88results about How to "Has generalization ability" patented technology

The invention relates to a face image gender recognition method based on stack type sparse self-coding, and belongs to the field of image recognition, machine learning, and computer vision. A training process of the method includes image graying, histogram equalization, geometric correction, image normalization, the training of a sparse self-coding model, logic regression classifier training, a fine tuning model, and model fusion of face standard databases FERET and CAS-PEAL-R1, and a prediction process comprises the capturing of natural scene images by a camera, image graying, histogram equalization, face detection, geometric correction, image normalization, the prediction by employing a stack type sparse self-coding model, and result marking. According to the method, the problem of face gender recognition is solved by employing the stack type sparse self-coding model, combination characteristics of the images can be learned layer by layer, original signals can be better represented in an abstract manner, characteristics extracted by a hiding unit are further adjusted by the adoption of fine tuning, and the recognition accuracy is higher.

The invention relates to a remote sensing image ground object labeling method based on an attention mechanism convolution neural network, which comprises the following four steps: a computer reads data, constructs convolution neural network of attention mechanism, trains network model, and tests network to obtain labeling result. By adding an attention mechanism module, the invention enables the network to pertinently extract the information of the key position, makes up for the deficiency of the lack of the spatial information at the network end, and improves the classification effect of thenetwork to the ground object details. By using the mechanism of in-depth monitoring and using the characteristics extracted from the middle of the network to supervise the classification, the trainingspeed of the network can be further increased and the comprehensive performance of the network can be improved; through the up-sampling module of deconvolution, the resolution of feature extraction is increased and the method can overcome the problem that small objects are difficult detect to a certain extent, and can automatically classify remote sensing image pixels into corresponding object categories, reduce the trouble of manual interpretation, greatly accelerate the interpretation process, and obtain refined labeling results.

The invention discloses a multi-agent federated cooperation method based on deep reinforcement learning. The method comprises the following steps: S1, establishing a deep reinforcement learning modelfor each agent; S2, establishing a corresponding neural network for the intelligent agent; S3, interacting the intelligent agent with the environment, storing the decision experience in an experiencepool, and updating a local neural network model according to a stochastic gradient descent method; S4, transmitting local neural network model parameters to a cooperation platform; S5, aggregating theparameters uploaded by the intelligent agents, and returning a result to each intelligent agent to update the parameters; S6, performing soft update by the intelligent agent to obtain latest local model parameters; and S7, repeating step S3 to S6 until the target task is completed. According to the intelligent agent, while environment exploration and decision making are carried out through deep reinforcement learning, learning experience of other intelligent agents is obtained through the federated learning technology, so that the learning efficiency of the intelligent agents is effectively improved, and the cooperation overhead between the intelligent agents is reduced.

The invention relates to an automatic sleep staging method based on multiple electroencephalogram and electromyography characteristics. The method comprises the following steps: collecting an electroencephalogram signal and an electromyography signal; utilizing wavelet decomposition to remove high-frequency noises from the electroencephalogram signal and the electromyography signal; extracting an energy ratio of alpha, beta, theta and delta characteristic waves of the electroencephalogram signal after removing the noise, thereby acquiring a first characteristic parameter; utilizing a sample entropy method to extract a sample entropy of the electroencephalogram signal, thereby acquiring a second characteristic parameter; utilizing a wavelet decomposition algorithm to extract a high-frequency characteristic energy ratio in the electromyography signal, thereby acquiring a third characteristic parameter; and inputting the first characteristic parameter, the second characteristic parameter and the third characteristic parameter to a support vector machine and performing training and testing, thereby acquiring a classifying result. According to the invention, the method for extracting multiple EEG and EMG characteristics is adopted and a support vector machine classifier is combined, so that the accuracy of the sleep staging is promoted; a cross validation result proves that the method has certain generalization ability; an experimental result is high in reliability; and the application prospect is excellent.

