516 results about "Supervised training" patented technology

Semantic indexing and retrieval of multimedia content requires that the content is sufficiently annotated. However, the great volumes of multimedia data and diversity of labels make annotation a difficult and costly process. Disclosed is an annotation framework in which supervised training with partially labeled data is facilitated using active learning. The system trains a classifier with a small set of labeled data and subsequently updates the classifier by selecting a subset of the available data-set according to optimization criteria. The process results in propagation of labels to unlabeled data and greatly facilitates the user in annotating large amounts of multimedia content.

The invention discloses a three-dimensional image segmentation method and system based on a full convolutional neural network, and the method comprises the following steps: step1, collecting and obtaining a sequence image, and carrying out the marking of the sequence image, and obtaining training sample data; step 2, performing normalization preprocessing on the training sample data obtained in the step 1; step 3, applying the sample data processed in the step 2 to carry out supervised training on pre-constructed 3-D full convolution residual U-net network model, training to a preset convergence condition, and obtaining a trained three-dimensional image segmentation model; and step 4, after normalization processing is carried out on the sequence image data to be segmented, inputting the sequence image data into the three-dimensional image segmentation model trained in the step 3, and obtaining a sequence image segmentation result. The continuity information of the sequence can be fullyutilized, and a relatively good result can be obtained in three-dimensional image segmentation.

The invention provides a semi-supervised X-ray automatic labeling method based on a generative adversarial network. A traditional training method is improved on the basis of an existing generative adversarial network method, and a semi-supervised training method combining supervised loss and unsupervised loss is used for carrying out image classification recognition based on a small number of labeled samples. The problem of data scarcity annotation of the X-ray image is studied. The method comprises: firstly, on the basis of a traditional unsupervised generative adversarial network, using a softmax for replacing a final output layer; expanding the X-ray image into a semi-supervised generative adversarial network, defining additional category label guide training for the generated sample, optimizing network parameters by adopting the semi-supervised training, and finally, automatically labeling the X-ray image by adopting a trained discriminant network. Compared with traditional supervised learning and other semi-supervised learning algorithms, the method has the advantage that in the aspect of medical X-ray image automatic labeling, the performance is improved.

The invention discloses a face identification method based on a random pooling convolutional neural network. According to the method, the characteristics of a face image are quickly extracted by the random pooling convolutional neural network and cascaded to realize face identification; selection strategies and steps of new pooling values are adopted in a process of creating the convolutional neural network and then supervised training is carried out by a softmax classifier; the probability distribution used in a sampling process is based on energy, and the effect of optimizing increment of the calculation speed of the characteristics extracted by the convolutional neural network and generalization application of a convolutional neural network training result can be achieved; the convolutional neural network training based on random pooling is simple and high in accuracy, and can promote wide application of random pooling in the process of extracting the face identification characteristics.

The invention discloses a multi-view SAR image target recognition method based on a depth neural network, and the method comprises three steps: image preprocessing, feature extraction based on CAE, and multi-view SAR image recognition based on RNN. The method specifically comprises the steps: firstly carrying out the cutting and energy normalization of an inputted image; secondly extracting the features of an original image through the non-supervision training of CAE; thirdly constructing a multi-view SAR image feature sequence through the above features; fourthly carrying out the supervised training of the RNN through employing a training feature sequence, wherein the RNN can be used for the recognition of a testing set feature sequence after training. The method can make the most of the capability of CNN in learning and extracting the general features of the image and the capability of RNN in fully extracting the context of the sequence, effectively improves the recognition rate of a multi-view SAR image target, and is higher in engineering value.

The invention provides a face recognition method based on a convolutional neural network. The face recognition method comprises the following steps: carrying out necessary pretreatment at early stage on a facial image to obtain an ideal facial image; selecting the ideal facial image as the input of the convolutional neural network to enter U0, wherein the output of the U0 enters UG, and the output of the UG is taken as the input of US1; extracting edge components in different directions in an input image as first-time feature extraction and outputting to the input of special UC1 through supervised training by the S nerve cell of the US1; taking the output of the UC1 as the input of US2, completing the second-time feature extraction by the US2 and taking as the input of UC2; taking the output of the UC2 as the input of US3, and completing the third-time feature extraction by the US3 and taking as the input of UC3; taking the output of the UC3 as the input of US4, and obtaining the weight, threshold value and neuron plane number in each layer in a supervision competitive learning mode by the US4 and taking as the input of the UC4; taking the UC4 as the output layer of the network, and outputting the final mode recognition result of the network determined by the maximum output result of the US4. According to the face recognition method, the recognition rate of faces in complex scenarios can be improved.

