Patents
Literature
36 results about "Manifold regularization" patented technology
Efficacy Topic
Property
Owner
Technical Advancement
Application Domain
Technology Topic
Technology Field Word
Patent Country/Region
Patent Type
Patent Status
Application Year
Inventor
In machine learning, Manifold regularization is a technique for using the shape of a dataset to constrain the functions that should be learned on that dataset. In many machine learning problems, the data to be learned do not cover the entire input space. For example, a facial recognition system may not need to classify any possible image, but only the subset of images that contain faces. The technique of manifold learning assumes that the relevant subset of data comes from a manifold, a mathematical structure with useful properties. The technique also assumes that the function to be learned is smooth: data with different labels are not likely to be close together, and so the labeling function should not change quickly in areas where there are likely to be many data points. Because of this assumption, a manifold regularization algorithm can use unlabeled data to inform where the learned function is allowed to change quickly and where it is not, using an extension of the technique of Tikhonov regularization. Manifold regularization algorithms can extend supervised learning algorithms in semi-supervised learning and transductive learning settings, where unlabeled data are available. The technique has been used for applications including medical imaging, geographical imaging, and object recognition.
A Label Noise Estimation Method Based on Manifold Regularized Transfer Matrix
A label noise estimation method based on a manifold regularization transfer matrix provided by the present invention pre-trains a first network in a second network, and after distilling a data set, inputs the obtained sub-data set into the second network to obtain the probability of the class to which the data instances in the sub-data set belong and obtain a transfer matrix related to the data instances; further calculates the cross-entropy loss of the second network according to the data instance labels, and combines an association matrix expressing the consistency of the data instances belonging to the same manifold and a penalty matrix of the data instances belonging to different manifolds to calculate the loss function of the second network; adjusts the loss function to reduce the training of the second network to obtain a trained second network, thereby completing the estimation of the class to which the data instances belong. The present invention can reduce the estimation error without affecting the approximation error of the transfer matrix, and experiments prove that the present invention can achieve excellent performance in label noise learning.
Owner:XIDIAN UNIV
A semi-supervised medical image segmentation method based on contrast manifold regularization and related devices
The present application discloses a semi-supervised medical image segmentation method and related devices based on contrast manifold regularization. The method includes: importing and preprocessing medical image datasets; initializing a teacher model and a student model, using labeled data to train the teacher model and generate pseudo labels; calculating the similarity between the labeled data and the pseudo labels to obtain a manifold regularization term; constructing positive and negative sample pairs and calculating the contrast loss term; weighted summation to obtain the contrast manifold regularization term, and combining it with the supervised loss as the loss function of the student model; iteratively training the student model to obtain the segmentation result. By combining contrast learning with manifold regularization, the present invention effectively alleviates the data dependency problem in semi-supervised medical image segmentation, improves the model generalization ability and segmentation accuracy, and performs particularly well in small target lesion segmentation scenarios.
Owner:ZHUHAI HENGQIN ALL-STAR MEDICAL TECHNOLOGY CO LTD
SAM-based semi-supervised adapter fine tuning and prompt learning medical image segmentation method and related device
The invention discloses a medical image segmentation method based on SAM semi-supervised adapter fine tuning and prompt learning and a related device. The method comprises the following steps: importing and preprocessing a medical image data set, loading a pre-training SAM model and a YOLO prompt model, inserting double adapters into each visual converter block of an SAM encoder, and embedding an over-prompt adapter, an MLP alignment adapter and a cross attention adapter into a decoder; training a prompt model based on the annotated data, screening qualified unannotated data through KL divergence, generating a click prompt, and performing pseudo tag segmentation by using an SAM model; a joint loss function containing a contrast manifold regularization item is constructed, and on the premise of freezing an SAM trunk, adapter fine tuning is executed in combination with a pseudo tag and annotation data. The device comprises a data import unit, a model loading unit and a prompt training unit. According to the method, the labeling dependence is effectively reduced, the model generalization ability and the segmentation precision are enhanced, and meanwhile, the original performance advantages of the SAM trunk are reserved.