The invention discloses an expression synthesis method and device based on phoneme driving and a computer storage medium, and the method mainly comprises the steps: recognizing a target voice text according to a preset database, so as to obtain a phoneme sequence, and converting the phoneme sequence into a replacement expression parameter sequence; extracting to-be-replaced original sub-video datafrom the original video data based on the voice duration of the target voice text; constructing a three-dimensional face model based on faces in the original sub-video data, extracting to-be-replacedexpression parameters of the three-dimensional face model frame by frame to generate a to-be-replaced expression parameter sequence, and replacing the to-be-replaced expression parameter sequence with the replaced expression parameter sequence; utilizing the replacement expression parameter sequence to drive a three-dimensional face model to generate a target two-dimensional image sequence, and rendering the target two-dimensional image sequence frame by frame; and splicing the rendered target two-dimensional image sequence to generate target sub-video data for replacing the original sub-video data. According to the invention, the expression synthesis video with a more real effect can be efficiently and accurately obtained.

A method and apparatus for home identification of address are provided in an embodiment of that present application, the method comprises splitting a receiving address of a logistics task into a plurality of sub-addresses corresponding to different geographical levels, by using a match data set, matching based on that plurality of sub-addresses, determining a target area to which the recipient address belongs in the divided plurality of areas, so that the multi-classifier based on a large number of samples is used to match the address of the receiver. The method can not only solve the problemthat the address adopts abbreviation or nickname to lead to the recognition failure, but also can learn the corresponding weights according to the wrong words and the wrong part of the address of theuser. Compared with the name matching of the campus, the method has more generalization ability and stronger robustness, and improves the accuracy of the home identification of the received address.

The invention provides a user identity recognition method and system combined with user gait information, and the method comprises the steps: carrying out the posture detection of each frame of pedestrian object in a video sequence of an original data set through a two-dimensional posture estimation system, and extracting the posture information; preprocessing the extracted joint coordinate sequence to generate a human skeleton data set; and finally, constructing a space-time diagram convolutional network model, dividing the skeleton diagram into six sub-diagrams, sharing joints among the sixsub-diagrams, learning an identification model by using the diagram convolutional network, training by using the constructed data set, and optimizing network parameters by using a multi-loss strategycombining classification loss and comparison loss and random gradient descent. And predicting the accuracy of the trained model by using the verification set. According to the method, effective information of the joint points is fully utilized, the motion state in the time dimension is reserved as much as possible, high robustness is achieved for clothes changes and carrying states, and good generalization capacity is achieved on a cross-view task.

The invention discloses a rock image pore type identification method based on semantic segmentation. The method comprises the following steps: S1, establishing an initial deep learning network model;S2, acquiring an original rock image, and performing image cutting and image enhancement to obtain preprocessed image data; S3, acquiring an original image of the rock, and performing manual marking of pore positions and shapes to obtain marked label image data; S4, carrying out One-Hot encoding on the labeled label image data to obtain coded label data; S5, training the initial deep learning network model by taking the preprocessed image data and the encoded tag data as training samples to obtain a trained model; and S6, identifying the to-be-identified image by using the trained model. The method has the advantages that the noise resistance is high, the generalization capability is realized, the pore recognition precision can be improved, and the pore category recognition is realized.

The invention relates to a fish body posture and length automatic analysis method based on key point detection and a deep convolutional neural network, and the method comprises the steps: S1, obtaining binocular images comprising a fish school through an underwater binocular camera, wherein the binocular images comprise a left image and a right image; and S2, performing calibrating in an underwater environment to obtain binocular camera parameters, and performing binocular correction on the obtained binocular image. The beneficial effects of the invention are that the method combines the deepconvolution neural network, and is high in adaptability to an application environment and a scene; a key point detection idea is introduced, and only the spatial positions of specific key points on fish bodies are concerned, so that the difficulty of global binocular matching in underwater application is avoided; required equipment is simple, and only an underwater binocular camera and an operation rear end are required; attitude estimation and length measurement can be carried out on multiple fishes with different positions and attitudes in the image in real time, and the accuracy and the efficiency are relatively high; and the model also has generalization ability for tasks, and is easy to migrate from one working scene to another.

The invention relates to a vortex structure recognition method based on Support Vector Machine (SVM), in aspect of feature extraction, statistics of coefficient matrix from wavelet transform of refractive index filed data is extracted, compared with traditional method of utilizing coefficient matrix from wavelet transform directly, the method winkles mass of redundant information, and decreases calculating amount. In aspect of classifier design, classification method based on support vector machine is brought forward, minimum structural risk principle is utilized to ensure that the vortex structure is correctly detached by classification surface and the classification interval is made maximum, in theory the invention decreases mis-recognition rate compared with the traditional classification method based on minimum experiential risk principle. The invention can represent and distinguish turbulence vortex structure more accurately; the invention establishes bases of optics effect accuracy for airborne optical equipment of missiles and airplanes, etc.