The invention discloses a cross-domain pedestrian re-identification method and system based on three stages, and the method comprises the steps: a domain adaptive learning stage: carrying out the processing of a source domain image and a target domain image through a domain adaptive network, calculating each loss, and updating the parameters of the domain adaptive network; a self-supervised training stage: carrying out the supervised training on the domain self-adaptive network through pseudo tags, calculating the loss of the triad difficult to sample, and updating network parameters; a jointloss training stage: constructing a joint loss training network, and defining label smooth regularization loss and difficult-to-sample triple loss; for a target domain image, inputting a joint loss training network, calculating the losses and updating the parameters of the joint loss training network, and carrying out the re-identification of cross-domain pedestrians. According to the cross-domainpedestrian re-identification method, the domain self-adaptation stage, the self-supervision clustering re-training stage and the joint loss learning stage are effectively integrated, and compared with a single training mode, the cross-domain pedestrian re-identification accuracy is further improved. The cross-domain pedestrian re-identification method has the advantages that the cross-domain self-adaptation stage, the self-supervision clustering re-training stage and the joint loss learning stage are effectively integrated.

The invention relates to a bearing fault diagnosis method based on a semi-supervised generative adversarial network, and the method comprises the following steps of obtaining vibration signals of thebearing in different states, and dividing the vibration signals into multiple samples; randomly dividing the samples into a training set and a test set; constructing a small number of label samples ofdifferent faults in the training set; constructing a one-dimensional semi-supervised generative adversarial network model; and inputting the training set into the adversarial network for training, wherein the trained adversarial network is used to test the diagnosis of centralized bearing faults. The method provided by the invention directly inputs the originally collected vibration signal and directly outputs the category of the bearing fault in the test set through training to achieve an end-to-end optimal diagnosis model, and uses a one-dimensional convolutional layer and a one-dimensionaldeconvolutional layer to enhance the ability of the one-dimensional semi-supervised generative adversarial network for extracting features. The invention is a semi-supervised training method, which does not require a large number of manual label samples, greatly saves time and labor costs, and has strong bearing fault diagnosis effect and anti-noise capability, and good stability.

The invention relates to a cross-modal joint hash retrieval method based on self-supervision. The method comprises the following steps: step 1, processing image modal data: carrying out feature extraction on the image modal data by adopting a deep convolutional neural network, carrying out Hash learning on the image data, and setting the number of nodes of the last full connection layer of the deep convolutional neural network as the length of a Hash code; step 2, processing the text modal data; using a word bag model for modeling text data, a two-layer full-connection neural network is established for feature extraction of text modal data, wherein the input of the neural network is a word vector represented by the word bag model, and the length of data of a first full-connection layer node is the same as that of data of a second full-connection layer node and a Hash code; step 3, for the neural network of category label processing, extracting semantic features from the label data by adopting a self-supervised training mode; and step 4, minimizing the distance between the features extracted from the image and the text network and the semantic features of the label network, so thatthe Hash model of the image and the text network can more fully learn the semantic features among different modals.

The invention relates to the technical field of text processing, in particular to a keyword extraction method, apparatus and device and a storage medium. The method includes: obtaining corpus text tobe extracted; inputting the to-be-extracted corpus text into a text labeling model for character type labeling processing to obtain a label corresponding to each character in the to-be-extracted corpus text, wherein the text labeling model is determined by performing supervised training based on a preset neural network model by using a training corpus text with a sample label, and the preset neural network model comprises a semantic representation model, a full connection layer connected with the semantic representation model, a conditional random field connected with the full connection layerand an output layer connected with the conditional random field; obtaining a character corresponding to a preset label in the corpus text to be extracted; and determining a keyword of the corpus textto be extracted according to the character corresponding to the preset label. According to the method, the accuracy and recall rate of keyword extraction can be improved.