Owner:ZHUHAI HENGQIN ALL-STAR MEDICAL TECHNOLOGY CO LTD
Photoplethysmography identity recognition method and system
PendingCN121191202ABiological modelsSubcutaneous biometric featuresSparse constraintIdentity recognition
The invention provides a photoelectric volume pulse wave identity recognition method and system, and relates to the technical field of identity recognition, and the method comprises the steps: obtaining a to-be-recognized PPG signal; and inputting the PPG signals into a trained identification model, firstly extracting linear features and nonlinear features, then projecting the features fused by the two features into a feature space by using a learned discriminant projection matrix to obtain multi-view features, and finally classifying the PPG signals by using the multi-view features to obtain an identity identification result. According to the method, manifold regularization and inter-class-error double sparse constraints are combined, the problems of intra-class discretization and inter-class overlapping of a linear model are solved, a graph structure learning method for adaptive local density adjustment is provided, and the manifold modeling precision of non-stationary PPG signals is improved.
Owner:XINJIANG UNIVERSITY
Emotion recognition method based on multi-band Riemannian manifold space learning
The invention discloses an emotion recognition method based on multi-band Riemannian manifold space learning, and belongs to the technical field of biological signal recognition, and the method comprises the steps: carrying out the self-adaptive frequency band division of an original signal through variational mode decomposition, restraining the mode aliasing, and extracting a multi-band sub-signal; further mapping the sub-signals to a Riemannian manifold space by using a multivariate phase space reconstruction + Riemannian manifold regularization Fisher discriminant analysis module, representing the geometrical characteristics of the signals through a covariance matrix, enhancing the characteristic separability by combining tangent space projection and discriminant analysis, and fusing the characteristic vectors by using an attention mechanism to obtain a multi-element phase space reconstruction and Riemannian manifold regularization Fisher discriminant analysis model; and finally, inputting the fused feature vectors into a classification model constructed by a graph convolutional neural network, and carrying out classification processing. According to the method, the problems of insufficient utilization of frequency band information and weak manifold feature discrimination in traditional emotion recognition are effectively solved, and the recognition precision and robustness are remarkably improved in scenes such as physiological signals.
Owner:DONGGUAN UNIV OF TECH +1
Low-rank multi-modal remote sensing image clustering method and device on superpixel manifold, and storage medium
The invention discloses a low-rank multi-modal remote sensing image clustering method and device on a superpixel manifold and a storage medium, and relates to the technical field of multi-modal remote sensing image clustering. The method comprises the following steps: splicing a multi-modal remote sensing image along a channel direction, segmenting the spliced image into a plurality of sub-regions through superpixel segmentation, and solving a mean value for each sub-region to obtain a multi-modal superpixel; embedding the Laplacian matrix of the multi-modal superpixels into manifold regularization about the superpixel clustering matrix, and capturing a local manifold structure of the multi-modal superpixels; under the constraint of manifold regularization, constructing a low-rank reconstruction model of a product of a single-mode clustering matrix and a unified clustering matrix; initializing and alternately optimizing the single-mode clustering matrix and the unified clustering matrix by using fuzzy clustering; and analyzing the super-pixel clustering result, and mapping the super-pixel clustering result into a clustering result of the original image. According to the invention, the accuracy and efficiency of remote sensing image clustering are improved.
Owner:CHENGDU TECH UNIV
Adaptive multi-domain disentanglement manifold network-based fault detection method for multi-working condition of inverter system
This invention discloses a multi-condition fault detection method for inverter systems based on adaptive multi-domain unentangled manifold networks. Addressing the issue of multi-condition operation and scarce samples for some conditions in inverter systems, the method first constructs a dual-path encoder comprising a shared encoder and a private encoder to explicitly separate common features that do not change with the operating conditions from unique features that do change with the operating conditions. Then, an adaptive weighting mechanism based on reconstruction differences is used to quantify the reference value of the source domains, selectively increasing the proportion of highly correlated source domains to suppress negative migration. Next, manifold regularization techniques are used to constrain the latent space, constructing a geometrically coherent healthy manifold benchmark. Finally, the distance of the test sample projected onto the local manifold space is calculated as a fault monitoring index to achieve the fault detection objective for the inverter system.
