65results about How to "Reduce training data" patented technology

The invention discloses a deep convolutional neural network-based human face occlusion detection method. The method comprises the steps of performing block segmentation on an input image to obtain a target pre-selected region; constructing a first deep convolutional neural network, training the first deep convolutional neural network comprising a first deep convolutional network and a first multilayer perceptron connected with the first deep convolutional neural network to obtain required parameters, extracting features of the target pre-selected region, and performing classification; predicting the position of a human head through a second multilayer perceptron according to the extracted features; filtering the credibility of a classification type which is the human head and the predicted position of the human head through non-maximum suppression to remove an overlapped duplicate detection box; and obtaining a human head block in combination with original image segmentation, constructing a multi-task learning policy-based second deep convolutional neural network, and judging whether the left eye, the right eye, the nose and the mouth of the human head block are occluded or not. According to the method, the occluded human face can be accurately detected and the specific occluded part of the human face can be judged; and the method is mainly used for crime pre-warning of videos of a camera in front of an automatic teller machine.

Insulator defect identification and positioning device and method based on capsule network, Deep learning model is established. Target detection based on capsule network structure is adopted. Parameters of convolution layer, capsule layer and decision layer of capsule network are initially configured. Existence and posture of insulators are detected by training set and patrol image training, and parameters of decision layer are predicted. The original machine patrols the image to be discriminated, outputs the specific position of the insulator through the capsule network, segments the image and extracts the image for standby, and calls the insulator feature extraction module to the segmented image. The extracted insulator characteristic string is divided into upper and lower parts, and theinsulator characteristic map obtained after the threshold value division is scanned line by line, the pixel points of each line are counted, the maximum connecting area where the insulator is locatedis determined, the main axis direction of the insulator is calculated, the insulator is rotated to the horizontal direction, and the abnormal position of the insulator is obtained. The invention greatly improves the efficiency of insulator defect identification, and provides reliable technical support for the power network maintenance department.

The invention discloses an image defogging method based on a multi-scale dense connection network. According to the method, images with different degrees of fog are reconstructed as relatively clear images, so that the quality and vision feeling of the images can be greatly improved. The invention discloses a self-adaptive histogram equalization mode for image preprocessing to increase contrast ratios of images, and a defogging effect can be remarkably improved; characteristics of different degrees of fog can be described by using a multi-scale dense connection convolutional nerve net, in addition, with effective combination of the characteristics, a most effective defogging effect can be achieved; an equation based on Retinex defogging problems is provided, so that end-to-end deep learning defogging can be carried out relatively concisely and effectively; compared with other defogging algorithms based on deep learning, the method has the effects that not only is the number of model parameters greatly reduced, but also an ideal defogging effect can be achieved under a condition of a very small amount of training data.

The invention discloses a wind power plant short-term wind power prediction method based on VMD-FCM-GRU, and belongs to the technical field of wind power prediction. The method comprises the followingsteps: firstly, obtaining a series of finite stable subsequence components based on a VMD normalized complex wind power time sequence, clustering the subsequence components by using an FCM algorithm,clustering the subsequences with similar fluctuation trends, and superposing the subsequences so as to shorten the training time; and training a GRU neural network for prediction by using the clustered sequences, and superposing and reversely normalizing prediction results of the sub-sequences to obtain a final prediction result. According to the method, in combination with the strong calculationefficiency of the GRU, the VMD-FCM-GRU combined model can realize relatively good prediction precision and relatively short training time, and the precision and speed of short-term wind power prediction can be remarkably improved.

The invention discloses an active learning and deep learning combined aluminum material surface defect detection method. The method comprises the steps of performing data enhancement on data in a training set by randomly adjusting image saturation, adjusting image brightness, adjusting image contrast and randomly rotating an image; adopting a Weighted-Entry evaluation standard for active learning,ranking the samples to be labeled according to a Weighted-Entry value in an increasing sequence, selecting K highest samples for labeling, and adding the samples into a training set to serve as training samples; and meanwhile, sequencing the samples to be labeled according to an increasing sequence of Entropy (information entropy) by adopting a pseudo labeling algorithm, selecting the lowest H samples, and taking a prediction result of the model as a pseudo label to serve as additional temporary training data of next training. An SEResNet-152 neural network structure is adopted, the neural network structure is based on a ResNet-152 network model, an SE module is added behind each Resinual module, and the SE module is used for calculating the weight proportion relation between channels ofa feature map.

