45 results about "Nearest neighbor graph" patented technology

The invention relates to a point cloud data based single tree three-dimensional modeling and morphological parameter extracting method. The point cloud data based single tree three-dimensional modeling and morphological parameter extracting method comprises obtaining three-dimensional surface point cloud data of high density standing trees through a three-dimensional scanner or other live-action measuring modes, calculating the shortest distance from points to root nodes through a k-nearest neighbor graph, performing hierarchical clustering on the data according to distance, enabling centers of clustering hierarchies to be served as framework points of a limb system and meanwhile extracting corresponding semi-diameter of the framework points; connecting the framework points to establish a topological structure of branches and grading the branches; performing three-dimensional geometrical reconstruction on branches through a generalized cylinder body; adding leaf models to the limb system to form into a vivid three-dimensional single tree model; extracting height of trees, diameter of breast height and crown breadth of the standing trees in the point cloud. The point cloud data based single tree three-dimensional modeling and morphological parameter extracting method can rapidly and semi-automatically extract tree important geometrical parameters and topological information to form into the high vivid single tree geometric model and has wide application prospects and values in fields such as agriculture and forestry survey, ecological research and landscape planning.

The present invention has disclosed a process of image segmentation, which comprises applying edge detection to an image to obtain an edge image and preprocessing the image; oversegmentating the preprocessed image to obtain the plurality of initial partitions; constructing k-NN Graph for the oversegmented image based on the similarity between the initial partitions; and using k-NN Graph to merge the initial partitions. With the present invention, the merging process can be accelerated and the segmentation accuracy can be improved.

The invention discloses a gear fault diagnosis platform and a gear fault diagnosis method. On the platform, gear faults are simulated and vibration signals are collected, the improved local preserving projection algorithm is combined with a Bayes classifier, the mode recognition effect is judged according to the correct classification rate of the Bayes classifier, the vibration signals of the gear faults are measured by a vibration acceleration sensor, the principal component analysis is conducted firstly, and then, kernel transformation, construction of the nearest neighbor graph, solving of mapping space and the like are conducted; and the Bayes classifier is used for recognition in classification according to multi-fault classification. Compared with the principal component analysis, the laplace algorithm and the local preserving projection, the improved local preserving projection fault recognition rate can be greatly increased. According to the improved local preserving projection algorithm and the Bayes classifier combined fault mode recognition method, the fault recognition rate is increased and the accuracy is improved, and the effect of the fault mode recognition of the gear can be improved. The gear fault diagnosis platform has a simple structure, and a high-accuracy diagnosis platform can be provided for the fault recognition of the gear.

The invention discloses an image marking method based on semi-supervised subject modeling. The method comprises the following steps of: firstly, obtaining images from the Internet, including images with text marks and unmarked images; then, modeling the relation between the visual features and text marks of all images through latent subjects by use of a model similar to probabilistic latent semantic analysis; establishing the nearest-neighbor graphs of all images, and adjusting the model according to the manifold structure obtained by modeling the nearest-neighbor graphs; learning the model by an expectation maximization algorithm, and calculating the probability of matching the latent subjects with the images respectively; and finally, calculating the probability of matching each text mark with the unmarked images according to the probability of matching the latent subjects with the images, and selecting the text mark with the highest probability to mark the unmarked images.

The invention discloses a region-based multi-feature fusion high-resolution remote sensing image segmentation method. The method comprises the steps of firstly, performing initial segmentation on an initial high-resolution image; secondly, calculating a texture feature distance, a spectral feature distance and a shape feature distance of any adjacent region in initially segmented regions; and finally, performing region combination based on an RAG (Region Adjacency Graph) and an NNG (Nearest Neighbor Graph). According to the method, a combination rule is established by comprehensively adopting a spectral feature, a texture feature, a shape feature and the like; compared with a rule established by separately adopting a feature, the combination rule better conforms to semantic description of an object, so that the segmentation precision is higher; an adjacency relation of the regions is maintained by jointly adopting two data structures of the RAG and the NNG, so that higher execution efficiency can be obtained by an algorithm; and compared with the prior art, a segmentation result can be obtained more quickly.

