114 results about "Similarity matrix" patented technology

The application provides a high-efficiency expression migration method using base expression space transformation, comprising a preprocessing stage and a facial expression migration stage; the preprocessing stage comprises: obtaining a source model O, the source model O being a base expression model comprising expression meshes under multiple different expressions; performing facial expression reconstruction on the source model O and a target model T to obtain expression description parameters of the two; performing similarity estimation according to the expression description parameters to obtain an expression parameter conversion matrix and a similarity matrix; the facial expression migration stage comprises: inputting a character picture into the base expression model, calculating expression description parameters of a target expression model according to the expression description parameters in the base expression model and the similarity matrix, and completing expression migration.

This application relates to a method and system for detecting art and design feature conflicts based on deep learning. The method includes: acquiring image data of the target artwork; extracting multi-dimensional feature vectors using a multimodal feature extraction network; generating a multi-level feature matrix through multi-scale decomposition; mapping the matrix to a preset semantic space to construct a semantic feature matrix; constructing positive and negative sample pairs using a triplet sampling strategy; calculating the cross-modal feature similarity matrix between the semantic feature matrix of the target artwork and a benchmark feature matrix library using a contrastive loss function; and generating a visualized detection report based on a similarity threshold. This technology solves the technical problems of low efficiency, strong subjectivity, and difficulty in quantifying and analyzing complex conflict relationships between design elements in traditional manual detection, achieving automated, multi-dimensional, accurate identification, and visualized presentation of art and design feature conflicts.

The application provides a three-dimensional point cloud registration method and system based on robust correspondence screening, which comprises the following steps: extracting original local features and superpoint sets of source point cloud and target point cloud respectively; selecting multiple guide points which are dispersed in space from the superpoint set, calculating geometric position relationship between each superpoint and guide point, fusing the geometric position relationship as position embedding information with the original local features to obtain enhanced features; directly screening a first matching pair set from a similarity matrix according to similarity; and screening dispersed matching pairs as a second matching pair set based on spatial position distance constraint of the selected matching pairs; and estimating and optimizing a transformation matrix between the source point cloud and the target point cloud based on at least the first matching pair set and the second matching pair set. The application can obtain high-confidence matching pairs which can completely express the overlapping area through distance constraint screening, and effectively avoid the collapse and failure of a large number of matching pairs caused by the aggregation of point pairs with high feature similarity.

The application discloses a source-load joint probability prediction method and system of a physically constrained graph attention network. The method collects multi-dimensional feature data of source-load nodes in a prediction area to construct an initial node feature matrix. A similarity matrix is generated through differentiable graph structure learning. A dynamic adjacency matrix is generated through normalization and introduction of a sparse mask. Spatial feature aggregation is performed through a multi-head graph attention network to obtain node spatial encoding. The node spatial-temporal hidden state is output through an encoder. The node spatial-temporal hidden state is input into a probability prediction head to output Gaussian distribution parameters of the source-load node power. A joint loss function is constructed. The joint loss function is used for soft constraint training to output a probability prediction result. Posterior projection hard constraint correction is performed in an inference stage to obtain a corrected prediction result. The application solves the problems of lack of physical consistency and inability to quantify uncertainty in the prior art.

The invention discloses an automobile demand prediction method and system based on large model time sequence cross-modal alignment, and belongs to the technical field of automobile demand prediction. The automobile demand prediction method based on large model time sequence cross-modal alignment comprises the following steps: acquiring time sequence data and prompt data related to a hydrogen fuel cell automobile; inputting the time sequence data into the time sequence coding branch in a reverse embedding manner to generate a time sequence embedding representation; inputting the prompt data into an LLM enhanced coding branch for coding, and generating a prompt embedding representation; and calculating a channel-level similarity matrix between the time sequence embedding representation and the prompt embedding representation, aggregating to the time sequence embedding representation based on the channel-level similarity matrix, generating an aligned cross-modal feature, and generating a demand prediction result of the hydrogen fuel cell vehicle based on a decoder. According to the method, multi-source heterogeneous information is effectively fused, the precision and robustness of demand prediction of the hydrogen fuel cell vehicle are improved, and the capability of responding to emergencies caused by external factors is enhanced.