The invention provides an image classifier generating method and device as well as an image classifying method and device. The method for generating an image classifier for form sorting equipment comprises the following steps: extracting a plurality of local features from each sample image in a plurality of sample images; dividing the sample image into a plurality of sub-areas; grouping local features which are extracted from the sample images and at least associated with the same sub-area into a same group; generating a search tree corresponding to the same sub-area according to the local features in the same group to serve as the image classifier.

The embodiment of the invention provides a training method and system for a language model, wherein the method specifically comprises the following steps: acquiring seed corpuses of all fields; screening big data corpuses according to a vector space model of seed corpuses of all fields to obtain seed screening corpuses in the corresponding fields; respectively training by utilizing the seed screening corpuses of all fields to obtain the screening model of the corresponding field; fusing the screening models of all the fields to obtain corresponding screening fusion model. According to the training method and system for language model disclosed by the embodiment of the invention, the parameter reasonableness of the language model can be improved on the premise of reducing a computation burden and saving time.

The invention discloses a hand exoskeleton rehabilitation training device and method based on surface electromyogram signals. The device comprises a mechanical glove, a surface electromyographic signal acquisition system and a driving control system, the surface electromyographic signal acquisition system comprises two electromyographic signal acquisition instruments, and each electromyographic signal acquisition instrument transmits an electromyographic signal to the drive control system through a Bluetooth module; the driving control system comprises an FPGA chip, a first rechargeable lithium battery, a Bluetooth module and a digital filter; the FPGA chip performs secondary filtering processing on the electromyographic signal after receiving the electromyographic signal of the electromyographic signal acquisition instrument; an active section signal is acquired by using a standard deviation threshold detection method; time domain features of the signals are extracted; the time domaincharacteristic parameters are input into the optimized BP neural network; gesture actions of various fingers are recognized through the BP neural network, the FPGA chip adds labels to the different gesture actions, surface electromyogram signals are converted into control signals, and then the control signals are sent to the mechanical glove through the Bluetooth module so as to drive the fingersto move.

The invention provides a robustness method used for voice recognition on mobile equipment during agricultural field data acquisition. The robustness method is based on a noiseproof method which combines mobile equipment voice recognition characteristic compensation with model compensation, a non-stationary noise environment of a typical application scene is acquired according to the agricultural field data, steady noise-resistant voice characteristic parameters are searched, characteristic extracted from noise-containing voice is processed, and deviation, caused by noise, between the noise-containing voice characteristic and pure voice characteristic is removed, thus the accuracy rate of system recognition is effectively improved. The robustness method provided by the invention has low complexity and is easier to implement; and meanwhile, training data required by a deviation mode which is based on an agricultural specific scene is less, and the instantaneity is good, thus the robustness method provided by the invention is more applicable to application under the condition that the calculation and storage resources of the mobile equipment are limited.

The invention discloses a wireless cellular network flow prediction method based on deep transfer learning and cross-domain data fusion, and belongs to the technical field of intelligent communication. Similarity among three services of short messages, telephones and the Internet and the similarity among different regions are analyzed, a plurality of cross-domain data sets is fused, and a space-time cross-domain neural network model is adopted to predict the wireless cellular traffic; a cross-service and region fusion transfer learning strategy based on a space-time cross-domain neural networkmodel (STC-N) is provided, and the prediction precision of a target domain is improved according to data characteristics of a source domain. The method can verify that the more comprehensive the considered data set is, the higher the prediction precision of the model is; in addition, the proposed transfer learning strategy can reduce the training data, calculation capability and generalization capability required for constructing the deep learning model.

The invention relates to a self-encoding neural network-based symbol diagram node classification method. The method includes the following steps that: S1, an adjacency matrix is constructed on the basis of a symbol diagram structure; S2, the representation vectors of symbol diagram nodes are learned on the basis of a self-encoding model, so that the representation vectors of two nodes connected bya positive edge are adjacent to each other, and the representation vectors of two nodes connected by a negative edge are far away from each other; S3, on the basis of the obtained node representationvectors, a small number of nodes with type tags are adopted to train a classifier; and S4, the trained classifier is utilized to predict the type of an unknown node, a node type is outputted, and therefore, the node can be classified according to the node type. Compared with the prior art, the method of the invention has the advantages of low computational complexity, fewer required training data, no need for retraining new nodes and the like.

