Product
Eureka
For R&D, Eureka makes reading and utilizing patents & technical documents easy.
Eureka AIR
Designed for self-driven R&D workflows. Generate viable solutions, solve complex R&D challenges, empower your innovation with AI.
Eureka Materials
Designed for material experts only. Revolutionize your material R&D, from search, analyze, to developing new materials.
Feature
TechResearch
Generate reliable direction feasibility study reports for your R&D in just a few steps.
TechSeek
Discover and master advanced knowledge NOW. Basics, ideas, possibilities, all at once.
TechMind
As an expert in R&D Theories, TechMind can generates customized viable solutions instantly.
TechRisk
Analyze your overall solution with one click, know your potential R&D risks in advance.
TechMonitor
Get weekly tech updates, stay abreast of the latest tech innovations and key insights.
Patents
Literature
Hiro is an intelligent assistant for R&D personnel, combined with Patent DNA, to facilitate innovative research.
681 results about "Sample graph" 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
Zero sample image classification method and system
InactiveCN107563444AImprove discriminationImprove adaptabilityCharacter and pattern recognitionFeature extractionNear neighbor
The invention discloses a zero sample image classification method and system. The method includes the following steps of: S10, inputting training data belonging to a visible category and a category tag thereof for feature extraction; S20, inputting the semantic assistance information of all category tags to obtain the semantic embedded representations of respective tags and semantic difference measure between the tags; S30, establishing a zero sample classification model, performing semantic consistency regularization on the model based on the semantic differences between the categories; S40,iteratively updating model parameters until convergence; and S50, predicting the category tag of a test image. The method and system establish a semantic consistency regularized zero sample classification model so that the output of the model conforms to the semantic neighborhood relation between categories to adapt to the semantic structure of the target category so as to obtain high classification accuracy.
Owner:ZHEJIANG UNIV
Method to quantify discrete pore shapes, volumes, and surface areas using confocal profilometry
Owner:SCHLUMBERGER TECH CORP
Data processing method and device, medium and computing equipment
PendingCN109934249ADiscriminative features that help distinguish whether an image is a positive sample or a negative sampleDiscriminative featuresCharacter and pattern recognitionStill image data queryingPositive sampleSample image
The embodiment of the invention provides a data processing method. The data processing method comprises the following steps: acquiring a plurality of sample images; Adding a label to the plurality ofsample images, adding a positive sample label to the sample image including a predetermined feature, and adding a negative sample label to the sample image not including the predetermined feature; Establishing a neural network classification model based on an attention mechanism; And training the neural network classification model by using the sample image added with the label to obtain an optimal classification model. According to the scheme, an attention mechanism is introduced into a neural network classification model as an initial training model; A neural network classification model with an attention mechanism introduced in the training process can extract discriminative features which are more beneficial to distinguishing whether the image is a positive sample or a negative sample,and then an optimal classification model which can more sensitively and accurately judge whether the image contains predetermined features or not is obtained. The embodiment of the invention furtherprovides a data processing device, a medium and computing equipment.
Owner:杭州网易智企科技有限公司
Image semantic segmentation method and device based on codec
PendingCN111292330AImprove Segmentation AccuracyImage enhancementImage analysisSample graphRadiology
The embodiment of the invention provides an image semantic segmentation method and device based on a codec. The method comprises the following steps: inputting a to-be-detected image into an encoder of a preset image semantic segmentation network model, extracting features by using a convolutional network, respectively inputting the features into a plurality of pooling layers with different sizes,and performing feature fusion according to output results of the plurality of pooling layers with different sizes to obtain a high-level semantic feature map of the to-be-detected image; inputting the feature map into a decoder of an image semantic segmentation network model to obtain a detection result of semantic analysis, wherein the image semantic segmentation network model is obtained by training a sample image with a determined semantic label. Due to the fact that local and global information is fused in the pooling layers of different sizes, learning of targets of different sizes is facilitated through the multi-scale feeling domain, and therefore an accurate high-level semantic feature map of the image to be detected can be obtained. After the decoder is used for analysis, the segmentation precision of the target boundary in the detection result of semantic analysis is higher.