The invention discloses an optimal path training method for an unmanned aerial vehicle to avoid a columnar obstacle to reach a target point. The method comprises the following steps of (1) constructing an environment simulator based on aerodynamics, (2) initializing an unmanned aerial vehicle strategy model by using a deep neural network, (3) enabling the unmanned aerial vehicle to obtain observation at the current moment in the environment simulator and making actions, and (4) enabling the unmanned aerial vehicle to interact with the environment simulator, generating and storing training data, sampling the training data, learning the unmanned aerial vehicle to avoid columnar obstacle through a reinforcement learning algorithm by using the training data, and selecting an optimal path to reach a target point. According to the method, the unmanned aerial vehicle can obtain learning ability similar to human beings, and the task is completed in tasks which are high in difficulty coefficient, unknown and complex in environment and have uncertain factors.

The invention discloses a continuous zooming target recognition system and method with an adjustable field of view, and belongs to the technical field of target recognition. The continuous zooming target recognition system with the adjustable field of view comprises a continuous zooming first-order subsystem with an adjustable field of view, an image acquisition first-order subsystem and a targetrecognition first-order subsystem, the continuous zooming first-order subsystem with the adjustable field of view is used for accurate focusing and capturing of a moving target in a variable field ofview, the image acquisition first-order subsystem is used for acquiring infrared and visible lights focused on the target and generating a target image database, and the target recognition first-ordersubsystem employs the target image database to train a deep neural network and realizes accurate recognition of acquired target images. The invention also discloses a continuous zooming target recognition method with an adjustable field of view. According to the system and the method, in the condition of target moving, clear acquisition of the moving target can be realized, and the acquired target can be accurately recognized.

The invention belongs to the field of intelligent cell biological recognition, and relates to an RNA binding protein recognition method based on multi-view depth features and multi-label learning. Themethod comprises a training stage and a use stage, wherein the training stage comprises initial multi-view data construction, a depth multi-view feature extraction model and multi-label classifier training, wherein the initial multi-view data construction comprises the following steps: converting an original RNA sequence into an amino acid sequence and a dipeptide component by using a molecular biology principle and a statistics principle to obtain the characteristics of the amino acid sequence and the dipeptide component, and then constructing an initial multi-view characteristic together with the original RNA sequence, and constructing a model for the initial multi-view characteristic. According to the method, based on the initial multi-view data, the CNN is used for deep learning to construct a deep multi-view characteristic, and compared with the original multi-view characteristic, the multi-view characteristic extracted based on the deep features have smaller data dimensions anda higher classification effect.

The invention discloses a rockburst grade prediction method and system based on principal component analysis and a BP neural network. The rockburst grade prediction method comprises the steps: gradingrockburst according to rockburst strength degrees; determining all influence factor indexes of rockburst grading in the high ground stress area; obtaining index variables in actual engineering and corresponding actual rockburst grade data, and performing mean variance normalization on index variable values; carrying out principal component analysis on the excavated index variables by adopting a principal component analysis method to obtain a plurality of principal component variables, and enabling the principal component variables to correspond to the rockburst grade determined according to the rockburst strength degree; taking the plurality of obtained index variables as input indexes, taking the corresponding rockburst levels as output values, carrying out training learning on the databy adopting a BP neural network algorithm, and establishing a mathematical model of each index-rockburst level; and obtaining an index variable value near an unexcavated tunnel face, carrying out principal component analysis based on the average value, the standard deviation and the like of the training data, extracting a corresponding principal component variable, obtaining a principal component,and then carrying out rockburst grade prediction by using the obtained mathematical model.