The invention relates to a sensor exception detection method based on a regularized vector autoregression model. The sensor exception detection method comprises the following steps: (1) establishing a multielement linear regression model, and determining a target function; collecting data by a sensor; establishing a nearest neighbor graph of data points, wherein the graph consists of n vertexes, each vertex corresponds to one data point, and the weight matrix of edges is defined by considering a relationship between fixed points; constructing a bound term in order to keep similarity among original data points while the obstacles of high dimension and overfitting can be overcome; and utilizing the target function to train a model parameter to obtain an optimal parameter coefficient, and utilizing the model obtained by training to carry out exception detection. The sensor exception detection method can better predict original data.

The invention provides a robust image clustering method. Firstly, the original image data is normalized and the features are extracted. The K-nearest neighbor graph G is constructed for the processedimage data, and the Laplace matrix L is calculated according to the adjacency matrix W and the corresponding degree matrix D of the K-nearest neighbor graph G; The Laplace matrix L is decomposed intoeigenvalues, and the obtained eigenvectors are sorted from small to large according to the corresponding eigenvalues, and p eigenvectors are sequentially selected to form a candidate eigenvector matrix U, and p is the number of clusters; Then, the objective function is constructed according to the Laplace matrix and the candidate eigenvector set, and the objective function is solved to obtain thediscriminant representation Y* of the original image data. Finally, the distinguishing representation Y* is given by k-mean value clustering, the final clustering results are obtained. The invention can cluster not only the simple image data set, but also the complex image data with divorced points.

The invention discloses a k neighbor graph-based rare class detection method, which comprises the following steps of: firstly, constructing a k neighbor graph of a given unlabeled data set S, and automatically selecting a k value by an algorithm; and then, based on the constructed k neighbor graph, giving a definition of a change coefficient Vc, and calculating a change coefficient Vc value of each node in the data set, finding out the node x with the maximum change coefficient from all the nodes, inquiring a labeler to obtain a category label y of the node x, and respectively adding the x andthe y into the selected data sample set I and the selected data sample real category label set L; carrying out rare class detection by utilizing a method for detecting local mutation of data sample distribution in the data set, and compared with other priori-free rare class detection methods, the KRED method is higher in efficiency and lower in algorithm overhead. And meanwhile, through a methodof automatically selecting the k value, the discovery efficiency of each class in the data set is effectively improved, and the inquiry frequency required for discovering all classes in the data set is remarkably reduced.

The invention discloses a method for generating TSV (through-silicon via) interconnection oriented three-dimensional integrated circuit clock topology structure, which comprises the following steps: inputting clock endpoints, a clock source, a buffer library and TSV information of a three-dimensional integrated circuit; circling a large-density area for the clock endpoints on each layer by using a classification algorithm, and establishing a subtree; mapping unclassified clock endpoints on all layers and root nodes of the clock tree established in each classified area to a 2D (two-dimensional) plane; finding the nearest neighbor node of each node by using a method of establishing the nearest neighbor graphs through tube decomposition, and carrying out pairing on the nodes so as to generate a father node according to a nearest distance principle; and determining whether unpaired nodes exist, if unpaired nodes do not exist, inserting the nodes into the buffer library and the TSV information from top to bottom so as to generate a three-dimensional clock topological structure. The method disclosed by the invention ensures the uniform distribution of TSV based on a clock endpoint density classification algorithm, and avoids the over-dense insertion of TSV, thereby increasing the manufacturability and the reliability to some extent.

The invention provides a face clustering method and device based on structure perception, and the method comprises the steps: obtaining a plurality of to-be-processed face images, extracting the face features of each to-be-processed face image based on a pre-trained convolutional neural network model, and constructing a K-nearest neighbor graph according to the face features of each to-be-processed face image; inputting the K neighbor graph into a pre-trained edge score prediction model to obtain the score of each edge in the K neighbor graph, wherein the edge score prediction model is obtained by sampling a K neighbor graph by using a structure retention sub-graph sampling strategy and training a graph convolutional neural network by using a sub-graph obtained by sampling; and according to the score of each edge in the K neighbor graph, performing a first pruning operation on the K neighbor graph to obtain a face cluster for the plurality of to-be-processed face images. The technical problem of insufficient face clustering accuracy in the prior art is solved.