Owner:BEIJING UNIV OF TECH
Zero-sample image classification method based on regression variation auto-encoder
ActiveCN111563554AMake up for the problem of missing samples of unknown classesEasy to classifyCharacter and pattern recognitionNeural architecturesSample graphGraphics
The invention discloses a zero-sample image classification method based on a regression variation auto-encoder. The invention relates to a method for identifying graphs by using electronic equipment.A variational auto-encoder is reconstructed by training, aligning and crossing; calculating regression network loss, the global loss L of the whole model network is calculated, the classifier is trained, the classification accuracy is calculated, zero-sample image classification based on the regression variation auto-encoder is completed, and the defects that in the prior art of generalized zero-sample classification, training samples are deficient, generated samples lack semantics, and a generative adversarial network is prone to collapse are overcome.
Owner:HEBEI UNIV OF TECH
Face identification method and device based on multiple models
InactiveCN107563280AImprove the correct recognition rateThe correct recognition rate is lowCharacter and pattern recognitionConcurrent computingSample image
The invention discloses a face identification method and device based on multiple models. The face identification method comprises the steps of firstly dividing the sample data according to the defined classification, and generating a plurality of sample files, each of which includes a kind of sample data; secondly conducting preprocessing on the data samples of all the sample files, and detectingfaces in the sample images and conducting normalization; thirdly obtaining different models from the data sample of each sample file after preprocessing by using different training methods like different face identification algorithms or the same face identification algorithm, wherein the data sample of the same sample file corresponds to multiple models; finally adopting multiple models to conduct identification on faces to be identified, and obtaining multiple face identification results and selecting the final face identification result. The invention is advantageous in that the face identification accuracy can be increased; model training time can be reduced; with the adoption of multiple models for concurrent computing, time of searching and contrasting can be reduced.
Owner:南京道熵信息技术有限公司
All-weather video monitoring method based on deep learning
ActiveCN104320617AAdaptableImprove accuracyImage analysisClosed circuit television systemsPattern recognitionVideo monitoring
The invention discloses an all-weather video monitoring method based on deep learning. The all-weather video monitoring method based on deep learning includes the following steps that video streaming is real-timely collected, and multiple original sampled graph samples and speed sampled graph samples are obtained through line sampling on basis of the obtained video streaming; the obtained speed sampled graph samples are subjected to space-time correction; on basis of original sampled graphs and speed sampled graphs, off-line training is performed to obtain a deep learning model, and the deep learning model comprises a classification model and a statistical model; the real-time video streaming is subjected to crowd state analysis by means of the obtained deep learning model. According to the all-weather video monitoring method based on deep learning, good adaptability can be achieved in terms of different environments, illumination intensities, weather situations and camera angles, high accuracy can be guaranteed in terms of crowding environments such as rushing out of mass flow crowds, the calculated amount is small, requirements of real-time video processing can be met, and the all-weather video monitoring method based on deep learning is widely applicable to monitoring and managing of public places such as buses, subways and squares where stranded people are dense.
Owner:INST OF AUTOMATION CHINESE ACAD OF SCI
Network training method, image processing method, network, terminal device and medium
ActiveCN110660066APrecise contour edgesGuaranteed approximationImage enhancementImage analysisNetwork terminationSample graph
The invention provides a network training method, an image processing method, a network, terminal equipment and a medium. The training method comprises the following steps: S1, acquiring a sample image accommodating a target object, a sample mask corresponding to the sample image and sample edge information corresponding to the sample mask; s2, inputting the sample image into an image segmentationnetwork to obtain a generated mask output by the image segmentation network; s3, inputting the generated mask into the trained edge neural network to obtain generated edge information output by the edge neural network; s4, determining a loss function according to the difference between the sample mask and the generated mask and the difference between the generated edge information and the sampleedge information; and S5, adjusting each parameter of the image segmentation network, and returning to the step S2 until the loss function is smaller than the threshold. According to the invention, the mask image output by the image segmentation network can represent the contour edge of the target object more accurately.