The invention relates to a ground source heat pump system performance prediction method. The method comprises the steps of (1) preprocessing and carrying out seasonal decomposition on the heat transfer performance data of a GSHP system; (2) constructing a data structure with GSHP system working parameter input in a seasonal month-by-month manner; (3) constructing a decision tree set model which has external system working parameter input and can be used for time sequence regression prediction, namely a prediction model; (4) carrying out data reconstruction on test set data by adopting the samemethod as the step (2), and carrying out performance prediction by adopting an autoregressive mode and the decision tree set model obtained by training in the step (3); and (5) comparing the prediction data obtained by the test set through the decision tree set model with the real data, and measuring the prediction effect of the decision tree set model. The prediction method is simple and is easyto operate. The method also solves the problem that the traditional time sequence analysis method cannot predict the time sequence with out-of-band partial parameters.

The invention relates to the technical field of multi-modal industrial process fault detection, and discloses a multi-modal industrial process fault detection method of a weighted k-nearest neighbor standardized method. The detection method comprises a modelling stage and a detection stage; the modelling stage comprises the following steps: collecting normal data of different modals in the process, and serving the normal data as the integrity to form a training set X belonging to Rnxm. Through the multi-modal industrial process fault detection method of the weighted k-nearest neighbor standardized method, the information from the same modal can be intensified and the different modal information can be weakened in the application process of the WKNS method through the importing of a weightof the distance; and meanwhile, an operation of determining the nearest neighbor parameter k value according to the experience is avoided in the computation process, and the modal effect between different modals and stages is effectively eliminated; and the single-modal independent modelling and the model attributive division for the new testing sample are avoided by combining the WKNS-PCA method;and the traditional single-modal fault detection method is applied, a certain generalization capacity is provided, and the fault detection precision is improved.

The invention provides a near-field source positioning method based on deep learning. According to the technical scheme, the method comprises the following steps: firstly, obtaining a near-field source phase difference matrix by utilizing a radar antenna array; then, inputting the information of the near-field source phase difference matrix into an automatic encoder, and calculating the direction of arrival of the near-field source; secondly, the output of the automatic encoder is input into a first-class convolutional neural network, and the direction of arrival of a near-field source is calculated; and finally, removing direction-of-arrival parameters contained in the near-field source phase difference matrix information by utilizing output information of the first type of convolutional neural network, inputting the direction-of-arrival parameters into the second type of convolutional neural network, and calculating the distance of the near-field source. The direction of arrival and distance parameters of multiple near-field sources can be effectively separated and calculated, the positioning precision is high, and the generalization ability is high.

The invention discloses a pushing and grabbing collaborative sorting network based on double viewing angles and a sorting method and system thereof. The trained pushing and grabbing collaborative sorting network comprises a pushing full convolutional network and a grabbing full convolutional network, and the network is applied to robot pushing and grabbing collaborative sorting. The sorting methodcomprises the following steps of correspondingly acquiring point cloud graphs of an object scene to be sorted from two viewing angles, rotating a top view of the point cloud graphs, correspondingly inputting a plurality of rotating images into the pushing full convolutional network and the grabbing full convolutional network to obtain two thermodynamic graphs with Q values output by the networks,and selecting the thermodynamic diagram with the larger Q value as a final thermodynamic diagram; and according to the pixel point corresponding to the maximum Q value in the thermodynamic diagram and the rotation angle of the rotation image corresponding to the thermodynamic diagram, controlling the robot to execute the sorting action of the network corresponding to the thermodynamic diagram, and then completing sorting. According to the sorting method, double viewing angles are combined with deep Q learning, so that the grabbing success rate is high and the generalization ability is high inthe face of a disordered stacking scene.

The invention provides a friction noise prediction method based on a BP (Back Propagation) neural network, which is used for predicting the frequency of friction noise generated in the braking processof a brake on a vehicle and comprises the following steps: step 1, acquiring the friction noise and a data sample p1i associated with the friction noise; Step 2, performing normalization processing on the data samples p1i to obtain corresponding normalized data samples p2i respectively, and dividing the data samples p2i into two groups, namely a first normalized data sample and a second normalized data sample; Step 3, establishing a friction noise prediction model by using a function newff specially establishing a BP neural network in MATLAB; 4, inputting the first normalized data sample intoa friction noise prediction model to obtain an updated network weight and an updated network threshold; And 5, inputting the second normalized data sample into the friction noise prediction model until the training error is 0.01, and determining a final network weight and a final network threshold to obtain a qualified friction noise prediction model.