The embodiment of the invention provides a method and device for identifying a user and a computer device. The method comprises the steps of marking different labels for normal users and abnormal users in user samples; determining a similar value between any two users in the user samples; establishing a K nearest neighbor graph for all users in the user samples through utilization of a K nearest neighbor algorithm; determining a weight value of a side formed by the any two users in the K nearest neighbor graph according to the similar value between any two users in the user samples; determining a first weight sum of the sides formed by carrying out connection on each rest user in the K nearest neighbor graph and the normal users, and a second weight sum of the sides formed by carrying outconnection on each rest user in the K nearest neighbor graph and the abnormal users according to a preset K value and the weight value of the side formed by the any two users in the K nearest neighborgraph; determining a label of each residual user according to the first weight sum and the second weight sum; and identifying each user according to the label of the user.

The invention provides a robust structure non-negative matrix factorization clustering method based on graph regularization, and the method comprises the steps: S10, obtaining m to-be-clustered images, and constructing k nearest neighbor graphs according to the to-be-clustered images; S20, a corresponding data matrix Y is obtained for each nearest neighbor graph, the data matrix Y comprises n datapoints, and a non-negative matrix decomposition method is used for decomposing the data matrix Y to obtain a feature matrix W and a coefficient matrix H; S30, establishing an objective function O ofrobust structure non-negative matrix decomposition based on graph regularization based on l2 and p norms; S40, according to the objective function O, using an iterative weighting method to iterate preset times, and updating the feature matrix W, the coefficient item and the graph regular item; S50 analyzing and clustering the feature matrix W obtained by each nearest neighbor graph by using a k-means clustering algorithm. According to the method, a robust loss function is adopted to measure a reconstruction error therein, mark data is not used in the robust loss function for judgment, and after a semi-supervised method of non-negative matrix factorization is introduced, the efficiency and the accuracy rate can be effectively improved.

The invention relates to a local spline embedding-based linear classification method. The method comprises the steps of inputting training data and test data; and performing supervised local spline embedding dimension reduction of the training data: constructing an intra-class graph and an inter-class graph, and selecting neighborhoods; constructing an intra-class local tangent space and an inter-class local tangent space of training data points; obtaining global low-dimensional coordinates of the training data points and obtaining corresponding optimal linear mapping; mapping the test data into a low-dimensional manifold; and classifying the test data subjected to dimension reduction by K-nearest neighbor classification to obtain category labels of the test data. According to the method, the defects that the intra-class nearest neighbor graph and the inter-class nearest neighbor graph of the data are not comprehensively analyzed and category information of the training data is not utilized in the past are overcome; and in addition, the data complexity is greatly lowered, the observability and distinguishability of the data are enhanced, and the classification accuracy of the data on a high-dimensional manifold is greatly improved.

The invention provides a density peak value clustering method and system based on principal component analysis and a nearest neighbor graph. The method comprises steps of using the principal componentanalysis to carry out characteristic conversion and characteristic extraction on original data, i.e., carrying out dimension reduction on the original data; then, using an improved local density calculation formula, i.e., by use of the nearest neighbor graph to replace the original mode, solving local density; and by use of solving steps in the original algorithm, finding a clustering center point, thereby finishing the clustering. According to the invention, effects on the algorithm imposed by high-dimension data and local structures in the data are fully considered; and the method and the system have quite high robustness and generalization ability.