Owner:GUANGDONG OPPO MOBILE TELECOMM CORP LTD
Training sample generation method for industrial big data processing
ActiveCN109583474AReduce mistakesAvoid manual interventionCharacter and pattern recognitionSample graphData set
Owner:HUAZHONG UNIV OF SCI & TECH
Image classification method based on reliable weight optimal transmission
PendingCN111814871AImprove efficiencyReduce negative transferCharacter and pattern recognitionNeural architecturesSample graphData set
The invention discloses an image classification method based on reliable weight optimal transmission, and the method comprises the following steps: firstly carrying out the preprocessing of source domain data, and enabling a deep neural network to fit a sample label of a source domain sample image; marking a picture, marking a pseudo label on the target domain data sample, pairing nodes to realizepairing of associated pictures in a source domain and a target domain, and finally realizing automatic analysis through a feature extractor and a self-adaptive discriminator to classify the images. The invention provides a subspace reliability method for dynamically measuring sample inter-domain differences by utilizing space prototype information and an intra-domain structure. The method can beused as a pretreatment step of an adaptive technology in the prior art, and the efficiency is greatly improved. According to the method, the reliability of the contraction subspace is combined with the optimal transportation strategy, so that the depth characteristics are more obvious, and the robustness and effectiveness of the model are enhanced. The deep neural network works stably on various data sets, and the performance of the deep neural network is superior to that of an existing method.
Owner:ZHEJIANG UNIV
Target detection model training method and device and electronic equipment
PendingCN110827253AImprove detection accuracyImage enhancementImage analysisPositive sampleSample graph
The embodiment of the invention provides a target detection model training method and device and electronic equipment, and the method comprises the steps: selecting a sample image from a preset training image set, inputting the sample image into a to-be-trained network model, and obtaining the prediction position information and prediction probability of a prediction target in the sample image; calculating a first loss of the to-be-trained network model according to the prediction position information and a preset regression function; calculating a second loss of the to-be-trained network model according to the prediction probability and a preset focus loss function; calculating target loss of the to-be-trained network model through a preset target loss function; and adjusting the to-be-trained network model according to the target loss to obtain a trained network model. According to the technical scheme of the invention, the weight correction is carried out according to the proportions of different positive samples through the preset focus loss function, so that the model obtained through training can adapt to different positive sample proportions, and the detection precision of the target in the to-be-detected image is improved.
Owner:BEIJING DAJIA INTERNET INFORMATION TECH CO LTD
Model generation method, target detection method, device, electronic equipment and medium
PendingCN112257815AImprove generalization abilityOptimize detection resultsCharacter and pattern recognitionNeural architecturesSample graphManual annotation
The embodiment of the invention discloses a model generation method, a target detection method, a device, electronic equipment and a medium. The model generation method comprises the steps of trainingan original detection model based on multiple groups of teacher training samples including a first sample image and a sample labeling result of a known target in the first sample image to obtain a teacher network; taking the first sample image and a first detection result obtained after the first sample image is input into a teacher network as a first training sample, and taking the second sampleimage and a second detection result obtained after the second sample image is input into the teacher network as a second training sample; and training a student network having the same network type as the teacher network based on the plurality of groups of first training samples and the plurality of groups of second training samples to generate a target detection model. According to the technicalscheme provided by the embodiment of the invention, the effect of improving the generalization performance of the target detection model is achieved under the condition of not additionally increasingthe manual annotation cost.
Owner:BEIJING WODONG TIANJUN INFORMATION TECH CO LTD +1
Instance segmentation model training method and device and image processing method and device
InactiveCN112016559AImprove accuracyImprove efficiencyCharacter and pattern recognitionNeural architecturesSample graphImaging processing
The invention discloses an instance segmentation model training method and device and an image processing method and device. The training method comprises the following steps: determining a sample image containing at least one target object, wherein the sample image comprises a shape edge point label, a target center point label and a category label of each target object in the at least one targetobject; training a neural network with a branch structure based on the sample image, obtaining the instance segmentation model, wherein the neural network model with the branch structure comprises abackbone network used for feature extraction and a plurality of parallel branch networks located behind the backbone network, the plurality of branch networks comprise a first branch network and a second branch network, the first branch network is used for outputting a classification result of each target object in the at least one target object, and the second branch network is used for outputting a segmentation result of the target object, so that the accuracy and efficiency of an instance segmentation result can be improved.