The invention discloses a negative selection intrusion detection method based on the variable self-body radius, and the method is mainly used for solving the problem of poor detection effect because the self-body space formed by setting a fixed autologous radius can not be better covered in the traditional method. The negative selection intrusion detection method is implemented through the following steps: 1) pretreating a KDD (Knowledge Discovery In Database) 99data set; 2) selecting parts of normal data from the data set to serve as an self-body so as to form a self-body set; 3) randomly generating a foreign body, and setting the variable self-body radius for all self-bodies by utilizing a distance characteristic between the self-body and the generated foreign body; (4) training a detector set D; and (5) detecting the test data by use of the detector set D so as to judge whether the test data is normal or abnormal. The negative selection intrusion detection method based on the variable self-body radius has the advantages of high positive detection rate and low misinformation rate, the effect of the negative selection intrusion detection method can be effectively improved under the condition that the self-body data amount is small, the negative selection intrusion detection method is used for identifying the abnormal network data, and ensuring the network safety.

The invention relates to an identity authentication method and device for wearable intelligent equipment. The method and the device have the beneficial effects that on the basis of the principle of intra-body communication and in combination with the bioidentification technology, when users use the wearable intelligent equipment, S21 parameters after test signals flow through user bodies are acquired in a test frequency band to determine the optimal frequency band, the S21 parameters are acquired and processed in the optimal frequency band to obtain final authentication templates used for representing user identifies; authentication of the user identifies is achieved by comparing acquired user identity authentication data with the authentication templates; the privacy information of the users is effectively protected by utilizing nonreproducible uniqueness of the biological characteristics of human bodies and the advantages that the biological characteristics of human bodies can not be lost or forgotten and are difficult to counterfeit, thus well ensuring the privacy and safety of the users, avoiding potential risks caused when the users use the wearable intelligent equipment and being greatly beneficial to popularization and application of the wearable intelligent equipment.

The invention discloses a natural language task processing method and device combining a regular expression and a neural network model. The method comprises the following steps of: 1) giving a task related to a natural language, and acquiring domain knowledge in a regular expression form for a specific domain related to the task to obtain a regular expression set; 2) selecting a basic neural network model capable of solving the task for the task; and 3) introducing domain knowledge contained in the regular expression set in the training process of the neural network model, thereby improving the performance of the neural network model for the task. The combination of the regular expression and the neural network model can be carried out on different levels, including an input layer, a module layer and an output layer. According to the method, the specific module of the neural network model can be finely adjusted based on the domain knowledge in the regular expression form, and the training of the neural network is assisted, so that the performance of the neural network model on the task is improved.

The invention discloses a method for intrusion detection base on negative selection and information gain, which mainly aims to solve the problems of poor effects and instability of the conventional negative selection algorithm (NSA) method. The method is implemented by the following steps of: (1) reading a self-mode set in; (2) performing conversion, normalization and discretization processing on data in the read-in self-mode set; (3) calculating the information gain of characteristics of 41 dimensions of the processed data; (4) selecting the characteristics of former N dimensions with higher information gain values according to calculated information gain value results, forming the self-mode set by using the data in the characteristics of the N dimensions, and training the formed self-mode set to generate a detector set; and (5) detecting the data in a test set by utilizing the generated detector set. The method has the advantages of high accurate detection rate, low rate of false alarm and capability of achieving satisfactory detection effects even with relatively less training data, and can be used for timely identifying abnormal network data and ensuring network security.

The invention belongs to the field of power battery management systems, and particularly relates to a battery SOC dual-state switching estimation method based on a BAS optimized ElmanNN-AH method. Thebattery SOC dual-state switching estimation method adopts a BAS algorithm for optimizing weight in ElmanNN, and switches different estimation algorithms for different battery states according to a switching method of the ElmanNN and a correction ampere-hour method. The battery SOC dual-state switching estimation method eliminates the defects that the initial SOC is difficult to measure and errorsare accumulated when an ampere-hour integral method is adopted, reduces the training data and training times of a neural network and reduces the workload.

The invention relates to a data processing technology, and provides an image structured data extraction method, an electronic device and a storage medium. The method comprises the following steps: inputting an image of to-be-extracted structural data into a position detection model; obtaining coordinates of all to-be-identified areas in the image; cutting the to-be-identified area and then executing transformation processing; inputting the to-be-identified areas before and after transformation into an identification model; obtaining an initial recognition result, screening a target recognitionresult of each to-be-recognized area from the initial recognition result, selecting a category result according to the similarity between the character of each target recognition result and each category of characters in the word bank, and generating a structured data file according to each category result and the target recognition result. According to the invention, the structured data in the image recognition result can be accurately extracted. Moreover, the invention also relates to an image recognition technology and a block chain technology in artificial intelligence, and the method canbe applied to the fields of smart government affairs, smart education and the like, thereby promoting the construction of smart cities.