The invention discloses a cross-modal medical image registration method. The method comprises the following steps: providing a training set comprising a floating image of a first modal and a referenceimage of a second modal; inputting the floating image into an image conversion network, and converting the floating image into a conversion image of a second modal; inputting the floating image and the reference image into a cross-modal flow sub-network to output a first deformation field; inputting the converted image and the reference image into the single-mode flow sub-network to output a second deformation field; inputting the first deformation field and the second deformation field into a deformation field fusion network to output a final deformation field; inputting the floating image and the final deformation field into a space transformation network to obtain a floating image after distortion transformation of the final deformation field; obtaining a first total loss function according to the transformed floating image and the reference image, and performing supervised training on the network by taking minimization of the first total loss function as a target; and inputting ato-be-registered image into the trained network to obtain a registered image. The cross-modal medical image registration method can greatly improve the cross-modal medical image registration effect.

The invention discloses a self-coding network and a training method thereof, and a method and a system for detecting abnormal power consumption, wherein the training method comprises the following steps: splicing the sample power data by a sliding window to obtain a training sample set, and marking the training sample set containing the surveyed user to obtain a labeled sample, wherein the unlabeled training sample is a non-labeled sample; carrying out unsupervised training on the self-coding network by unlabeled samples, and obtaining the initialization parameters of the self-coding network,then taking the discrete class tags obtained from the coding layer of the coding network as classifiers, carrying out supervised training on the classifiers by the labeled samples, and updating the parameters of the coding layer to obtain the trained self-coding network; then, utilizing the trained self-coding network to detect the power data of the user to be measured, and judging whether the user to be measured abnormally uses power. The invention can mine abnormal information in low-density electric power data, avoid noise data interference, and improve abnormal detection accuracy.

The invention discloses a micro-blog emotion prediction method based on weak supervised type multi-modal deep learning and relates to the field of multi-modal emotion analysis. The method comprises the following steps of preprocessing micro-blog multi-modal data; carrying out the weak supervised training of a multi-modal deep learning model; and carrying out the micro-blog emotion prediction of the multi-modal deep learning model. The method solves the problems of the multi-modal discriminant expression and the data label limitation existing in the emotion prediction of the micro-blog multi-channel content in the prior art, and realizes the final multi-modal emotion class prediction, wherein the accuracy is adopted as the experiment evaluation standard. The consistency degree between the predicted micro-blog emotion polarity category and the pre-marked emotion category is reflected. The performance of the method is greatly improved and the correlation among multiple modals is considered. As a result, an optimal effect is achieved in the aspect of the overall multi-modal performance. An ideal classification effect is achieved for different emotion categories. Through the weak supervised training, an initial model for text and image modals is obviously improved in the aspect of emotion classification effect.

The invention discloses a method for detecting a text in a natural scene with a robust shape. The method comprises the following steps: step 1, preprocessing a training picture in a text data set; step 2, establishing a PSENet progressive scale growth network, and finishing feature extraction, feature fusion and segmentation prediction on the training picture by utilizing the progressive scale growth network to obtain segmentation results of a plurality of prediction scales; step 3, carrying out supervised training on the PSENet progressive scale growth network established in the step 2 to obtain a detector model; step 4, detecting the to-be-detected picture; and step 5, obtaining a final detection result by using a scale growth algorithm.

The invention provides a fundus retinal blood vessel segmentation method and system based on K-Means clustering annotation and naive Bayesian model. The method of the invention comprises the followingsteps: randomly extracting color fundus images in a data set to construct a training set and a test set; extracting the gray image of G channel of the color fundus image as the object of feature extraction; carrying out feature extraction on the gray image, and representing each pixel in the image by a multi-dimensional feature vector; for each image in the training set after feature extraction,using K-Means clustering algorithm to label eigenvectors by clustering; training a naive Bayesian model based on training set data labelled based on K-Means clustering; segmenting the blood vessels ofeach image in the test set using the trained naive Bayesian model. The invention regards the result of clustering as the label with supervised training, and trains the naive Bayesian classification model for retinal blood vessel segmentation by using the label. The whole process does not require human to participate in the label, saves time and labor, and greatly improves the processing efficiency of the machine learning model.