Owner:INFERVISION MEDICAL TECH CO LTD
Method and device for carrying out analysis on sensitive images
ActiveCN108734184AIncrease the level of automationReduce manual review costsCharacter and pattern recognitionCommerceSample imagePicture recognition
The embodiment of the invention discloses a method and device for carrying out analysis on sensitive images, relates to the technical field of image identification, and can promote an automation levelof carrying out identification detection on advertising images and reduce manual review cost. The method comprises the steps of: carrying out clustering on sample images in a training sample set; then according to the clustered sample images, by a convolutional neural network, training an identification model corresponding to each category; and then by utilizing the identification model obtainedby training, identifying sensitive images corresponding to each category from a to-be-detected image library. The method and device disclosed by the invention are applicable to identification on the sensitive images on an online platform.
Owner:SUNING COM CO LTD
Lane line detection method and device, terminal equipment and readable storage medium
PendingCN112528878ASolve the large amount of calculationSolve the speed problemCharacter and pattern recognitionNeural architecturesSample graphComputer graphics (images)
The invention is applicable to the technical field of computer vision and image processing, and provides a lane line detection method and device, terminal equipment and a readable storage medium. Inputting the road image into a trained neural network model for processing, and outputting a detection result of a lane line in the road image of the current scene; wherein the trained neural network model is obtained by training according to sample images in a training set and a semantic segmentation model, and the sample images in the training set comprise collected road images of a plurality of scenes and marked images corresponding to the road images of the plurality of scenes. According to the method and the device, the problems that most of deep learning models used for lane recognition atpresent are relatively large in calculation amount, relatively complex in model and unfavorable for meeting the requirement on real-time performance in an actual application scene of an automatic driving task can be solved.
Owner:SHENZHEN INST OF ADVANCED TECH CHINESE ACAD OF SCI
Book content distribution system and content server
InactiveUS20120136892A1Web data indexingDigital data processing detailsContent distributionThumbnail
A content server retrieves a book content from a database in response to a book content search request, the book content meeting the book content search request transmitted from a user terminal; creates sample image data having a plurality of pages, with regard to the retrieved book content; and distributes search result information including the sample image data to the user terminal. The user terminal displays, when the search result information is received, a search result screen on which an image of a first page of a sample image of the book content is displayed as a thumbnail based on the sample image data of the book content; and, when a user provides a page-turning operation instruction to the sample image of the book content displayed on the search result screen, makes pages of the sample image turned page by page on the search result screen.
Owner:FOR SIDE COM
Method for acquiring sample images for inspecting label among auto-labeled images to be used for learning of neural network and sample image acquiring device using the same
ActiveUS10373027B1Improve throughputLarge throughputCharacter and pattern recognitionNeural architecturesSample imageAnnotation
A method for acquiring a sample image for label-inspecting among auto-labeled images for learning a deep learning network, optimizing sampling processes for manual labeling, and reducing annotation costs is provided. The method includes steps of: a sample image acquiring device, generating a first and a second images, instructing convolutional layers to generate a first and a second feature maps, instructing pooling layers to generate a first and a second pooled feature maps, and generating concatenated feature maps; instructing a deep learning classifier to acquire the concatenated feature maps, to thereby generate class information; and calculating probabilities of abnormal class elements in an abnormal class group, determining whether the auto-labeled image is a difficult image, and selecting the auto-labeled image as the sample image for label-inspecting. Further, the method can be performed by using a robust algorithm with multiple transform pairs. By the method, hazardous situations are detected more accurately.