The invention relates to the field of artificial intelligence, discloses an intention classification method, device and equipment based on voting decision, and a storage medium, and can improve the convenience of intention classification. The method comprises the steps: carrying out keyword extraction on a to-be-processed text, and obtaining a target keyword, wherein the to-be-processed text is used for indicating input consultation information; taking the target keyword as an index to retrieve a preset historical text to obtain a candidate historical text corresponding to the target keyword,the historical text being used for indicating consultation information stored in a database before receiving the to-be-processed text; calculating scores of the candidate historical texts through a template semantic similarity algorithm and a syntax similarity algorithm based on deep learning; carrying out sorting and voting decisions on the candidate historical texts according to the scores, obtaining target historical texts, obtaining pre-marked intention categories in the target historical texts, and taking the intention categories as target intention categories of the to-be-processed texts.

The invention discloses a dynamic service function chain arrangement method and system based on deep reinforcement learning. The method comprises the following steps: acquiring a historical network state according to an SDN controller; the network state comprises service function chain request stream information generated in the Internet of Things supporting mobile edge computing and corresponding network resource state information; setting deep reinforcement learning parameters and initializing the weight of the neural network; training a neural network according to an experience sample generated by interaction of the intelligent agent and the environment; for the service function chain request flow obtained in real time, the trained neural network is utilized, a heuristic algorithm is adopted, the placement and routing path of the virtualized network function meeting the requirement of the service function chain request flow is determined and deployed, and network resource state information is comprehensively considered; the load balancing of the network is realized while the resource consumption cost and the time delay of the request stream of the Internet of Things are reduced, and the network flow receiving rate is improved.

The invention relates to a graph query method and device, electronic equipment and a computer readable storage medium, and belongs to the technical field of graph query, the method comprises the stepsof performing graph embedding mapping on a first graph to a vector space to obtain a first vector space; performing graph embedding mapping on the second graph to the vector space to obtain a secondvector space; calculating and clustering the second vector space to obtain a node clustering cluster of the second graph; based on the first vector space and the node clustering cluster, calculating acandidate node set similar to the node of the second graph in the nodes of the first graph; calculating the maximum connected subgraph of each node in the candidate node set to obtain a plurality ofconnected subgraphs; and obtaining graphs with similar topological structures of the second graph in the first graph from the plurality of connected sub-graphs, and taking the graphs as query resultsof the second graph in the first graph. According to the embodiment of the invention, the graph query efficiency is effectively improved, and the generalization ability of the query result is improved.

The invention discloses a pavement long-term performance prediction model based on deep learning and a construction method thereof. The model is an LSTM-BPNN feature fusion model for fusing an LSTM long-term and short-term memory network and a BPNN back propagation neural network through an attention method. The LSTM-BPNN neural network long-term performance prediction model is constructed by analyzing the prediction model of domestic and overseas scholars in road performance based on deep learning, dividing data features into basic features and time sequence features according to data features of a road surface long-term performance database and combining different feature data, so that the LSTM-BPNN neural network long-term performance prediction model can be used for predicting the road performance, gives play to the advantages of the network, performs respective functions, achieves the correlation coupling, and fully captures the potential correlation between different types of climatic changes and road performance change trends, thereby achieving the more accurate prediction of the long-term performance of a pavement in the future.

The invention discloses an A3C-SRU-based intelligent vehicle traffic flow converging method and system. The implementation method comprises the following steps of 1, adopting environmental parametersand vehicle parameters by devices such as a digital camera, a multi-line laser radar, a millimeter-wave radar and a gps positioning system; 2, establishing a simulation environment platform by utilizing simulation software according to the environment parameters and the vehicle parameters extracted in the step ; 3, setting parameters and constraint conditions of a reinforcement learning algorithmaccording to the simulation environment in the step 2; 4, training by using an A3C-SRU algorithm according to the simulation environment constructed in the step 2 to obtain a decision of an imported traffic flow scene; and 5, obtaining the optimal action sequence obtained in the step 2 according to the model obtained in the step 4, storing the trained model, and inputting the model into the intelligent vehicle to realize a traffic flow importing task. According to the A3C-SRU-based intelligent vehicle afflux traffic flow algorithm of the invention, real-time afflux traffic flow tasks can be effectively realized according to the settings of the steps 1-5.

