiple coded views in a media file are described herein. In one example, a method includes parsing a track of video data, wherein the track includes one or more views. The method further includes parsing information to determine whether a texture view or a depth view of a reference view is required for decoding at least one of the one or more views in the track. Another example method includes composing a track of video data, wherein the track includes one or more views and composing information that indicates whether a texture view or a depth view of a reference view is required for decoding at least one of the one or more views in the track.

A method and device for coding screen content into a bitstream by selecting a color palette table for a coding unit (CU) of screen content, creating a color index map having indices for the coding unit (CU), and encoding the selected color palette table and the color index map for the CU into a bitstream.

Techniques for encapsulating video streams containing multiple coded views in a media file are described herein. In one example, a method includes parsing a track of multiview video data, wherein the track includes one or more views, including only one of a texture view of a particular view and a depth view of the particular view. The method further includes parsing a track reference to determine a dependency of the track to a referenced track indicated in the track reference. Track reference types include ‘deps’ that indicates that the track includes the depth view of the particular view and the reference track includes the texture, ‘tref’ that indicates that the track depends on the texture view which is stored in the referenced track, and ‘dref’ that indicates that the track depends on the depth view which is stored in the referenced track.

The invention discloses an unsupervised target detection method and a system based on a variational auto-encoder and a Gaussian mixture model, and the method comprises the steps: converting an input image into a feature map of H * W dimensions, namely H * W cells, through a backbone network, coding the feature map into a hidden variable of which the prior distribution accords with the Gaussian mixture model, and carrying out the image reconstruction through a decoder according to the hidden variable, and comparing the reconstructed image with the input image, and calculating a loss function, so that a neural network is trained, an encoder obtains information such as the category and the position of an object in the image, and unsupervised target detection is realized. According to the method, the space attention mechanism and the Gaussian mixture model are combined, end-to-end target detection and classification can be achieved, meanwhile, good performance is still achieved under the condition that a large number of objects exist, and good expansibility is achieved.

The invention discloses an abnormality detection method based on attribute graph representation learning. The method comprises the following steps: acquiring an attribute graph data set; for the similarity between the nodes in the attribute graphs, expanding an attribute graph topological structure in the data set; importing the topological structure data in the attribute graphs into a TransE module to obtain an embedded vector set of nodes; taking the expanded attribute graph data set and the embedded vector set obtained in the previous two steps as input, and operating a coding module to carry out attribute graph coding; performing structure reconstruction decoding on a coded data set obtained by coding; carrying out attribute reconstruction decoding on the coded data set obtained by coding; and predicting and sorting abnormality nodes according to structure reconstruction error and attribute reconstruction error obtained by coding and decoding. According to the method, the problem that node attributes and an attribute graph topological structure are not closely associated is solved. The detection performance of the abnormality detection method based on attribute graph representation learning is significantly improved compared with the performance of the abnormality detection method based on graph convolution in the prior art.

Embodiments of the present invention provide a system for fast parallel graph compression based on identifying a set of large cliques, which is used to encode the graph. The system provides both permanently-stored and in-memory graph encoding and reduces the space needed to represent and store a graph, the I / O traffic to use the graph, and the computation needed to perform algorithms involving the graph. The system thereby improves computing technology and graph computation. During operation, the system obtains data indicating vertices and edges of a graph. The system executes a clique-finding method to identify a maximum clique in the graph. The system then removes the clique from the graph, adds the clique to a set of found cliques, and generates a compressed representation of the graph based on the set of found cliques.

The invention relates to a social media rumor detection method and system based on a hierarchical heterogeneous graph neural network, and the system comprises: a data preprocessing module which is used for preprocessing text data, extracting user static features, and constructing a hierarchical heterogeneous graph structure; an event encoding module which is used for encoding a text containing propagation and diffusion structure information; a user coding module which is used for learning user behavior features; a global heterogeneous graph coding module which is used for capturing rich global structure information between the event and the user; and a rumor detection label output module which is used for fusing the text information, the user behavior features and the global heterogeneous graph information to complete label prediction work of rumor detection. According to the method, local inline relationships between users and between texts can be effectively learned, user and text representations containing adjacent node information are generated, a global structure relationship between the user and an event is modeled and learned, and finally the authenticity of the event is identified.