Owner:UNIV OF ELECTRONICS SCI & TECH OF CHINA
A manifold constraint-based elastic body erosion profile multi-modal self-attention fusion prediction method
The application discloses a kind of based on manifold constraint's elastic body erosion profile multimodal self-attention fusion prediction method, comprising: data set construction and pre-processing;Build multimodal self-attention fusion network model;Based on the geometric manifold characteristics and physical evolution law of elastic body erosion profile, build the dual compound constraint loss function including data loss, geometric manifold regularization loss and physical consistency loss;Model training and hyperparameter optimization;The microstructure image and physical parameter under the working condition to be predicted are input into the trained model, and the discrete coordinate points and continuous profile curve of the elastic body erosion profile are output, and the importance of different input features to the prediction result is quantified by the integrated explainability analysis module.The application realizes the high-precision and rapid prediction of the elastic body erosion profile, and has the advantages of strong generalization ability and explainability, which provides important technical support for the completion of ammunition warhead design, protection engineering optimization and damage efficiency evaluation tasks.
Owner:NANJING UNIV OF SCI & TECH
A mashup service multi-label classification method based on double manifold regularization width learning
ActiveCN116150669BNatural language data processingEnergy efficient computingActivation functionMulti-label classification
The application provides a Mashup service multi-label classification method based on double manifold regularization width learning, mainly comprising: using a hidden Dirichlet distribution topic model to extract features from preprocessed Mashup description documents; linearly mapping a Mashup description document topic feature matrix into n groups of feature nodes respectively; processing the feature nodes through an activation function to generate enhanced nodes; splicing the feature nodes and the enhanced nodes to generate enhanced feature nodes as the input of the model; constructing a target function of a Mashup service multi-label classification model based on double manifold regularization width learning; using a least square method to solve the target function to obtain a weight matrix of a double manifold regularization width learning network; obtaining a description document of a test Mashup service and sending it into a trained model to predict a multi-label classification result. The application improves a width learning model by using double manifold regularization, and realizes a Mashup service multi-label classification function by using an improved BLS model.
Owner:DALIAN MARITIME UNIVERSITY
Balanced multi-view clustering method and system for discrete weighted pseudo labels
The invention provides a balanced multi-view clustering method and system for discrete weighted pseudo tags, and relates to the technical field of multi-view clustering analysis, and the method comprises the steps: firstly obtaining a multi-view data set; then, according to the multi-view data set, constructing a weighted label indication matrix of the multi-view data set; constructing a balanced clustering model according to the multi-view data set and the weighted label indication matrix, and training the balanced clustering model to obtain a trained balanced clustering model; and finally, inputting multi-view data to be clustered into the trained balanced clustering model to obtain a clustering result. According to the method, the clustering model is normalized through balance constraint, the discrete pseudo-label matrix is weighted to avoid trivial solutions, norms are introduced to realize feature selection, manifold regularization is combined to relieve overfitting, and the matrix is circularly updated, so that higher clustering efficiency, accuracy, balance and feature selection capability are realized.
Owner:GUANGDONG UNIV OF TECH
Cancer subtype detection method based on multi-view clustering under diversity criterion
The present invention discloses a method for detecting cancer subtypes based on multi-view clustering under a diversity criterion, which relates to the field of data processing technology and solves the problem that existing multi-view clustering methods ignore the differences between multiple views and the problems existing in the representation tensor. The present invention improves on the classic multi-view method, divides the representation tensor into a clean part and a noisy part, adds a weighted tensor nuclear norm to the clean representation tensor to take into account the prior knowledge of singular values, and applies l to the noisy part. 2,1 Norm constraints make the learned representation tensor clearer and enhance the robustness of the algorithm. A difference term is introduced into the objective function to describe the diversity of the multi-view representation matrix, and a manifold regularization term is introduced to preserve the local structure of the data. This improves the effectiveness and robustness of multi-view clustering, achieving excellent performance on various datasets.