The application relates to an industrial process state monitoring method based on robust space-time joint projection, which comprises the following steps: data preprocessing and standardization; data space manifold structure and time dynamic information capturing: calculating the similarity degree between samples according to neighborhood overlapping order similarity, obtaining a data similarity matrix, and thus obtaining the overall space manifold information of process data; subsequently, calculating a data time derivative matrix, extracting slow change characteristics in the data, and obtaining key dynamic information; feature extraction and monitoring model building based on robust space-time joint projection: comprehensively integrating space-time information, performing dimension reduction projection on the data, obtaining key low-dimensional characteristics for retaining the representation of system operation states, constructing monitoring statistics and monitoring thresholds based on the key low-dimensional characteristics, and building an offline process monitoring model; real-time online monitoring of the state of an industrial system.

The application discloses a kind of gear engagement loop resistance measurement based on gear face lubrication state judging method, belong to gear face lubrication technical field.This method includes: installation angle sensor identifies the engagement start-stop point of each gear tooth;Constant current is applied to collect loop resistance signal, and is divided into the engagement period resistance waveform of each gear tooth;The similarity matrix between waveform and the discrete degree value of its row mean are calculated, compared with the reference value under initial good lubrication state to obtain relative deviation degree;According to relative deviation degree, elastohydrodynamic lubrication, mixed lubrication or boundary lubrication state is judged, and the gear tooth number of similarity abnormality is output simultaneously.The application can also extract oil film extrusion depth coefficient and recovery residual coefficient, and identify lubricating film dynamic instability through inter-circle cross-correlation analysis.The method realizes gear tooth level lubrication state evaluation and abnormal positioning, is robust to working condition change, and is convenient for online monitoring and predictive maintenance.

This invention discloses a graph attribute prediction system based on manifold learning, relating to the fields of graph neural networks and machine learning technology. It includes: a feature smoothing module, a graph structure feature extraction module, a manifold similarity calculation module, a deep graph representation learning module, a manifold structure preservation loss calculation module, a graph attribute prediction module, and a joint optimization module. The invention also discloses a graph attribute prediction method based on manifold learning, comprising the following steps: S100, system initialization; S200, data input; S300, generating a node embedding matrix; S400, calculating a node similarity matrix; S500, latent space structure feature extraction; S600, manifold learning loss calculation; S700, graph attribute training and prediction; S800, joint optimization; and S900, graph attribute inference and prediction. This invention effectively maintains the topological authenticity of the latent space and significantly improves performance in downstream tasks such as node classification and link prediction.

The present application relates to the technical field of industrial vision detection, and discloses a green color printing surface defect detection method based on computer vision. The method comprises the following steps: collecting real-time image flow and performing multi-scale decomposition to generate a feature vector sequence; constructing a dynamic defect template set by using a generative adversarial network; encoding features by using a convolutional neural network, calculating a similarity matrix with the template set, and generating a defect matching index; adjusting the detection threshold value adaptively according to the defect matching index, and obtaining a defect probability distribution map through clustering analysis; identifying a defect aggregation area through space-time domain analysis, and starting a local re-inspection mechanism; updating a defect feature prototype library and optimizing network weights by using re-inspection data, and forming a self-learning closed loop; and finally performing full-process quality consistency inspection and outputting a report. The present application realizes dynamic adaptation and continuous optimization of the detection process.

The application discloses a course teaching resource recommendation method, comprising the following steps: establishing a course teaching resource recommendation dataset, learning text and image features of course content and course teaching resources in a training set, and calculating a similarity matrix between the text and image features; based on the similarity matrix and existing scoring samples, a low-rank sparse matrix decomposition recommendation model of the course teaching resources is constructed, and a scoring matrix is decomposed into a low-rank latent matrix related to course content and a sparse latent matrix related to course teaching resources; the low-rank matrix and the sparse matrix are initialized by using random numbers, the low-rank matrix and the sparse matrix are iteratively optimized by using a gradient descent method, and the optimization is updated until a target function based on the similarity matrix, the low-rank matrix and the sparse matrix converges; course content features in test data are extracted, corresponding course teaching resources are found from approximate estimation of the scoring matrix, and a recommendation list is returned according to scoring item sorting.