Owner:STRADVISION
Neural network training method and device, electronic equipment and storage medium
PendingCN111368923AEasy to classifyImprove classification efficiencyNeural architecturesNeural learning methodsSample graphEngineering
The invention relates to a neural network training method and apparatus, an electronic device and a storage medium. The method comprises the steps of obtaining position information and category information of a target area in a sample image; cutting out at least one target area according to the position information of the target area; according to the category information, classifying the at leastone cut target area to obtain N types of sample image blocks; and inputting the N types of sample image blocks into a neural network for training. According to the neural network training method provided by the embodiment of the invention, the fine classification of the sample image blocks can be obtained, and the neural network is trained, so that the neural network can finely classify the images, the classification efficiency is improved, and the medical diagnosis accuracy is improved.
Owner:SHANGHAI SENSETIME INTELLIGENT TECH CO LTD
Product edge defect detection method
ActiveCN113450307ADetection clearImprove accuracyImage enhancementImage analysisSample graphData set
The invention relates to a product edge defect detection method, which comprises the following steps that firstly, a template image and a sample image are input, and then a feature map is constructed by acquiring information such as a centroid, a target contour, a centroid and a deflection angle from the template image and the sample image; an iterative optimization method of a mapping model from a sample image to a template image is constructed by taking a minimum residual sum as a loss function, so that feature map matching is carried out, a global mapping matrix is obtained according to a matching result, a difference image process is completed, and a coarse segmentation region is obtained by using an adaptive threshold value; secondly, for the problem of scarcity of industrial image samples, a large number of artificial defect sample pre-training data sets are obtained through modeling defects, the sample sets are used for pre-training the multi-scale integrated residual neural network, and then a real defect sample set is used for migration training; the obtained result is used for carrying out defect type identification on the coarse segmentation region.
Owner:XIDIAN UNIV
Image scene classification method based on target and space relationship characteristics
InactiveCN102902976AImprove classification accuracyCharacter and pattern recognitionClassification methodsSample image
The invention discloses an image scene classification method based on target and space relationship characteristics and relates to image scene classification technologies. The method comprises the steps of: defining a space relationship histogram, conducting representation on the space relationship between targets, comprising left, right, top, bottom, far, near, including and excluding, and giving a calculation method; labeling a target in a sample image, assigning the membership degree of the space relationship between any two targets, counting mathematical features of the membership degree of the space relationship between any two targets in the scene, classifying the space relationship histogram between the targets by using a fuzzy K neighbor classifier according to test images, and calculating the membership degree of the space relationship; establishing an image model by employing a probability latent semantic analysis model of the space relationship characteristics between fusion themes; and classifying the scene images by using a support vector machine. According to the method, the image is modeled by employing the probability latent semantic analysis model of the space relationship characteristics between fusion themes, and the scene images are classified through input of the support vector machine.
Owner:INST OF ELECTRONICS CHINESE ACAD OF SCI
Method and device for recognizing text in image, computer equipment and computer storage medium
PendingCN111291629AImprove accuracyHigh scene recognition abilityCharacter and pattern recognitionNeural architecturesSample graphText recognition
The invention discloses a method and device for recognizing a text in an image, and a computer storage medium, relates to the technical field of text recognition, and aims to expand sample data collected in an actual scene, so that a trained model can well fit the actual scene, and the accuracy of recognizing the text in the image is improved. The method comprises the steps of obtaining a character sample image of a stylus-like printing font after scene processing; respectively inputting the character sample images of the needle-like printed fonts into network models of different architecturesas training data for training to obtain a text region detection model and a text recognition model; when an image text detection request is received, inputting an image requested to be detected intothe text area recognition model, and determining position information of a text area corresponding to the image; and jointly inputting the position information of the text area corresponding to the image and the image requested to be detected into the text recognition model to obtain text information in the image.