The invention discloses a three-dimensional graph recognition method, device and apparatus and a storage medium. The method comprises: obtaining point cloud of a target three-dimensional model, preprocessing the point cloud of the target three-dimensional model to obtain a target two-dimensional array and a mapping relationship between points in the target three-dimensional model and two-dimensional pixels, obtaining a target two-dimensional image corresponding to the target three-dimensional model based on the target two-dimensional array and the feature data of each point in the point cloudof the target three-dimensional model, performing image recognition on the target two-dimensional image based on a two-dimensional image recognition network, and mapping the target two-dimensional image recognition result to a target three-dimensional model to obtain a target three-dimensional graph recognition result. By utilizing the technical scheme provided by the invention, a three-dimensional graph processing problem can be converted into a two-dimensional image processing problem. By utilizing the characteristics of more mature two-dimensional image recognition, high efficiency and accurate recognition, a more reliable three-dimensional graph recognition result is obtained, less training data is needed, and the cost and time are saved.

The invention discloses a high-resolution SAR image classification method based on a depth ladder network, mainly to solve the problem that the high-resolution SAR image has few data with class identifiers and a network can not be trained effectively. The method comprises steps: a to-be-classified high-resolution SAR image and identifier information thereof are inputted; a training data set D1 and a test data set D2 are constructed; features of the data sets D1 and D2 are normalized to obtain data sets D3 and D4; a classifier model based on the depth ladder network is constructed; the training data set D3 is used for training the network; and the well-trained classifier model is used for classifying the test data set D4. The few training samples with class identifiers can be made full use of, and the high classification precision can be acquired.

The invention relates to an implicit triad neural network and an optical fiber nonlinear damage equalization method, and belongs to the technical field of optical fiber communication and equalization.The neural network comprises an implicit triad neural network and an optical fiber nonlinear damage equalization method based on the implicit triad neural network. The method comprises the followingsteps: 1) generating a training set, a verification set and a test set, specifically, generating a binary bit stream, generating a label symbol stream, obtaining a symbol stream to be processed, generating a sample and dividing a data set; 2) optimizing an implicit triad neural network to obtain an optimal implicit triad neural network, initializing a hyper-parameter search process, initializing an adjustable parameter iteration process, calculating a loss function and gradients of all adjustable parameters, updating the adjustable parameters, iterating and evaluating an optimization result, and selecting an optimal implicit triple neural network; 3) testing the implicit triple neural network to obtain an equalized signal. Compared with the prior art, the neural network and the method arelower in calculation cost, and the equalization effect can be further improved.

The invention discloses a small hydropower station group generating capacity prediction method based on big data driving, and the method comprises the steps: dividing small hydropower station groups according to regions for the current situation that the small hydropower station groups are wide in point range and are managed disorderly; obtaining historical meteorological data and generating capacity data of each region, preprocessing bad data by adopting a data mining technology, and then performing normalization processing; and training the historical meteorological data and the power generation data by using two layers of long-term and short-term memory neural networks and one layer of full-connection neural network, and performing optimization calculation by taking the minimum root-mean-square error of the power generation true value of the test data and the power generation predicted value of the training model as an optimization target to finally obtain the power generation prediction model of the small hydropower station group. According to the method, the generating capacity of the small hydropower station group can be accurately predicted, comprehensive and ordered management of the small hydropower station is realized, and safe and stable operation of a power grid is ensured.

The invention provides a voice processing model training method and device, equipment and a computer readable storage medium. The method comprises the steps: obtaining sample data comprising voice of a source language and a target language sample text corresponding to the voice of the source language; inputting the voice of the source language into a voice recognition sub-model of a voice processing model to obtain a source language text; inputting the source language text into the word database of the voice processing model for traversal to obtain a word vector corresponding to the source language text; inputting the word vector and the target language sample text into the machine translation sub-model of the voice processing model to obtain a target language translated text; based on a preset loss function, calculating a loss value of the voice processing model according to the target language translation text and the target language sample text; and performing parameter adjustment on the voice processing model according to the loss value to obtain a trained voice processing model. According to the invention, the training data of the model can be reduced, and the training efficiency is improved. The invention further relates to a block chain technology.

The invention discloses a small hydropower station generating capacity prediction system based on a cloud edge fusion architecture and an implementation method thereof, and the method comprises the steps: collecting the real-time meteorological data and small hydropower station generating data of a measured region from an external system, directly uploading the real-time meteorological data and small hydropower station generating data to a cloud computing server through a VPN channel to achieve the human-computer interface of the system through the Internet; enabling the cloud computing serverto transmit the meteorological data and the small hydropower station power generation data to an edge server for storage. A small hydropower station generating capacity prediction model is constructed, small hydropower station generating capacity prediction is carried out by utilizing the model, then a prediction result is returned to a cloud computing server, system core analysis functions of data analysis, mining, prediction and the like are completed, and finally the prediction and analysis results are transmitted to terminal equipment through the Internet to be displayed. The method is atypical case of organic combination application of cloud computing and edge computing technologies, and effectively solves the problems of storage, transmission and display of power grid sensitive data while improving the reliability of the system.