Owner:SOUTHWEST UNIV
Dual relaxation image classification method based on a width learning system
The application provides a double relaxation image classification method based on a width learning system, and steps are as follows: firstly, a feature data set and a corresponding class label matrix are acquired, and the feature data set generates width conversion features through a standard width learning network; secondly, a double relaxation technique and a graph regularization technique are introduced, and a double relaxation image classification optimization objective function based on the width conversion features is constructed; finally, the double relaxation image classification optimization objective function is solved by using iterative optimization, a classification result is obtained, and the classification result is evaluated. The manifold regularization technique is applied to the width learning network, and a double relaxation method is used to obtain greater freedom, so that the data geometric structure is mined, and the learning of the intra-class similarity realizes the relaxed regression of the target. The application has the characteristics of higher classification precision, relatively less training time, higher model flexibility and the like, and the introduction of the double relaxation method makes the model have stronger discrimination ability.
Owner:HENAN UNIVERSITY OF TECHNOLOGY
Deep learning-based logistics system visual modeling and simulation method and system
InactiveCN120764356ADesign optimisation/simulationNeural learning methodsLogistics managementControl engineering
The invention discloses a logistics system visual modeling and simulation method and system based on deep learning, and the method comprises the following steps: collecting various types of logistics data, including transportation, traffic, storage and environment information, and constructing an adjustable modeling mechanism; an improved neural differential equation model is adopted, so that the system state continuously changes along with time, and a manifold regularization kernel regression method is introduced for dynamic adjustment. The system also utilizes a specific data structure to learn and restrain control input, keeps the modeling process stable and consistent, trains each module through a unified optimization strategy, and realizes dynamic simulation and visual output of the operation state of the logistics system. The invention aims to improve the modeling efficiency and response capability of the logistics system in a complex environment.
Owner:GUOLIAN (SHANDONG) LOGISTICS TECHNOLOGY CO LTD
A graphics card workstation intelligent configuration method based on big data analysis
PendingCN122173150AImprove stabilityImprove adaptabilityBootstrappingEnergy efficient computingGraphic cardData set
This invention discloses an intelligent configuration method for graphics card workstations based on big data analysis, comprising: S1, collecting data and features to generate an original holographic runtime dataset; S2, constructing a three-dimensional heterogeneous tensor matrix, extracting the kernel tensor and factor matrix using HOSVD decomposition, and generating a deep computing power feature vector; S3, constructing a computing power feature profile library, calculating similarity to identify task load pressure patterns, and outputting a target task feature profile; S4, generating a preliminary configuration scheme using an improved WGAN-GP model, and introducing a manifold regularization projection mechanism to filter candidate configuration subsets; S5, parsing and obtaining hard constraints, and matching the final configuration parameters; S6, issuing configuration parameters and collecting real-time load data for dynamic correction; S7, constructing incremental samples to feed back into the original dataset, and performing continuous iterative self-optimization. This invention achieves accurate matching of graphics card resources and task load, effectively improving the computing power utilization efficiency of the workstation.
Owner:NANJING YAOZHUO NEW MATERIAL TECHNOLOGY CO LTD
Console login identity information acquisition and verification method based on deep learning
ActiveCN120263423AUser identity/authority verificationBiological modelsFeature vectorComputation process
The invention relates to the technical field of information security, and discloses a console login identity information acquisition and verification method based on deep learning, which comprises the following steps: S1, acquiring biological characteristics and behavior characteristics of a user, extracting a feature vector of the biological characteristics by adopting a deep learning model, extracting a feature vector of the behavior characteristics by adopting a time sequence model, and acquiring a user login identity information database; the feature vectors of the biological features and the feature vectors of the behavior features are spliced to obtain identity feature vectors, and the feature vectors of the biological features and the feature vectors of the behavior features are spliced to construct the identity feature vectors; s2, carrying out dimensionality reduction on the identity feature vector by adopting a topology preserving projection method, and constructing a low-dimensional manifold embedding representation; and S3, adding a manifold regularization item in the loss calculation process of the deep learning model. According to the method, deep learning and a time sequence model are combined to extract identity features, the effect of quickly and accurately collecting biological and behavior information is achieved, and compared with a single feature recognition scheme in the prior art, the problems of high false recognition rate and low recognition efficiency are solved.