The invention provides a constrained concept decomposition clustering method based on a depth matrix. The method comprises the following steps: S10 obtaining m images to be clustered and constructs knearest neighbor graphs according to the images to be clustered; S20, obtaining a corresponding data matrix X for each nearest neighbor graph, wherein the data matrix X comprises n data points, and decomposing the data matrix X by a non-negative matrix decomposition method to obtain a characteristic matrix W; S30 marks p data points in the data matrix X; S40, based on the characteristic matrix W and the marked data points, the objective function sigma of the constraint conceptual decomposition is established; S50, according to the objective function sigma, iteratively weighting the preset number of times to obtain the eigenmatrix Wi and the weight matrix Hi of each layer of the depth decomposition of the data matrix X, 1 <= i <= k; S60 Using k -Means clustering algorithm analyzes and clusters the eigenmatrices of each nearest neighbor graph. This clustering method uses marked and unmarked data points, and imposes the marked information on the objective function as a constraint, whichgreatly improves the recognition ability and the clustering accuracy.

The invention discloses a low-rank projection feature extraction method under a label missing condition, and the method comprises the steps: carrying out the detection of sample data and labels one byone, and obtaining a current sample and a corresponding label; and if the current sample label belongs to the normal range, continuing to detect the next sample, and if the current sample label is missing, calculating the Euclidean distance between each sample and the sample in the original space, solving the label with the maximum possibility through a k-nearest neighbor principle, and writing the label corresponding to the sample into the original data set. After label compensation, a nearest neighbor graph matrix is constructed, a projection matrix is learned through a PCA algorithm, the projection matrix is applied to the test set, and classification is performed by using a classifier. The method provided by the invention can adaptively provide more accurate and reasonable training data for various classification models so as to help a classifier to generate a better classification model and can improve the accuracy and robustness of image classification.

The invention discloses a dynamic chain graph model-based earthquake damage remote sensing image segmentation method and system. The method includes the following steps that: a multi-spectral earthquake damage remote sensing image is segmented initially, so that the initial segmentation regions of the multi-spectral earthquake damage remote sensing image can be obtained; heterogeneities of all the initial segmentation regions are calculated; a chain graph model is constructed according to the heterogeneities of the segmentation regions and the adjacency relations among the segmentation regions, wherein the chain graph model includes a region adjoining graph and a nearest neighbor graph which are linked to each other; and red-black tree-based priority queues are constructed with edge lengths in the nearest neighbor graph adopted as primary keys, the red-black tree-based priority queues are dynamically merged according to rule that priority queues with lowest heterogeneity are merged first, and multi-spectral earthquake damage remote sensing image segmentation results matched with earthquake damage surface features can be obtained. With the method and system adopted, wrong segmentation in complex earthquake damage remote sensing image segmentation can be avoided, the correctness of segmentation can be improved, and the segmentation results can be better matched with the earthquake damage surface features.

The invention provides a sparse dual constrained hyperspectral image unmixing method, which comprises the following steps: S10, obtaining a hyperspectral image to be unmixed, and constructing a nearest neighbor graph by searching k nearest neighbor points of each pixel point in a high-dimensional data matrix; S20, taking each pixel point in the nearest neighbor graph constructed in the step S10 asa vector, and constructing another graph in the sparse region through a method of setting a sparse distance between the pixel points; S30, constructing two images according to the step S10 and the step S20; S40, decomposing the data matrix N layer by layer to obtain a corresponding matrix factor, adjusting the matrix factor by using a fine adjustment rule when the data matrix N is decomposed to the last layer, and performing iterative updating after adjustment; and S50, outputting a final iteration result, obtaining an end member matrix U and an abundance matrix V after the hyperspectral image is unmixed, and completing the unmixing of the hyperspectral image. The whole speed is high, the occupied storage space is small, a certain processing speed is maintained while the low complexity isensured, and the unmixing efficiency is improved.

The invention discloses a tree skeleton extraction method based on 3D point clouds, and relates to the technical field of big data processing and tree 3D modeling. The method comprises the following steps: reducing a scale of 3D point cloud data by using an octree method and an L1-median point method; constructing a ground root node; constructing a K-nearest neighbor graph, and dividing horizontal hierarchy sets; performing clustering to obtain point cloud data of different branches in the same horizontal hierarchy set; performing division by using the octree method, and calculating an L1-median point of each division to obtain skeleton points; and determining whether two class clusters are adjacent or not according to whether K-neighbors of members in the two class clusters have intersections or not, connecting the skeleton points of the adjacent class clusters to obtain an initial skeleton, and calculating a B-spline curve of the initial skeleton to obtain an optimized skeleton. Through the steps, influences of uneven 3D point cloud distribution on skeleton extraction can be effectively reduced, so that an extracted tree skeleton result is more accurate.