The invention relates to a face shielding detection algorithm, which comprises the following steps of: face image preprocessing: cutting a plurality of collected face images according to face positioninformation and face key point coordinates to obtain a plurality of face five-sense-organ images as a basic data set; data augmentation: performing data augmentation on the basic data set; data annotation and division: carrying out classification annotation on different types of augmented images, and taking the annotated data as a training sample which comprises a training set, a verification setand a test set; building a convolutional neural network model; setting hyper-parameters, training the convolutional neural network model, determining the hyper-parameters, configuring the training set to simulate training, testing the model by the verification set, observing whether the accuracy of the model obviously floats or not, if the accuracy obviously floats, returning to setting the hyper-parameters, and otherwise, outputting the hyper-parameters as a face shielding detection model; and testing the model, verifying the accuracy, determining the accuracy of the model by using the testset, and evaluating the generalization ability of the optimal hyper-parameter training model.

The invention discloses a 3D point cloud semantic segmentation and migration method based on meta-learning, and relates to the technical field of robot navigation. The method comprises the following steps: constructing a Point Net network model; selecting a training data set; for each training data set, forming a training task set by utilizing different types of data; constructing a meta-learningframework; according to the meta-learning framework, training the Point Net network model through each training task set; selecting a test task set; and inputting the test task set into the trained Point Net network model for testing until the gradient update value of the model converges. According to the method, a trained model is loaded in a new environment task, optimal similar task parametersare used, and the training efficiency of the indoor scene semantic segmentation method of the new task is high; the semantic segmentation capability of different tasks is learned through a meta-learning framework, it is ensured that learning features can be suitable for different migration environments, and the generalization performance of the model is improved.

The invention discloses a multi-feature fusion sign language recognition method based on adaptive hidden markov, which comprises the following steps: 1. Firstly, extracting multiple features from a sign language video database and performing front-end fusion, namely constructing a feature pool set; then, an adaptive hidden Markov model of sign language video under different features in the featurepool set is constructed, and a feature selection strategy to obtain the appropriate back-end score fusion features. After selecting the back-end scoring fusion features, the scoring vectors under each back-end scoring fusion feature are calculated, and different weights are assigned to them, and then the back-end scoring fusion is carried out, so as to obtain the optimal fusion effect. The invention can realize the accurate recognition of sign language categories in sign language video and improve the robustness of the recognition.

The invention discloses an anomaly detection method for an intelligent power system, and effectively improves the accuracy of anomaly detection and classification of the power system. When training data is insufficient, compared with a traditional supervised classification method, the method apparently improves the detection effect and the classification precision; when the training data is sufficient, the method employs an improved multi-granularity cascade forest for replacing traditional deep learning, so the problems that a deep learning algorithm model is complex and the training time istoo long are solved. Compared with a traditional detection classifier, the method has the advantages of less required training data, lower algorithm complexity, shorter training time, faster convergence speed and faster response time.

Various embodiments set forth systems and techniques for training a personalized prediction model. The techniques include generating, based on interaction data associated with one or more users and a first weight associated with the interaction data, a first set of training data; generating, based on the personalized prediction model, a predicted enjoyment signal associated with playback of a digital content item; generating, based on the first set of training data and the predicted enjoyment signal, a second set of training data; and updating one or more parameters of a personalized ranking model based on the second set of training data.

The invention relates to an annular RGV trolley scheduling method based on SVR model prediction, and the method comprises the steps that RGV trolleys in an annular RGV trolley system start to go to all stations for work, and the specific work and position information of the trolley is obtained through an RIFD and an encoder, which are disposed on the RGV trolley; each station sends an applicationsignal request to call the RGV trolley to reach a corresponding position of the station to work, a pre-established SVR prediction model starts to record an application sending time point of each station after sending an application from each station for the first time, self training is completed, and prediction of an application signal generation moment of the station is realized; and reasonable scheduling is realized according to the predicted occurrence time of the application signal of each station and the position information and the working information of each current RGV. According to the invention, the uncertainty of station application signal generation is reduced, and the reasonable scheduling of RGV trolleys is feasible. And meanwhile, the trolleys on the annular rail can be reasonably dispatched, and the production efficiency of an assembly line is improved.