The invention relates to the technical field of semantic synthesis, and discloses a speech synthesis method and device, computer equipment and a computer readable storage medium. The method comprisesthe steps: acquiring a to-be-synthesized text, and converting the to-be-synthesized text into graph embedding vector information through a speech synthesis model, encoding the graph embedding vector information according to a graph encoder to generate corresponding first intermediate vector information, generating corresponding Mel language spectrum information according to the first intermediatevector information, and outputting voice synthesis information corresponding to the Mel language spectrum information. The specific semantic information of the text information is analyzed through thegraph auxiliary encoder to be mapped to different voice rhythms, so that the rhythm adjustment process becomes a full-automatic process, and the voice synthesis accuracy is improved. Meanwhile, the invention also relates to a blockchain technology, and the method can be applied to the fields of smart government affairs, smart education, smart medical treatment and the like, thereby further promoting the construction of smart cities.

The invention discloses a dialogue generation method fusing knowledge. The dialogue generation method comprises the following steps that S1, a knowledge graph is constructed; and S2, a dialogue modelis constructed, the dialogue model is composed of an encoder and a decoder, and the construction process comprises word encoding, knowledge encoding, double-hop entity encoding, weighted merging and decoding. According to the method, graph coding and a graph attention mechanism are introduced to carry out double-hop entity coding, and entity semantics in dialogue are better captured based on the relationship between adjacent entities; and meanwhile, in combination with concept knowledge involved in dialogue surrounded by the knowledge graph, a more reasonable reply rich in information is given, and the problem of topic concept drifting and extension in the current dialogue process is solved.

The invention provides cross-language unsupervised classification with multi-view transfer learning. Embodiments of an unsupervised cross-language sentiment classification model, which may be referred to as a multi-view encoder classifier (MVEC), utilizing an unsupervised machine translation (UMT) system and a language discriminator are presented. Unlike previous Language Model (LM)-based fine tuning methods that adjust parameters based only on classification errors of training data, embodiments employ an encoder-decoder framework of UMT as a regularization component on shared network parameters. In one or more embodiments, a cross-language encoder of an embodiment learns a shared representation, which is effective for reconstructing input sentences of both languages and generating more representative views from the input for classification. Experiments on five language pairs prove that the embodiment of the MVEC is obviously superior to other models for 8 / 11 sentiment classification tasks.

A computer-implemented method for securing software installation through deep graph learning includes extracting a new software installation graph (SIG) corresponding to a new software installation based on installation data associated with the new software installation, using at least two node embedding models to generate a first vector representation by embedding the nodes of the new SIG and inferring any embeddings for out-of-vocabulary (OOV) words corresponding to unseen pathnames, utilizing a deep graph autoencoder to reconstruct nodes of the new SIG from latent vector representations encoded by the graph LSTM, wherein reconstruction losses resulting from a difference of a second vector representation generated by the deep graph autoencoder and the first vector representation represent anomaly scores for each node, and performing anomaly detection by comparing an overall anomaly score of the anomaly scores to a threshold of normal software installation.

An Information supply method for displaying a message received or generated by a first electronic device on a display of a second electronic device. The information supply method includes the steps of: receiving a request of displaying the message; generating a word string according to the request, wherein the word string includes at least one word pattern and at least one graph encoding correspondingly representing the at least one word pattern; transmitting the word pattern and the graph encoding from the first electronic device to the second electronic device; and receiving the at least one word pattern and the at least one graph encoding, and displaying the at least one word pattern corresponding to the message received or generated by the first electronic device on the display according to the graph encoding.

The invention provides an information processing model generation method based on target attribute decoupling and related equipment, and the method comprises the steps of obtaining feature maps outputby a hidden layer, carrying out the coding of the feature maps through Hash codes, and obtaining coordinate values corresponding to all feature maps; clustering the feature maps according to the coordinate values to obtain a feature map group, respectively calculating orthogonal loss and / or inhibition loss corresponding to the feature graphs in each feature graph group, obtaining a total loss value of the model according to the calculated orthogonal loss and / or inhibition loss, adjusting model parameters by utilizing the total loss value of the model, and repeating the above steps until the training is completed, so as to obtain a generated information processing model. According to the method provided by the embodiment of the invention, the attribute coupling is reduced by mining the semantic attributes of the latent layer and constructing the orthogonal loss of the clustering group, and the cross coupling according to the attributes is reduced by performing intersection suppressionon the feature maps in the intersection region, so that the attribute coupling between the feature maps is reduced, and the generalization ability of the network is improved.

The invention discloses a convolutional neural network accelerator based on feature map sparsity. The convolutional neural network accelerator comprises an input feature map coding module, a weight coding module, a data flow control module, a sparse matrix calculation module and a bus. The bus is respectively connected with the data flow control module, the input feature map coding module, the weight coding module and the sparse matrix calculation module; the input feature map encoding module encodes the feature map according to the 0 element in the feature map; the weight coding module provides corresponding weight data for the sparse matrix calculation module according to coding information input into the feature map coding module; the data flow control module controls the working modes of other modules according to the register information; and the sparse matrix calculation module performs convolution calculation by using data provided by the input feature map coding module and the weight coding module. According to the method, the accelerator can be switched to use the sparsity in the weight, the method can be flexibly applied, and the sparse weight is supported.