The present invention discloses a training method of a convolutional neural network and a face identification method and device. The training method is characterized by determining a first error by adopting a first joint training supervision function composed of a cross-entropy loss function and a comparison loss function after the normalization and having two threshold values and according to thecharacteristic vectors of the face images in the samples of a source domain training sample set, and adjusting the network parameters of the convolutional neural network via the first error, whereinthe first threshold value is used to compare with the Euclidean distance of the characteristic vectors of the two face images in a positive sample pair, and the second threshold value is used to compare with the Euclidean distance of the characteristic vectors of the two face images in a negative sample pair, so that the supervised training of the negative sample pair can be controlled, and the supervised training of the positive sample pair also can be controlled, and the training efficiency and the accuracy of the CNN are improved, and accordingly, the generalization ability of the face identification method can be improved when the trained CNN is applied to the face identification method.

The invention discloses an intrusion detection method based on semi-supervised learning. The method comprises the steps of selecting an initial mixed sample set with samples with labels and unlabeled samples to be tested, calculating information gain of each characteristic value in a characteristic space, and completing characteristic selection based on information entropy; then, screening the samples with the labels based on the characteristic selection of the information entropy, using new screened training data for semi-supervised training of a classifier based on LapSVM, and utilizing the classifier after training is finished to classify the unlabeled samples to be tested; according to a detection index, determining the best evaluation value of the detection index, and outputting a classification result corresponding to the best evaluation value of the detection index. According to the intrusion detection method based on semi-supervised learning, the characteristic selection method is adopted to deal with redundancy phenomena easily occurring in network environment data, a semi-supervised learning model is established by utilizing a small number of samples with labels and a large amount of unlabeled data, the false alarm rate is reduced, and the detection rate is increased; meanwhile, the data redundancy can be reduced, and the detection efficiency is improved.

The invention discloses a transverse federated learning optimization method and device based on semi-supervision and a storage medium. The method comprises the steps that global model parameters updated by a global model this time and issued by a server side are received; after the first model is updated based on the global model parameters, self-supervised training on the first model is performedbased on a local label-free sample and an augmented sample of the label-free sample to obtain local model parameters; the local model parameters are sent to a server, so that the server performs supervised training on the second model according to the labeled samples and the local model parameters received from each client to obtain new global model parameters updated by the global model and issues the global model parameters to each client; and the process is repeated until a preset condition is met to stop training to obtain a target model. According to the invention, transverse federatedlearning can be carried out when only a small number of labeled samples exist at the server side and no label data exists at the client side, so that the method is suitable for a real scene lacking label data and the labor cost is saved.

The invention discloses a point cloud semantic segmentation method based on point global context relation reasoning. The method comprises the steps of obtaining a training set T and a test set V; constructing a point cloud data semantic segmentation network of deep learning and global context reasoning; using a multi-classification cross entropy loss function as a loss function of the point cloudsemantic segmentation network; performing P rounds of supervised training on the point cloud data semantic segmentation network by using the training set; and inputting the test set into the trained network model for semantic segmentation to obtain a segmentation result of each point. The method has the beneficial effect that the problem of insufficient global information extraction of 3D point cloud semantic segmentation is solved by utilizing a method based on deep learning and global context reasoning. On the basis of deep learning, an added global context inference module models the relationship between feature channels by using a channel attention mechanism, global information of the relationship between the channels is further transmitted and aggregated through graph convolution, andthe global information can be obtained so as to refine the result of point cloud semantic segmentation.

The invention discloses a transverse federated learning optimization method and device based on semi-supervision and a storage medium. The method comprises the steps of receiving global model parameters which are issued by a server and updated by a label-free global model this time; performing self-supervised training on a local model to be trained according to the global model parameters and thetraining samples, and updating encoder parameters and decoder parameters in the model to be trained to obtain local model parameters; sending the local model parameters to a server, so that the serverperforms supervised training on the to-be-trained model according to the local model parameters sent by each client to obtain global model parameters updated by the new label-free global model and issues the global model parameters to each client; and repeating until a preset condition is met to stop training to obtain a target model. According to the invention, transverse federated learning canbe carried out when the client only has a small amount of label data or even does not have label data completely, so that the method is suitable for a real scene lacking label data, and the labor costis saved.