Owner:MT TITLIS BEIJING CONTROL TECH
A multi-view based news theme mining method
This invention discloses a news topic mining method based on multiple views, belonging to the technical field of text analysis and data mining. First, the invention constructs an LT-MSC model, then builds manifold regularization, sparse constraint, and diversity regularization terms. By introducing the manifold regularization term, geometric information in the multiple-view news data is mined; by using the sparse constraint term, the block diagonal structure of the subspace representation matrix is enhanced; and by incorporating the diversity regularization term, complementary information between different views in the news data is captured. Finally, spectral clustering is used to cluster the data. Compared to existing single-view and multi-view mining schemes, the method of this invention achieves the best clustering effect and recognition performance in news topic mining scenarios, thereby effectively improving the efficiency and accuracy of news topic mining.
Owner:JIANGXI JILUO SCIENTIFIC & TECHNOLOGICAL ACHIEVEMENTS TRANSFORMATION SERVICE CO LTD
Robust physical extreme learning machine modeling method for lithium battery temperature prediction
The embodiment of the invention provides a robust physical extreme learning machine modeling method for lithium battery temperature prediction, and belongs to the technical field of data processing, and the method specifically comprises the steps: 1, building a lithium battery heat conduction model corresponding to a lithium battery of an electric camellia fruit picking machine based on the law of conservation of energy; 2, constructing a lithium battery temperature prediction model fused with physical data, substituting the model into the lithium battery heat conduction model, and constructing a weight optimization model for lithium battery temperature prediction; 3, on the basis of the weight optimization model, introducing a structure risk item, an error item based on adaptive error weighting and a manifold regularization item, and constructing a robust optimization model for lithium battery temperature prediction; and 4, by constructing a Lagrange function and utilizing a KKT condition to solve the robust optimization model, obtaining an analytic solution of an output weight vector, and reconstructing temperature field distribution of the lithium battery according to the analytic solution. Through the scheme of the invention, the prediction accuracy and robustness are improved.
Owner:CENTRAL SOUTH UNIVERSITY OF FORESTRY AND TECHNOLOGY
A hypergraph representation method of brain functional network
This invention discloses a hypergraph representation method for brain functional networks. The steps include: preprocessing resting-state functional magnetic resonance imaging (fMRI) to obtain time series data for all brain regions; dividing the entire time series into multiple overlapping sub-sequence segments using a sliding window; constructing a dynamic brain functional network and transforming it into an optimization model; constructing a hypergraph of the dynamic brain functional network using the nearest neighbor algorithm; dynamically modifying the hypergraph structure through convolution operations and extracting features to obtain a new dynamic hypergraph; extracting the Laplacian matrix of the dynamic hypergraph; constructing the manifold regularization term of the Laplacian matrix and simultaneously introducing the manifold regularization term and the L1 norm regularization term into the optimization model to obtain the hypergraph representation of the brain functional network. This invention is used to represent functional interactions and higher-order relationships between multiple brain regions, determine discriminative brain functional network classification features, and effectively improve the classification performance of brain disease features.
Owner:CHANGZHOU UNIV
Data-driven distributed fault detection method and device
The invention discloses a data-driven distributed fault detection method and a data-driven distributed fault detection device, which are used for analyzing and extracting more valuable information from two levels of global variables and local sub-block variables to implement distributed fault detection. The distributed fault detection method disclosed by the invention not only relates to the application of a brand new algorithm of manifold regularization slow feature analysis, but also provides a plurality of improved technical schemes related to variable quantum block division, neighborhood and reconstruction modes thereof, and whitening processing. In order to overcome the defect that feature analysis is only carried out on each sub-block after blocking is completed in the traditional method, the method provided by the invention inherits the modeling thought of mutual staggering from local sub-blocks to global sub-blocks and then to the local sub-blocks, and comprehensively represents the time sequence and spatial distribution features of learning data changes. Besides, the invention further designs a distributed fault detection device based on the same concept, and the central control module can execute the execution process of a manifold regularization slow feature analysis algorithm and also can autonomously divide variable sub-blocks.