The invention discloses a pavement disease detection method based on deep learning. A large number of normal road pictures are input, road disease image samples are not needed, a system result contains two neural networks, the structure of the first neural network is 'encoder-decoder-encoder', the first neural network is mainly responsible for encoding and decoding pictures, the encoding process is to convert the pictures into a group of vector data through the neural network, and finally, the decoded vector is compared with the decoded vector of the original picture. According to the pavementdisease detection method based on deep learning, which relates to the technical field of traffic pavement detection, a road disease sample does not need to be used for analysis, a large amount of manpower and material resources are saved, a plurality of neural networks are used, a large number of normal road pictures are analyzed for coding, decoding and re-coding training, the training quality is judged through a discrimination network, and finally, the difference between coded pictures and original picture codes is analyzed to detect whether a road has diseases or not.

The embodiment of the invention discloses an event extraction method, a related device, equipment and a storage medium, which are used for converting a sentence-level natural language into nodes and edges and then converting the nodes and the edges into semantic features to perform event extraction, so that the accuracy of event acquisition can be ensured. The method of the embodiment comprises: obtaining a to-be-processed text; generating abstract semantic representation according to the to-be-processed text, wherein the abstract semantic representation comprises nodes in one-to-one correspondence with the words and edges used for connecting the nodes; carrying out semantic coding processing on the abstract semantics and the text representation to obtain a semantic embedding vector, wherein the semantic embedding vector is used for representing semantic features between each word and the event; carrying out graph coding processing on the abstract semantic representation to obtain a graph embedding vector, wherein the graph embedding vector is used for representing structural features of nodes connected through edges; splicing the semantic embedding vector and the graph embedding vector to obtain a spliced feature vector; and identifying the spliced feature vector, and outputting a target event.

The invention belongs to the field of computer artificial intelligence of natural language generation, and discloses a multi-round dialogue reply generation system and method based on a relational graph attention network. Comprising the steps of obtaining multi-round dialogue input content for preprocessing, obtaining semantic information representation of each round of utterance, and encoding statement semantic information of each round of utterance to obtain semantic representation of a dialogue context; then, a graph attention network is adopted to capture autocorrelation in multiple rounds of dialogues and related features among dialogues, and relation position codes are introduced into the graph attention network to illustrate sequence information containing utterances, so that high-level semantic representation of a graph coding layer is obtained; and finally, taking the context semantic information representation of the dialogue and the advanced semantic representation of the attention coding of the relational graph as input, and decoding by using a GRU model to generate a final dialogue reply output representation. According to the method, the generation quality of the multi-round dialogue reply is remarkably improved, and the generated reply is more coherent and meaningful.

The invention discloses a textile material CT image segmentation method and device based on a convolutional neural network. The method comprises the steps: firstly building a segmentation model comprising an encoder and a decoder; secondly, acquiring a CT image of the textile material to be segmented; inputting the data into a trained segmentation model; performing feature extraction on the inputimage to obtain an encoding feature map via the encoder; inputting the encoding feature maps of the multiple levels into the corresponding levels of the decoder for feature fusion to obtain a first fusion feature map via the encoder; decoding the first fusion feature map to obtain a decoded feature map via the decoder; fusing the decoding feature maps of the middle level and the deep level to obtain a second fused feature map. The textile material CT image segmentation method solves the technical problems that in an existing textile material CT image segmentation method, manual segmentation isadopted, the segmentation difficulty is large, the segmentation process is tedious, time and labor are wasted, the segmentation result depends on experience and knowledge of operators to a large extent, and the segmentation result is difficult to reappear.

The invention discloses a motion blur restoration method based on a contour enhancement strategy, and relates to the technical field of image processing, and the method comprises the following steps:(1) encoding an original blur image, removing noise information through fine adjustment of a restoration network, and finally restoring a clear contour image through a decoder; respectively carrying out contour extraction on the image by using a Sober operator and a Canny operator, forming a sequence with the contour recovered by the decoder, carrying out further restoration extraction on the image contour by using LSTM, and finally generating a sharp edge; (2) respectively sampling and encoding the original fuzzy graph and the sharp edge, pairing the graph codes and the sharp edge codes withthe same size one by one, and outputting graph code and sharp edge code pairs; (3) generating a potential clear graph by using a multi-scale repair framework; according to the invention, the sharp edge in the image with serious motion blur can be extracted, so that the generated sharp edge effectively assists a multi-scale framework in removing the motion blur, and the efficiency of removing the motion blur is effectively improved.