The invention discloses a method for accurately positioning candidate bounding boxes in a target segmentation network, and the method comprises the steps: 1, constructing a lightweight convolutional neural network model, and designing network parameters; 2, solving an intersection-to-parallel ratio IoU of an IoU layer in the network model; 3, solving confidence scores of candidate bounding boxes in the network model; and 4, removing the redundancy of the candidate bounding box according to the setting of the intersection-parallel ratio and the confidence score, and finally enabling the confidence score of the output candidate bounding box to be highest and closest to a real calibration box. According to the method, a balance position can be found between target missing detection and errordetection, the candidate bounding box is closer to a real calibration box, and meanwhile, the network structure and the confidence coefficient solving method designed by the invention have certain generalization ability and can be applied to a relatively complex scene.

The invention provides a communication waveform comprehensive transmission performance evaluation method and system. The method comprises the following steps: a system classification step: classifyinga communication waveform efficiency evaluation system; a performance index simulation step: setting a transmission waveform and parameters, setting a simulation scene, and simulating each waveform performance index; an efficiency evaluation step: evaluating the comprehensive efficiency of the waveform by using a TOPSIS method; and a waveform index evaluation step: evaluating the waveform index byusing the gray hierarchical model. According to the invention, transmission effectiveness, resource effectiveness and anti-interception indexes of waveform indexes are simulated, a TOPSIS method is used, the waveform is comprehensively evaluated by adopting a gray multi-level evaluation method and a fuzzy neural network evaluation method, and the three evaluation means are unified by using the algorithm, so that the defects that the TOPSIS method is not easy to meet the actual demand, the gray multi-level evaluation method is too subjective and the fuzzy neural network is easy to overfit areavoided, and the evaluation system meets the user demand and has generalization ability.

The invention discloses a rotating machinery small sample fault diagnosis method based on a generative adversarial network. The method comprises the following steps: firstly, converting an acquired rotating machinery time domain signal into a frequency spectrum signal of which the characteristics are easy to observe through fast Fourier transform (FFT); and then, on the basis of the ACWGAN-GP, introducing an 'argmax' multi-classification thought so that an ACWGAN-GP-ARGMAX diagnosis model is obtained. According to the diagnosis model, a discriminator is endowed with classification and recognition capability, the classification and recognition capability of the discriminator is enhanced, and the accuracy and efficiency of fault diagnosis of the small sample of the rotating machinery can be effectively improved.

The invention discloses a refrigerant charge fault diagnosis method and device for a multi-split air conditioner. The method comprises the steps that S1, operation data of the multi-split air conditioner with different refrigerant charge are collected; S2, the collected operation data are processed through a principal component analysis (PCA) method to obtain characteristic operation data; S3, aBP neural network is trained by using the characteristic operation data to obtain a refrigerant charge fault detection model; and S4, whether the refrigerant charge of the multi-split air conditionermeets a preset standard or not is judged by utilizing the fault detection model. According to the refrigerant charge fault diagnosis method and device for the multi-split air conditioner, redundant parts in original data can be removed, so that the input data dimension is reduced; and the nonlinear modeling capability of the BP neural network can be exerted, so that the generalization capability and learning stability of the model are improved, and the convergence of the neural network is greatly improved.

The invention belongs to the field of radar image recognition, and relates to an HRRP target-free adversarial sample generation method based on deep learning. The method comprises the steps of training a deep neural network model by using a data set, and obtaining parameters thereof; selecting a sample and initializing algorithm parameters; for all sample categories, based on an FGSM algorithm, obtaining a disturbance scaling factor of each category by adopting a binary search method; selecting a minimum scaling factor from the disturbance scaling factors obtained by all the categories, calculating the gradient direction of the category corresponding to the scaling factor, and obtaining target-free fine-grained confrontation disturbance of n samples; adding the target-free fine-grained adversarial disturbance to the original sample to generate an adversarial sample; aggregating the target-free fine-grained adversarial disturbances of the n samples to obtain target-free general disturbances; and adding the target-free general disturbance to any sample to generate an adversarial sample. According to the method, target-free fine-grained disturbance and general disturbance can be obtained, corresponding adversarial samples are generated, and the safety of radar target recognition is improved.