Owner:NINGBO POLYTECHNIC
A training method, device and equipment for an illegal fund-raising risk prediction model
The embodiments of this specification provide a training method, device, and equipment for an illegal fund-raising risk prediction model, which can be used in the field of machine learning technology. The method includes obtaining a training data set related to the enterprise's fund-raising risk; wherein, the training data set includes labeled samples and unlabeled samples; calculating the local density value and clustering membership degree of each sample in the training data set; constructing a pointwise manifold regularization constraint term according to the local density and clustering membership degree of each sample and the discriminant function of a preset classifier; wherein, the pointwise manifold regularization constraint term is used to constrain the relationship between each sample and its neighboring samples; determining the loss function of the preset classifier based on the pointwise manifold regularization constraint term; training the preset classifier based on the training data set and the loss function to obtain an illegal fund-raising risk prediction model. Using the embodiments of this specification can improve the accuracy of predicting the illegal fund-raising risk of enterprises.
Owner:INDUSTRIAL AND COMMERCIAL BANK OF CHINA
An Incomplete Multi-view Animal Image Clustering Method Based on Graph Convolutional Manifold Regularized Deep Spectral Clustering Network
An incomplete multi-view animal image clustering method based on a graph convolutional manifold regularization deep spectral clustering network belongs to the field of animal image clustering processing in image information processing. First, the present invention extracts existing sample features from an incomplete multi-view large batch of animal image datasets and normalizes the feature vectors. Then, an incomplete multi-view deep spectral clustering network (DSCN-IMC) is constructed. This network first uses a multi-view encoder to extract the common features of incomplete multi-view images, obtains an unregularized spectral embedding through a linear layer, and constructs an orthonormalization layer to obtain a regularized constrained view common spectral embedding. Finally, the mini-batch stochastic gradient descent algorithm is used to optimize the manifold-regularized spectral clustering overall loss function until convergence. And k-means clustering is performed on the spectral embedding learned by the network to obtain the final clustering result. Compared with other methods, the present invention has a faster computing speed, a reusable network model, and is relatively less affected by the data missing rate, and has better clustering performance.
Owner:HARBIN UNIV OF SCI & TECH
Hyperspectral image classification method of semantic-guided kernel low-rank sparse preserving projection
PendingCN121937762AImprove class separabilityOvercoming the limitations of fixed neighborhood representationClimate change adaptationCharacter and pattern recognitionImaging processingKernel method
The invention belongs to the technical field of image processing, and particularly relates to a hyper-spectral image classification method for semantic-guided kernel low-rank sparse preserving projection, which comprises the following steps: firstly, acquiring a training sample set and a test sample set of a hyper-spectral image, and calculating a kernel matrix and a cross kernel matrix of the hyper-spectral image; further constructing a kernel method dimension reduction model objective function fusing a sparse reconstruction error term, a sparse regularization term, a semantic guidance low-rank constraint term and a spatial adaptive manifold regularization term; the complex objective function is efficiently solved by adopting an alternating direction multiplier method, and an optimal projection matrix is finally obtained by introducing an auxiliary variable and alternately updating a projection matrix coefficient, a sparse coding matrix and other variables; and mapping the test sample to a low-dimensional feature space by using the projection matrix so as to finish classification. According to the method, through semantic guidance and spatial-spectral information adaptive fusion, the discrimination ability and classification precision of dimension reduction features are significantly improved.