The invention discloses a self-supervised learning rumor detection method and system, equipment and a storage medium, and the related method comprises the steps: modeling a message propagation process as a propagation tree, employing a graph encoder to encode the graph structure information of the propagation tree, and obtaining a propagation representation; encoding text content information of the message by adopting a text encoder, and obtaining text representation; on the basis of a self-supervised learning mode, improving the similarity between the propagation representation and the text representation of the same message, and reducing the similarity between the propagation representation and the text representation of different messages; and performing rumor detection by using the message representation obtained based on the self-supervised learning mode. According to the scheme, information of multiple sources is synthesized through self-supervised learning, more comprehensive expression of the message is obtained, and then the accuracy and stability of a downstream automatic rumor detection task are improved.

The invention discloses a depth unsupervised single cell clustering method based on a Gaussian mixture graph variational auto-encoder. The method comprises the following steps: initializing a gene regulatory network A by using a protein-protein interaction relationship PPIs (or regulatory element interaction HiChIP); initializing a cell cluster C of each cell by using a K-means method; enabling the gene regulatory network A and single-cell gene expression data X (or regulatory element opening degree data X) to pass through a graph encoder to obtain a hidden layer; obtaining a cell cluster C, and sampling from a Gaussian mixture model GMM to obtain a cell low-dimensional representation Z; predicting a gene regulatory network A by using a decoder; calculating a loss function, performing back propagation to update A and GCN, and repeating the steps until convergence; and outputting a gene regulatory network A, a cell low-dimensional representation Z and a cell cluster C. Clustering of cells and dimensionality reduction of cell expression are completed in the process of constructing the gene regulatory network A.

The embodiment of the invention provides an method and device for training an interaction prediction model and a method and device for using the interaction prediction model. The training method comprises the steps: firstly, constructing a dynamic interaction graph based on an interaction event sequence,determining a first node and a second node corresponding to a current interaction event from the dynamic interaction graph, wherein the first node points to two historical nodes corresponding to a previous interaction event in which the first node participates; using a graph coding network to obtain coding vectors corresponding to the two historical nodes; and inputting the coding vectors and the respective occurrence moments of the two interaction events into a prediction representation network to obtain a first representation vector of the first node; inputting the attribute characteristics of the second node into an attribute coding network to obtain a second attribute vector; predicting the interaction probability according to the first representation vector and the second attribute vector, determining the loss according to the interaction probability, and updating the graph coding network, the prediction representation network and the attribute coding network.

The invention discloses a sparse graph coding-based hyperspectral remote sensing image eigen decomposition method and system, and relates to the field of image processing. According to the method, the problem that the precision of generating the reflectivity component of the hyperspectral image is low due to the fact that the surface feature boundary cannot be effectively kept when an existing intrinsic decomposition method is applied to the hyperspectral image is solved. The method comprises the following steps: acquiring a hyperspectral remote sensing image; averaging a hyperspectral remote sensing image geometrically in a spectral dimension to remove spectral change caused by geometric distribution on the surface of an object, so that the image is geometrically averaged in a spatial dimension; eliminating spectral change caused by illumination changing along with spatial distribution to obtain an image, and obtaining a sparse graph coding dictionary of each pixel in the hyperspectral remote sensing image; obtaining a similarity matrix of a sparse graph according to the sparse graph coding dictionary of the hyperspectral remote sensing image; obtaining a reflectivity component of the hyperspectral image according to the sparse image similarity matrix; the system comprises an acquisition module, a construction module, a calculation module and a decomposition module.

The invention discloses a human flow mode detection method based on a deep neural network sparse graph encoder, and the method comprises the following steps: collecting city data which comprises humanactivity data, city interest point data and city geographic information data; dividing a city into a plurality of regions by combining an improved OPTICS algorithm and city data; according to a region division result, constructing human activity flow matrixes in multiple time periods, and constructing a human activity tensor matrix based on the human activity flow matrixes; solving structural data of the human activity tensor matrix based on a sparse graph encoder of the deep neural network; and processing the structured data by using a clustering algorithm to obtain a human flow mode in eachscene. The method can be applied to more complex data structures, more effective urban area selection is realized, the structural characteristics of the network are better reflected, the preferencesof the user to the POI in different space-time environments are deduced, the effective detection of the urban human flow mode is realized, and suggestions are provided for urban planning and POI construction.