When repair / replacement information is acquired by a repair / replacement information acquisition device, feature quantities of vehicle state information acquired by an vehicle state information acquisition device are detected, and a content of repair or parts replacement is estimated using the detected feature quantities of the vehicle state information, and supervised training data. If the result of the estimation accords with the repair / replacement information acquired by the repair / replacement information acquisition device, a storage device is controlled so that the combination of the detected feature quantities of the vehicle state information and the repair / replacement information acquired by the repair / replacement information acquisition device is added to the supervised training data.

The invention discloses a papillary thyroid cancer pathological image classification method based on deep learning, and mainly solves the problem of poor classification effect on papillary thyroid cancer pathological images in the prior art. According to the scheme, the method comprises the following steps: 1) reading a papillary thyroid cancer pathological section image with an amplification factor of 20, and inputting the papillary thyroid cancer pathological section image into an improved VGG-f convolutional neural network to obtain an attention heat map; 2) normalizing the attention diagram to obtain a discrimination force region position; reading a 40-time amplified thyroid cancer pathological image and obtaining an image block according to the position of the discrimination area; 3)inputting the image blocks into an original VGG-f network, constructing a loss function, and performing supervised training on the network; 4) extracting trained VGG-f network convolution features andperforming classification processing to obtain categories of the image blocks, and 5) judging the categories of the thyroid cancer pathological images according to the categories of the image blocks.The classification accuracy is high, and the method can be used for classifying the thyroid cancer papillary cancer pathological images by a computer.

The invention discloses a robot stirring and grabbing combination method based on deep reinforcement learning. The robot stirring and grabbing combination method comprises the following steps: building a robot stirring and grabbing execution platform in a real environment and a robot stirring and grabbing execution platform in a simulation environment respectively; modeling a stirring and grabbingprocess of the robot into a Markov process, and constructing a state space, an action space and a reward function; constructing a robot stirring and grabbing learning framework based on deep reinforcement learning and a deep reinforcement learning network; then, repeatedly carrying out stirring and grabbing action attempts on the simulation platform, collecting experience sample data, and carrying out self-supervised training on the deep reinforcement learning network according to the Markov process; and finally, migrating the trained network model to a real environment, and carrying out actual training on the robot in the real environment. When facing a random grabbing scene and a novel grabbing object, the robot can sense and analyze the scene and decide to execute a stirring or grabbing action, an action position and an action direction.

A method includes receiving input data including a plurality of feature vectors and labeling each feature vector based on a temporal proximity of the feature vector to occurrence of a fault. Feature vectors that are within a threshold temporal proximity to the occurrence of the fault are labeled with a first label value and other feature vectors are labeled with a second label value. The method includes determining, for each feature vector of a subset, a probability that the label associated with the feature vector is correct. The subset includes feature vectors having labels that indicate the first label value. The method includes reassigning labels of one or more feature vectors of the subset having a probability that fails to satisfy a probability threshold and, after reassigning the labels, training an aircraft fault prediction classifier using supervised training data including the plurality of feature vectors and the labels.

The invention provides a training method for a bimodal emotion recognition model and a bimodal emotion recognition method. The training method for the bimodal emotion recognition model comprises the following steps: inputting voice training data into a first neural network model for training, so that a voice emotion recognition model is obtained; inputting image training data into a second neuralnetwork model, and carrying out supervised training of the first stage by adopting a first loss function, so that an initial image emotion recognition model of the first stage is obtained; and inputting the image training data into the initial image emotion recognition model of the first stage, carrying out supervised training of the second stage by adopting a second loss function, so that a target image emotion recognition model is obtained, and carrying out decision fusion on the voice emotion recognition model and the target image emotion recognition model, so that the bimodal emotion recognition model is obtained.