The application relates to a social network text target topic detection method based on sparse subspace clustering, and belongs to the technical field of computer data processing of natural language processing and social network data mining.The method writes text identification into target platform text content and interaction event records, carries out word segmentation, denoising and vectorization processing, reduces short text noise and event mismatch interference on subsequent analysis, constructs a social relationship graph, calculates edge confidence weight to form a graph regular constraint parameter, reduces the weight of a low-confidence screen edge in the constraint, suppresses relationship noise caused by organized manipulation from the source, solves sparse representation coefficients, constructs a similarity matrix, performs spectral clustering, enhances the separability of samples with similar semantics but different propagation modes, calculates semantic cohesion and propagation deviation, determines a target topic cluster by using a double threshold, extracts a key text set and a propagation evidence set, forms a reviewable target topic detection result, and improves target topic detection rate and reduces false positives.

The application discloses a special vehicle engine digital maintenance system based on big data analysis, comprising: a data set generation module for forming an engine time series data set; a feature generation module for forming a window feature sequence; a prototype construction module for forming a working condition prototype set and an ordered working condition label sequence; a similarity generation module for forming a directed transition similarity matrix; a cost generation module for forming an adaptive segmentation cost sequence; an initial segmentation module for forming an initial working condition state segment set and an initial state segment boundary by using an improved SegmenTier algorithm; an evidence interval generation module for forming a precursor evidence interval; and a re-segmentation output module for outputting a working condition recognition result, a precursor segment positioning result and a maintenance triggering result. The application improves the working condition segmentation accuracy and the maintenance triggering reliability.

The application provides a pre-training model structure pruning method and system for module clustering and centroid selection, comprising: obtaining hidden representations of each layer of a model using a pre-training language model and a downstream task dataset; calculating size limitation conditions for different modules according to given model size constraints; calculating cosine similarity between different modules using the hidden representations of each layer; clustering different modules according to the calculated cosine similarity matrix; determining the number of categories using a clustering tree and the size constraints of different modules; calculating the average amplitude size of different modules according to the hidden representations of each layer, and retaining the module with the largest amplitude in each category cluster and pruning other modules in the category cluster. The application proposes a structured pruning method based on module clustering and centroid selection for the proposed pre-pruning framework, which can efficiently measure the similarity of different modules of a language pre-training model.

The application discloses a large model-based information pushing strategy generation method and device, a storage medium and computer equipment. The method comprises the following steps: first, obtaining customer original behavior data and dividing time and space dimensions, and constructing a time-space fusion similarity matrix according to the time and space dimensions; second, constructing a graph structure with the customer as a node and extracting customer fusion features; third, searching and matching knowledge information in a preset knowledge base; and finally, constructing prompt words based on the knowledge information and the customer fusion features, and generating an information pushing strategy by a strategy generation large model. The method can be applied to the information pushing strategy generation scene of financial technology and medical health, can help to deeply mine the potential law of customer behavior, accurately grasp the needs of customers in different scenes, improve the accuracy and individuality of the information pushing strategy, improve the efficiency and effect of information pushing, and further improve the satisfaction of customers. The method can be applied to the financial technology and medical health scene.

The present application relates to the technical field of microanalysis, and particularly relates to a microanalysis similarity pixel-level quantitative characterization method based on intersection over union, which comprises the following steps: obtaining an element distribution map and extracting an effective area, establishing a mapping relationship between color and numerical value through a color scale, extracting element intensity values pixel by pixel, and generating a normalized spatial distribution matrix; generating a binary mask corresponding to each element according to a preset calculation mode, wherein the first mode performs binaryzation processing on the intensity matrix based on a threshold value, and the second mode obtains a binary mask through manual labeling; calculating the intersection over union between binary masks of any two elements to represent the similarity of the spatial distribution thereof; and outputting a similarity matrix and a visualization result. The present application realizes objective and reproducible pixel-level quantitative evaluation of the similarity of the spatial distribution of elements in microanalysis, and is suitable for various scenes such as batch comparison and small sample fine analysis.