Owner:ANHUI UNIV OF SCI & TECH
Network traffic abnormal behavior identification method based on deep learning
The invention discloses a network traffic abnormal behavior identification method based on deep learning, and relates to the technical field of network security, and the method comprises the steps: constructing a traffic state matrix through traffic sample features, extracting a high-frequency traffic slice, and calculating a maximum Lyapunov index to generate a self-adaptive numerical integration step length; dynamically generating a channel mask vector by combining the power spectrum entropy of the frequency domain transformation; according to the method, an anomaly perception model containing flow inertia and a frequency domain analysis branch is constructed, an inertia branch executes Euler discretization operation by utilizing integral step length to capture cross-scale time domain dynamic characteristics, and a frequency domain branch accurately filters mimicry noise through mask vector gating and an activation energy sorting mechanism; and jointly training the model by using the total loss function fusing manifold regularization and elastic filtering, and outputting a judgment result. According to the method, numerical divergence under chaotic burst traffic is effectively overcome, and the recognition robustness in the face of unknown bandwidth attacks and feature cheating is remarkably improved.
Owner:SHAANXI SCI TECH UNIV
Medical image segmentation method based on contrast manifold regularization and related device
The invention discloses a medical image segmentation method based on contrast manifold regularization and a related device. The method comprises the following steps: after importing medical image data and completing preprocessing, calculating the similarity between samples to construct a manifold regularization item; performing feature decomposition and enhancement on the data to obtain anatomical structure features, pathological semantic features and enhanced versions thereof; for each sample, taking the original-enhanced feature pair of the same sample as a positive sample, taking the cross combination of different sample features as a negative sample, and calculating a comparison loss item; performing dynamic weighted summation on the manifold regularization item and the comparison loss item to form a dynamic comparison manifold regularization item; and an image segmentation model is trained and the weight is obtained, so that accurate segmentation is realized. According to the method, the manifold structure and the cross-sample contrast constraint are fused, so that the image segmentation precision, robustness and generalization ability in a small sample scene are remarkably improved, and meanwhile, the dependence on a large amount of labeled data is reduced.
Owner:ZHUHAI HENGQIN ALL-STAR MEDICAL TECHNOLOGY CO LTD
A medical image segmentation method and related device based on semi-supervised adapter fine-tuning and prompt learning based on SAM
This application discloses a medical image segmentation method and related device based on semi-supervised adapter fine-tuning and prompt learning based on SAM. The method includes: importing and preprocessing a medical image dataset, loading a pre-trained SAM model and a YOLO prompt model, inserting dual adapters into each visual transformer block of the SAM encoder, and embedding a super-prompt adapter, an MLP alignment adapter, and a cross-attention adapter into the decoder; training the prompt model based on labeled data, screening qualified unlabeled data through KL divergence, generating click prompts, and performing pseudo-label segmentation using the SAM model; constructing a joint loss function containing a contrast manifold regularization term, and performing adapter fine-tuning based on the pseudo-label and labeled data while freezing the SAM backbone. The device includes a data import unit, a model loading unit, and a prompt training unit. The present invention effectively reduces labeling dependency, enhances model generalization and segmentation accuracy, and retains the original performance advantages of the SAM backbone.
Owner:ZHUHAI HENGQIN ALL-STAR MEDICAL TECHNOLOGY CO LTD
A console login identity information collection and verification method based on deep learning
ActiveCN120263423BUser identity/authority verificationBiological modelsFeature vectorComputation process
The application relates to the technical field of information security and discloses a console login identity information acquisition and verification method based on deep learning, which comprises the following steps: S1, collecting biological characteristics and behavioral characteristics of a user, extracting a feature vector of the biological characteristics by using a deep learning model, extracting a feature vector of the behavioral characteristics by using a time sequence model, splicing to obtain an identity feature vector, and splicing the feature vector of the biological characteristics and the feature vector of the behavioral characteristics to construct the identity feature vector; S2, dimensionally reducing the identity feature vector by using a topology-preserving projection method and constructing a low-dimensional manifold embedding representation; and S3, adding a manifold regularization term in a loss calculation process of the deep learning model. The application jointly extracts identity features by using deep learning and a time sequence model, achieves the effect of quickly and accurately acquiring biological and behavioral information, and solves the problems of high misrecognition rate and low recognition efficiency compared with a single feature recognition scheme in the prior art.