Owner:深圳平安医疗健康科技服务有限公司
Pedestrian re-identification model optimization processing method and device and computer equipment
ActiveCN111860147AOvercoming the problem of dividing features evenlyImprove recognition accuracyCharacter and pattern recognitionNeural architecturesSample graphData set
The invention relates to a pedestrian re-identification model optimization processing method and device and computer equipment. The method comprises the following steps: performing network layer deletion on an original pedestrian re-identification model corresponding to a model identifier and modifying a convolution stride of a specified network layer to obtain a backbone network model; performingfeature extraction on each sample image in the sample data set through a backbone network model to obtain initial feature data; performing batch standardization processing on the initial feature datato obtain a plurality of feature maps; constructing a plurality of attention branch network models according to the plurality of feature maps and a preset network layer of the backbone network model;and combining the backbone network model and the plurality of attention branch network models to obtain an optimized pedestrian re-identification model, training the optimized pedestrian re-identification model through the sample data set and the plurality of loss function relationships until a preset condition is met, stopping model training, and outputting the trained pedestrian re-identification model. By adopting the method, computing resources can be saved.
Owner:BEIJING WEIFU SECURITY & PROTECTION TECH CO LTD
Target detection model training method, target recognition method, devices and medium
PendingCN111814902AIncrease diversityImprove accuracyGeometric image transformationCharacter and pattern recognitionSample graphSample image
Owner:CHINA SOUTHERN POWER GRID DIGITAL GRID RES INST CO LTD
Method for distinguishing fog concentration in intelligent image monitoring of power transmission line
InactiveCN109961070AAccurate identificationHigh concentration classification accuracyCharacter and pattern recognitionSynthesis methodsSample image
The invention discloses a method for distinguishing fog concentration in intelligent image monitoring of a power transmission line, which adopts a multi-feature synthesis method and comprises the following steps of: extracting a plurality of features related to fog in a picture, and synthesizing the plurality of features extracted from each image into a feature parameter table to form a two-dimensional feature map; utilizing a deep learning convolutional neural network to train and model the characteristic parameter graphs of a large number of sample images; and judging whether the unknown image has fog or not by using the trained model, and distinguishing the concentration levels of the fog, namely clear fog, medium fog and dense fog. According to the method, a plurality of deep and irrelevant features of the provider are integrated to serve as image feature parameters, and then a nonlinear classification algorithm model is trained for a large number of sample images by using the convolutional neural network, so that the problem that the accuracy of detecting whether the image is foggy is low is solved.
Owner:STATE GRID HEBEI ELECTRIC POWER RES INST +2
Apparatus and method for training classifying model
An apparatus for training a classifying model comprises: a first obtaining unit configured to input a sample image to a first machine learning framework, to obtain a first classification probability and a first classification loss; a second obtaining unit configured to input a second image to a second machine learning framework, to obtain a second classification probability and a second classification loss, the two machine learning frameworks having identical structures and sharing identical parameters; a similarity loss calculating unit configured to calculate a similarity loss related to a similarity between the first classification probability and the second classification probability; a total loss calculating unit configured to calculate the sum of the similarity loss, the first classification loss and the second classification loss, as a total loss; and a training unit configured to adjust parameters of the two machine learning frameworks to obtain a trained classifying model.
Owner:FUJITSU LTD
And searching reliable semi-supervised few-sample image classification method of abnormal data center
PendingCN110942091AImprove accuracyImprove image recognition accuracyCharacter and pattern recognitionNeural architecturesSample graphData set
The invention discloses a semi-supervised few-sample image classification method for searching a reliable abnormal data center. The method specifically comprises the following steps: dividing a data set; sampling a semi-supervised few-sample classification task from the training set; extracting feature representation of the few-sample classification task samples by using a neural network; searching a reliable abnormal data clustering center; optimizing various image prototypes by utilizing label-free data; classifying to-be-classified samples in the task by utilizing the prototype, calculatingcross entropy loss, and performing back propagation to update network parameters; performing iterative training to obtain an ideal feature extraction network; and completing a semi-supervised few-sample classification task. According to the method, the feature extractor suitable for few-sample classification is trained, so that the classifier can still obtain relatively ideal classification performance under the condition of extremely few training data. And label-free data is added during training, a reliable abnormal data center searching method is utilized, information of the label-free data is reasonably utilized, and the performance of the classifier is improved.