The invention provides an approximate nearest neighbor search method combining a VP tree and a guide nearest neighbor graph. The method comprises two processes: an index construction process and a search process. The index construction process comprises the following steps: (1) constructing a VP tree for storing entry points for high-dimensional vector data; (2) constructing the high-dimensional vector data into a K-nearest neighbor graph by using a navigable extension graph; and (3) uniformly dividing the neighbor set of each point in the K neighbor graph to obtain a steerable neighbor graph.The search process comprises the following steps of: (1) searching an entry point, namely quickly obtaining the entry point approximately close to a query point by searching the VP tree; (2) a guidesearch stage: starting from the entry point, obtaining a local nearest point of a query point; and (3) an exhaustive search stage: exhaustively searching points near the query point by utilizing a range search algorithm and returning top-K nearest points of the query point. According to the method, the performance of approximate nearest neighbor search is effectively improved by using a hybrid search strategy.

The invention relates to a construction method of a dynamic k-nearest neighbor graph and a rapid image retrieval method based on the dynamic k-nearest neighbor graph, which realize online updating ofan approximate k-nearest neighbor graph and realize rapid image retrieval based on the dynamic k-nearest neighbor graph, and because distance measurement between vectors is not assumed, the method hasgood generalization, and moreover, the mapping efficiency exceeding that of a nearest neighbor descent method is displayed on most of data sets while the quality of the k-nearest neighbor graph is ensured, and the retrieval efficiency on a plurality of image feature data sets is better than that of widely accepted methods such as HNSW and the like.

The invention discloses a similarity measurement and truncation method in a chameleon algorithm, belonging to the technical field of agglomeration type hierarchical clustering algorithms. A chameleonruns on a sparse graph, and nodes represent data items and weighted edges represent similarities between the data items. The chameleon finds out clusters in a dataset through a two-stage algorithm, wherein in the first stage, a k-nearest neighbor graph Gk is constructed according to the dataset, and the data items are divided into several relatively small sub-clusters through a graph partitioningalgorithm; in the second stage, through an algorithm, real clusters are found by repeatedly combining the sub-clusters. The improved algorithm improves a conventional chameleon clustering algorithm byintroducing recursive dichotomy, flood fill and first hop truncation. The invention also provides a method that can automatically select the best clustering result from a modified chameleon tree diagram.

The invention discloses a nearest neighbor graph potential similarity optimization method based on distance transformation. The method comprises the following steps: step 1, constructing a nearest neighbor graph structure and a spectral space thereof; step 2, on the basis of the spectral space, constructing a new similarity distance function expression, namely distance transformation, through function analysis and derivation; step 3, constructing a global nearest neighbor graph and using the global nearest neighbor graph for distance transformation; step 4, constructing a local nearest domainmap based on the consistency punishment information and applying the local nearest domain map to distance transformation; and step 5, respectively constructing a gKNN image and an lKNN image by adopting public data, then optimizing a graph structure by utilizing a proposed distance transformation method, and outputting a final result. According to the method, common problems of potential similarity information hidden in a database are developed by using a gKNN image, and the punishment consensus information is further combined to construct an lKNN image. The robustness and the high performanceof the method are proved, and the superiority of PCI information is also proved.

The invention discloses an image marking method based on semi-supervised subject modeling. The method comprises the following steps of: firstly, obtaining images from the Internet, including images with text marks and unmarked images; then, modeling the relation between the visual features and text marks of all images through latent subjects by use of a model similar to probabilistic latent semantic analysis; establishing the nearest-neighbor graphs of all images, and adjusting the model according to the manifold structure obtained by modeling the nearest-neighbor graphs; learning the model by an expectation maximization algorithm, and calculating the probability of matching the latent subjects with the images respectively; and finally, calculating the probability of matching each text mark with the unmarked images according to the probability of matching the latent subjects with the images, and selecting the text mark with the highest probability to mark the unmarked images.