In methods of recommending exercise, first to n-th text combinations are generated for first to n-th questions, which are embedded into first to n-th question vectors, cosine similarity between the first to n-th question vectors is calculated to generate a question similarity matrix, first and second correlation matrices are generated based on information on whether other users answered the first to n-th questions correctly, first and second determination matrices are generated by multiplying the question similarity matrix to the first and second correlation matrices, respectively, answering results of a target user for some of the first to n-th questions are received, similarity scores for remaining questions except for a question that the target user answered incorrectly among the first to n-th questions is calculated using the determination matrices, and some of the first to n-th questions are recommended to the target user in descending order of the similarity scores.

The present application belongs to the field of data analysis, and particularly relates to a pre-school children picture book intelligent recommendation method, a picture book management system and a reading device thereof. The recommendation method recommends new books suitable for and meeting the current user's preferences to the current user according to all the user's personal information and reading records. The recommendation method comprises the following steps: S1: obtaining all the picture books and score records of the picture books that the current user has seen or shared, to obtain a seen list and a score table; S2: obtaining all the picture books and score records of the picture books that all the similar users have seen or shared, to obtain a reference list and a score table; S3: constructing a tag similarity matrix; S4: generating a user preference matrix; S5: calculating the final preference score of the current user for each picture book in the reference list; and S6: reordering the reference list to obtain a recommended book bag. The present application solves the problem that the existing picture book recommendation method is not suitable for pre-school children reading recommendation.

The application provides a kind of based on the incremental knowledge graph entity alignment method and device of multi-dimensional feature, the steps of the method include: for the entity in first knowledge graph and second knowledge graph, attribute vector is constructed;First similarity matrix is calculated based on attribute vector;For the neighbor entity of each entity of first knowledge graph and second knowledge graph, the number of entity that the entity of first knowledge graph and the neighbor entity of the entity of second knowledge graph are mutually aligned is calculated, and the matrix of aligned neighbor entity number is constructed;Second similarity matrix is calculated based on first similarity matrix and aligned neighbor entity number matrix;Information quantity parameter is calculated based on the entity information of entity pair, and the pre-screening comparison value of entity pair is calculated based on second similarity and information quantity parameter, to obtain entity pair screening set;For each entity pair in entity pair screening set, the third similarity of two entities is calculated, and the alignment result of the two entities of the entity pair is determined based on the third similarity.

This invention relates to the field of image recognition technology, specifically to a method and system for intelligent recognition of abnormal behavior at construction sites based on deep learning. The method acquires continuous image frames and extracts shallow feature maps, which are then input into a parallel decoupling network of a background dynamic decoupling network. A background motion estimation branch predicts the background pixel-level displacement field based on deformable convolution, and the shallow feature map of the current frame is reverse-distorted and aligned to generate an aligned feature map. A foreground human body separation branch extracts the foreground human body motion mask based on the aligned feature map. The spatiotemporal feature similarity matrix of the foreground human body motion masks across multiple consecutive frames is calculated, and pseudo-motion masks caused by sudden changes in illumination are removed, retaining temporally consistent motion masks. The temporally consistent motion masks are then input into a behavior classification network to output the recognition result. This invention removes non-human dynamic background interference at the feature level, eliminates mask anomalies caused by illumination changes, and ensures the purity of the input features to the behavior classification network.

The application relates to the technical field of power system network security, and discloses a digital power grid security situation awareness prevention and control method based on artificial intelligence, which comprises the following steps: storing historical power flow feature vectors into a physically isolated static storage area and performing hierarchical version management; constructing a dynamic cache area in an independent memory space, extracting real-time feature vectors through an adaptive sliding window, and completing cache eviction by attaching a time decay label; adopting Jensen-Shannon divergence and cosine similarity linear weighting to calculate a global similarity matrix, and generating a feature importance decay compensation weight vector; constructing a main gradient flow and compensation gradient flow double-channel architecture in the model back propagation stage, updating the model parameters after gradient gate fusion; and dynamically adjusting the threshold based on the gradient flow contribution degree, so that system closed-loop adaptive optimization is realized. The application can relieve the model catastrophic forgetting problem, and improve the power grid security risk identification precision and the adaptive ability of the prevention and control system.