Owner:MT TITLIS BEIJING CONTROL TECH
Context pre-training model depolarization framework and equipment based on prefix tuning
The invention discloses a context pre-training model depolarization framework and equipment based on prefix tuning, and belongs to the technical field of artificial intelligence. The depolarization framework freezes original pre-training model parameters, optimizes prefix parameters through a specific loss function, reduces deviation and maintains semantic modeling capability; the depolarization framework comprises: an embedding mapping module configured to input a sentence containing a specific attribute word, output a corresponding embedding vector, and capture a sentence semantic representation; the depolarization loss function module is configured to perform joint optimization by adopting manifold regularization loss, orthogonalization loss and regularization loss so as to balance a depolarization effect and semantic retention; and the prefix parameter optimization module is configured to be capable of freezing original pre-training model parameters and only optimizing prefix parameters so as to reduce consumption of computing resources. According to the method, the deviation problem in the pre-training language model can be solved, and meanwhile the semantic modeling capability of the model is kept.
Owner:GUANGDONG UNIV OF TECH
A medical image segmentation method based on contrast manifold regularization and related device
ActiveCN120526154BImage analysisCharacter and pattern recognitionAnatomical structuresDynamic contrast
The application discloses a medical image segmentation method based on contrast manifold regularization and related devices. The method comprises: after importing medical image data and completing preprocessing, calculating the similarity between samples to construct a manifold regularization term; performing feature decomposition and enhancement on the data to obtain anatomical structure features, pathological semantic features and their enhanced versions; for each sample, the original-enhanced feature pair of the same sample is taken as a positive sample, different sample feature cross combinations are taken as negative samples, and a contrast loss term is calculated; the manifold regularization term and the contrast loss term are dynamically weighted and summed to form a dynamic contrast manifold regularization term; the image segmentation model is trained and the weight is obtained, and then precise segmentation is realized. The application significantly improves the image segmentation accuracy, robustness and generalization ability in the small sample scene by fusing the manifold structure and cross-sample contrast constraint, and reduces the dependence on a large amount of labeled data.
Owner:ZHUHAI HENGQIN ALL-STAR MEDICAL TECHNOLOGY CO LTD
Textile multi-modal feature selection method and system based on hessian manifold regularization
The application discloses a textile multi-modal feature selection method and system based on a Hessian manifold regularization, and relates to the technical field of computer vision, which comprises the following steps: constructing multi-modal data by a jacquard chart image and a process single parameter, and obtaining each modality pseudo label matrix through subspace projection; introducing L2, 1 norm sparse constraint and Hessian manifold regularization to maintain local curvature structure and improve sparsity; generating cross-modal shared pseudo labels by using adaptive weighted fusion, and optimizing in cooperation to enhance consistency; further tensorizing the pseudo labels, applying tensor kernel norm low rank constraint, and taking into account modality commonality and specificity; finally, all components are optimized in a unified framework, and a high discriminability and structure robust feature subset is output. The application realizes efficient feature selection without manual annotation by fusing subspace learning, Hessian manifold regularization and tensor kernel norm analysis.
Owner:HUAQIAO UNIVERSITY
Semi-supervised medical image segmentation method based on contrast manifold regularization and related device
The invention discloses a semi-supervised medical image segmentation method based on contrast manifold regularization and a related device. The method comprises the following steps: importing and preprocessing a medical image data set; initializing a teacher model and a student model, training the teacher model by using the annotation data, and generating a pseudo tag; calculating the similarity between the annotation data and the pseudo tag to obtain a manifold regularization item; constructing positive and negative sample pairs and calculating and comparing loss items; performing weighted summation to obtain a contrast manifold regularization item, and combining the contrast manifold regularization item with supervision loss to serve as a loss function of the student model; and iteratively training the student model to obtain a segmentation result. Compared with the prior art, by combining comparative learning and manifold regularization, the problem of data dependence in semi-supervised medical image segmentation is effectively relieved, the model generalization ability and segmentation precision are improved, and the method is particularly excellent in performance in a small target focus segmentation scene.
Owner:ZHUHAI HENGQIN ALL-STAR MEDICAL TECHNOLOGY CO LTD