Owner:WUHAN UNIV OF TECH +1
Calibration method and device, roadside equipment and computer readable storage medium
ActiveCN112017251ANot affectedRealize CalibrationImage enhancementImage analysisSample graphPoint cloud
The invention provides a calibration method and device, roadside equipment and a computer readable storage medium, and the method comprises the steps: obtaining a sample image and sample point cloud data; calculating a first conversion matrix of the coordinate system of a to-be-calibrated object relative to the coordinate system of a data acquisition device; obtaining the overlap ratio of the edgepoints projected to the sample image and the edge features of the sample image; and optimizing the first conversion matrix to obtain an optimized first conversion matrix so as to calibrate the to-be-calibrated object. According to the technical scheme, point cloud data are acquired through the data acquisition device, and the alignment characteristic of the edge characteristics of the point clouddata and the sample image is utilized to optimize the first conversion matrix of the coordinate system of the to-be-calibrated object and the coordinate system of the data acquisition device, so thatthe to-be-calibrated object is calibrated, the whole calibration process is simple and rapid, is not influenced by the outdoor environment of a road, and the method is more flexible to use and has higher calibration accuracy.
Owner:杭州飞步科技有限公司
Training method of image label classification network, image label classification method and equipment
PendingCN111582409AImprove classification efficiencyImprove classification accuracyCharacter and pattern recognitionNeural architecturesSample graphFeature extraction
Owner:TENCENT TECH (SHENZHEN) CO LTD
Method for generating unbiased deep learning model based on transfer learning
PendingCN112115963AEnsure fairnessImprove accuracyCharacter and pattern recognitionNeural architecturesSample graphData set
The invention discloses a method for generating an unbiased deep learning model based on transfer learning. The method comprises the following steps: (1) constructing a task label with a sample imageand an original data set of a biased label; (2) training a biased deep learning model by using the original data set; (3) constructing and training an adversarial attack network, and attacking the original data set by using the trained adversarial network without prejudice; (4) training an initial unbiased deep learning model with the same structure as the biased deep learning model by utilizing the unbiased data set; and (5) preparing a third feature extractor, and forming an unbiased deep learning model by the third feature extractor with determined parameters of the third feature extractorand a second classifier included in the trained initial unbiased deep learning model based on a transfer learning strategy, so as to ensure the fairness of the deep learning model during automatic decision making according to the input image. The accuracy of image recognition is improved.
Owner:ZHEJIANG UNIV OF TECH
Image recognition method and device and terminal equipment
PendingCN111079785AEasy to trainImprove image recognition accuracyCharacter and pattern recognitionNeural architecturesSample graphFeature vector
The invention is suitable for the technical field of image processing, and provides an image recognition method and device, and terminal equipment, and the method comprises the steps: obtaining a sample training set; executing feature extraction operation, and inputting sample image groups in the sample training set into a twin neural network to obtain a first feature vector and a second feature vector, wherein the first feature vector and the second feature vector are normalized feature vectors; calculating a loss value according to the first feature vector, the second feature vector, samplemarks and a cosine distance-based comparison loss function; if the loss value is greater than a preset loss threshold and the iteration frequency is less than the preset iteration frequency, updatingthe twin neural network according to the loss value, adding 1 to the iteration frequency, and returning to the feature extraction operation; and if the loss value is less than or equal to the preset loss threshold or the number of iterations is greater than or equal to the preset number of iterations, performing image recognition on to-be-processed image groups by using the twin neural network. The problem that existing conventional image recognition method is low in accuracy can be solved.
Owner:SHENZHEN INTELLIFUSION TECHNOLOGIES CO LTD
Who we serve
- R&D Engineer
- R&D Manager
- IP Professional
Why Eureka
- Industry Leading Data Capabilities
- Powerful AI technology
- Patent DNA Extraction
Social media
Eureka Blog
Learn More Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic, Popular Technical Reports.
© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap|About US| Contact US: help@patsnap.com