The application discloses a risk account identification method and device, and relates to the technical field of risk control. A specific implementation manner of the method comprises the following steps: acquiring transaction data corresponding to two payment systems respectively, wherein the transaction data comprises a plurality of transaction accounts and transaction information of each transaction account; generating a first transaction matrix and a second transaction matrix corresponding to the two payment systems respectively according to the transaction data; calculating a first similarity matrix corresponding to the two payment systems according to the first transaction matrix and the second transaction matrix; determining an account matching pair from the two payment systems according to the first similarity matrix and a matching algorithm; and determining a risk account according to the account matching pair. The implementation manner improves the accuracy of identifying the risk account.

This invention relates to a multilingual generative retrieval method based on cross-language semantic compression, belonging to the field of information retrieval technology. The invention includes the following steps: constructing a multilingual document retrieval dataset; extracting keywords from multilingual documents from multiple perspectives using a keyword extraction model, and calculating the extracted keywords using semantic similarity to construct a similarity matrix; performing semantic clustering based on the similarity matrix, representing clusters using atomic IDs, and then assigning document identifiers to each multilingual document by the cluster containing the keywords; in the inference stage, after inputting a query, employing a dynamic multi-complement constraint decoding method, gradually narrowing the decoding range of the document identifier in the current step based on the decoding results of previous steps, thereby obtaining the final document identifier. The retrieval capability of this invention is significantly improved compared to other models.

The application provides a voice keyword recognition and positioning method and device, wherein the method comprises obtaining a to-be-tested audio and a candidate keyword table, inputting the to-be-tested audio and the candidate keyword into a trained voice keyword recognition and positioning model to obtain an audio feature sequence and a local text feature representation, obtaining a similarity matrix of a text token by using a monotonous alignment search algorithm according to the matching score of the text token and each audio feature, obtaining an optimal alignment path of the text token by using the similarity matrix, determining whether the candidate keyword is a keyword of the to-be-tested audio according to the score of each text token optimal alignment path and a set threshold, and determining the start and end time of the appearance of the keyword in the to-be-tested audio through the optimal alignment path of each text token. The application solves the problem that a traditional weak supervision model is difficult to consider both semantic detection and accurate timing positioning.

This invention discloses a multi-target discrimination method based on the spatiotemporal features and dynamic similarity measurement of flight tracks, comprising the following steps: acquiring raw flight track data; preprocessing the raw flight track data to generate multiple continuous flight track segments; extracting spatiotemporal features and statistical features to construct a similarity matrix; performing hierarchical flight track clustering analysis on the similarity matrix, the result of which is n clusters divided according to the similarity matrix; loading a relational model, performing conflict resolution and target identification, and calculating the flight track association confidence score; and generating a flight track set based on the flight track association confidence score. According to the above technical solution, multi-target echo signals in radar seeker tracking dense target scenarios can be distinguished, achieving high-accuracy target flight track association and high-precision low-latency processing.

This invention belongs to the field of power system monitoring technology and discloses a method, system, equipment, and medium for analyzing the icing state of transmission lines based on point cloud modeling and deep learning, thereby overcoming the limitations of existing methods in time series processing. The method includes: dividing the time series point cloud of the transmission line into multiple time subsequences; generating a topological persistence graph for each subsequence through voxelization and persistent coherence analysis; calculating the dynamic topological distance between subsequences based on the persistence graph; extracting topological features to construct a feature matrix, and fusing topological distance, feature space, and feature distribution three views to construct a similarity matrix; using multi-view joint learning to obtain a consensus subspace and performing subsequence clustering within this space to obtain cluster labels; combining physical context information and predefined rules to assign physical pattern labels to the clustering results, generating an icing state classification report and visualizing it. This invention achieves temporal topological quantitative analysis and highly robust state identification of icing morphology.