The invention provides a saliency map fusion method and system. The saliency map fusion method comprises preparation of an annotation data set, and also comprises the following steps: carrying out image appearance-based neighbor search of a test image I in an image set D; extracting an error corresponding to the nearest neighbor image from an absolute error matrix E of the image set D to form an error matrix P; calculating an average absolute error of each detection method of a K nearest neighbor set, K being the number of neighbor images; calculating the weight of each detection method according to the average absolute error; and performing salient region extraction and fusion on the test image I by applying M detection methods. According to the saliency map fusion method and system provided by the invention, image data does not need to be trained in advance, and the problem of parameter modification caused by different training sets is avoided, and the processing method is simple, and the accuracy and efficiency of saliency region detection are effectively improved.

The invention provides a face clustering method and device based on graph convolution and electronic equipment, and relates to the technical field of face clustering, and the method comprises the steps: obtaining a plurality of pieces of to-be-clustered face data, extracting the face features of the plurality of pieces of to-be-clustered face data, constructing a first K-nearest neighbor graph based on the extracted face features, and obtaining a first K-nearest neighbor graph based on the first K-nearest neighbor graph; and inputting the first K neighbor graph into the trained graph convolutional network to obtain a density prediction value of each node in the first K neighbor graph, performing directed connection on the node with the low density prediction value and the node with the high density prediction value and the maximum connection strength, and constructing a plurality of independent face clusters. The mode of combining the graph convolutional neural network and the clustering algorithm based on the density predicted value does not need to additionally construct a sub-graph, and the operation efficiency and accuracy of the clustering algorithm are effectively improved.

The invention relates to a high-dimensional data visualization method based on a probability multi-level graph structure, and belongs to the technical field of data visualization and dimension reduction. The high-dimensional data visualization method comprises the steps of: (1) giving a high-dimensional data set, wherein the data set comprises n data points, and the dimension of each data point isD; 2) calculating k neighbors of each data point, constructing a nearest neighbor graph structure G0, and constructing a probability multi-level graph structure based on the graph structure G0 to obtain a probability multi-level graph structure set; 3) laying out probability multi-level graphs layer by layer based on the probability multi-level graph structure set to obtain data low-dimensional representation, wherein the dimension of each data point is two-dimensional or three-dimensional; and 4) constructing a scatter view based on the low-dimensional data for data mining and analysis. According to the high-dimensional data visualization method, an optimization calculation process is accelerated by utilizing a hierarchical graph structure, and a visualization effect is optimized by introducing probability-based sampling.

The disclosure relates to an image retrieval method, device and storage medium, and belongs to the field of data retrieval. The method includes: obtaining the graph index structure corresponding to the image feature posting table, based on the target image feature and the graph index structure, obtaining the nearest neighbor image feature of the target image feature from the image feature posting table, and obtaining the nearest neighbor image feature corresponding Candidate image set; Based on the residual feature codebook, the residual feature between the target image feature and the nearest neighbor image feature is quantized by product to obtain the coding of the residual feature. The residual feature codebook adopts the optimal product quantization method The product is obtained by quantizing the sample residual features of a plurality of first sample images; based on the encoding of the residual features and the encoding of the candidate residual features of the candidate images in the candidate image set, the candidate image is obtained from the candidate image set as an image retrieval result . In this way, the accuracy of image retrieval can be improved on the premise of ensuring the efficiency of image retrieval.

A protein sequence design method of a fixed skeleton based on graph network mask node classification comprises the following steps: in an offline stage, training a graph neural network by constructing a protein nearest neighbor graph of a sample as a training set; in the online stage, the nearest graph of the protein structure of the deleted sequence is processed through the trained graph neural network, the category probability of the deleted amino acid is obtained, and then the category probability is sampled to obtain the predicted deleted amino acid sequence. According to the method, more structural constraints are added, distance and relative angle features and dihedral angle features of each amino acid are added to connected amino acid pairs in a protein map, and then a neighbor map is established, so that a more efficient map network based on a Transform multiple attention mechanism is realized, and the optimal amino acid mask proportion is explored.