The invention provides a 3D point cloud data semantic segmentation method based on deep learning and a self-attention mechanism, which is used for solving the technical problem of low segmentation precision in the prior art, and comprises the following implementation steps of: (1) obtaining a training set and a verification set; (2) constructing a 3D point cloud data semantic segmentation network of deep learning and a self-attention mechanism; (3) setting a loss function required by the 3D point cloud data semantic segmentation network for training deep learning and a self-attention mechanism; (4) carrying out supervised training on the 3D point cloud data semantic segmentation network of deep learning and a self-attention mechanism; and (5) obtaining a semantic segmentation result of the 3D point cloud data test set. According to the method, the self-attention module is added to the deep learning network, the deep features containing the relations between the feature channels can be better extracted, and therefore the segmentation precision is improved.

The invention discloses a CNN-DBN-based partial discharge fault diagnosis method, which comprises the following steps: constructing a deep belief learning network based on a deep convolution neural network and a restricted Boltzmann machine model; collecting partial discharge simulation data and actually-measured partial discharge data, mixing a part of the partial discharge simulation data and actually-measured partial discharge data to serve as a training sample set, and mixing the rest partial discharge simulation data and actually-measured partial discharge data to serve as a test sample set; carrying out unsupervised training on the deep belief learning network by utilizing the training sample set, and extracting cross-mode features; inputting the cross-mode features into a logical regression classifier, and carrying out supervised training on the regression classifier by utilizing the test sample set to obtain a trained deep belief learning network; and inputting partial discharge to be tested into the trained deep belief learning network to obtain a partial discharge fault diagnosis result. Accuracy of fault diagnosis is improved.

The invention discloses a method for identifying offline handwritten Chinese characters based on a deep separable convolutional neural network. According to the method, firstly, an image is subjectedto image cropping and image negative image preprocessing; then a convolutional neural network based on a depth separable convolution is designed, a random gradient descent method and a reverse propagation algorithm is used to carry out a supervised training, and model parameters are saved when a model is converged; finally, the saved model is used to identify a test image to verify a validity of the model. According to the method for identifying offline handwritten Chinese characters based on the deep separable convolutional neural network, the calculation amount of the model and the storage amount can be reduced, such that the offline handwritten Chinese characters identification model based on the neural network can be operated on the mobile terminal device offline; and the image preprocessing and convolutional neural network design are improved by mainly reducing the calculation complexity and the model capacity of the convolutional neural network, such that the method can be applied to computing equipment with limited computing resources and storage resources.

The invention discloses an artificial intelligence optimization system and method adopting adversarial training, and belongs to the field of machine learning. The method comprises the following steps:carrying out feature extraction on acquired data in an original data set, and training a neural network model which is not subjected to adversarial training to obtain an identification model; Takingthe identification model as a parameter for calculating the fitness, and generating confrontation samples by a genetic algorithm based on the calculated parameter and the extracted characteristics toobtain all the generated confrontation samples; Mixing all generated adversarial samples with the samples subjected to feature extraction, and performing supervised training of the identification model again as a new data set to obtain an identification model subjected to supervised training; Testing the supervised and trained recognition model, if the supervised and trained recognition model meets a given requirement, obtaining a finally trained recognition model, and if not, adjusting parameters of a genetic algorithm, and then generating a confrontation sample again for training. The methodis used for training the neural network model against resistance, and the model security is improved.

The invention discloses a semantic similarity calculation method based on deep learning, and relates to the field of semantic similarity calculation. The method comprises the following steps: step 1,constructing a training data set, and preprocessing training data to obtain one-hot sparse vectors; step 2, constructing a semantic similarity calculation network model comprising N layers of BI-LSTMnetworks, a residual network, a similarity matrix, a CNN convolutional neural network, a pooling layer and a full connection layer; step 3, inputting the one-hot sparse vector into the network model,and training parameters by using a training data set to complete supervised training; and step 4, inputting a text to be tested into the trained network model, judging whether the text to be tested isa similar text or not, and outputting a result. The semantic similarity calculation network model comprises a multi-layer BI-LSTM network, a residual network, a CNN convolutional neural network, a pooling layer and a full connection layer. Meanwhile, a BI-LSTM network and a CNN convolutional neural network are used, and a residual network is added into the BI-LSTM network, so that the problem ofgradient disappearance caused by a multi-layer network is solved, and the feature extraction capability of the model is enhanced.