The application discloses an interactive pathological section image segmentation method, device, equipment and medium. The method comprises the following steps: using a pre-trained feature extraction model to perform feature extraction on a plurality of foreground image blocks in a target pathological section image to construct an image block similarity matrix; in response to a paracarcinoma image block labeling operation, performing similarity retrieval on unlabeled image blocks in the foreground image blocks according to labeled image blocks in the foreground image blocks to obtain reference tumor image blocks; in response to a tumor image block labeling operation, constructing a directed graph according to the image block similarity matrix, the foreground image blocks and image block identifiers; and performing segmentation on the directed graph by using a minimum cut algorithm to obtain a target segmentation result. The technical scheme is based on self-supervised contrast learning to construct an interactive semi-automatic pathological section image segmentation mode, and through similarity analysis of image block features, the fast and accurate segmentation of the pathological section image can be realized without the need of labeled data training.

The application discloses a task-driven intelligent agent dynamic decision optimization method and system, which comprises the following steps: modeling graph structure data, constructing an adjacency matrix, a feature matrix and a similarity matrix; constructing neighborhood intelligent agents and layer number intelligent agents; based on the neighborhood intelligent agents, evaluating the importance of neighbor nodes through an attention mechanism and combining the upper confidence bound method to make reinforcement learning decisions, so as to obtain an optimal neighbor selection strategy; based on the layer number intelligent agents, dynamically determining the optimal aggregation network layer number for each node to obtain an optimal network layer number identification strategy; based on the above two strategies, re-normalizing the attention weight of the node and executing task-specific information aggregation under a graph neural network framework to generate a node embedding representation; constructing a joint loss function comprising a reconstruction loss and a self-supervised loss, optimizing the joint loss function, training the parameters of the neighborhood intelligent agents, the layer number intelligent agents and the graph neural network, and outputting an optimal node embedding result.

This invention proposes an enhancement method, apparatus, medium, and device based on underwater degraded images. By combining local feature modeling of image patches and global feature modeling of graph neural networks, it solves the problems of excessive local enhancement and global imbalance in underwater image enhancement. Through processing the cosine similarity matrix, the information transmission capability between image patches is further improved, achieving high-quality enhancement of underwater images. The enhancement method based on underwater degraded images can effectively restore details in underwater images, improve color consistency and brightness balance, and is particularly suitable for complex underwater environments.

The application discloses a kind of academic question and answer-oriented multi-model paper retrieval method, it is related to natural language processing and information retrieval technical field, including: first, construct unified corpus and training dataset, utilize the model in first model set and second target model respectively encode generation document vector set;Then, based on the initial retrieval result of second model, difficult negative sample is filtered, and the contrast learning sample pair is constructed to fine-tune and re-encode corpus;With the model group of the second target model after fine-tuning and the model in first model set, each model in model group is executed similarity retrieval in parallel respectively, and the corresponding original similarity matrix is obtained, based on the original similarity matrix, the document is filtered, and the first target document list of target query is generated.The application can effectively improve the accuracy and robustness of academic literature retrieval.

The application discloses a gas reservoir classification method and system, comprising: collecting characteristic data of all attributes of a combination of gas reservoirs to be classified, and dividing the characteristic data into numerical data and text data; determining a gas reservoir type matched with each text characteristic data according to a specified natural gas classification standard, and generating a one-hot encoding of each text characteristic data, further combining normalized numerical data to construct a similarity matrix used for calculating the similarity between gas reservoirs; obtaining a clustering result of the combination of the gas reservoirs to be classified by using the similarity matrix, and then determining a weight value representing the influence degree of each attribute on the clustering result and a best gas reservoir classification number; obtaining a classification result matched with the best gas reservoir classification number based on any attribute, and then evaluating the similarity between the gas reservoirs in each category in the current classification result by using the weight value to obtain a best classification result. The application realizes effective classification of the gas reservoirs under the condition of considering the overall characteristics of the gas reservoirs.