69results about How to "Strong generalization ability" patented technology

The invention discloses a deep learning-based image high-density population counting method. The method comprises the following steps of S1, establishing a depth complementation convolutional neural network by utilizing a deep learning framework caffe; S2, performing image data enhancement on an image according to operations of angle rotation, image multi-scale zooming, image mirroring and image pyramid zooming; S3, performing Gaussian kernel fussy normalization processing on the enhanced image data to obtain a real crowd density graph, outputting an estimated density graph and the real density graph by the network, and performing continuous iterative training optimization on the whole network structure according to a loss function; and S4, inputting a crowd picture and a tag picture to the network for training, and performing continuous iterative optimization to obtain a trained network model finally. According to the method, the end-to-end convolutional neural network is designed; a picture is given and input, and the estimated density graph corresponding to the picture is output, so that an estimated crowd number is obtained; and by outputting the density graph, more useful information is reserved.

The invention discloses an inland river water area abnormal behavior ship automatic identification system and method; the method comprises the following steps: obtaining inland river water area ship navigation AIS message information, hydrology meteorology department environment information, CCTV video images and depth of field images in real time; parsing ship abnormal behavior mode types, and building a ship abnormal behavior sample database; building a depth learning network model to parse ship behaviors, thus obtaining ship abnormal behavior mode and GPS positioning information; detecting the ship in the CCTV video image, using the depth of field image to obtain the ship three dimension space information, and thus obtaining the ship video positioning information; fusing the GPS positioning information, the video positioning information, the ship abnormal behavior mode and ship detected characteristics, carrying out ship target association, and automatically identifying the abnormal behavior ship in the CCTV video. The method and system can effectively solve the problems that the abnormal behavior ship cannot be easily identified in the CCTV video; the method and system are high in identification rate, can provide safety warning for waterborne traffic, thus further improving the maritime affair supervision intelligent level.

The invention discloses a method for screening the characteristic wavelength of a near infrared spectrum based on a heredity kernel partial least square method, which is used for detecting quality of food and farm products. The method comprises the following steps of: by utilizing a physicochemical analytical method, determining the concentration values of components to be detected of samples to be detected, and then dividing a calibration set and a predication set of the samples; by utilizing a genetic algorithm, carrying out global search on the preprocessed calibration set spectral data points; finally determining characteristic variable number participated in modeling according to a minimum cross-verification root-mean-square error value in the kernel partial least square method cross-verification process; by utilizing characteristic variables screened out from the genetic algorithm, forming a new data matrix again to be used as input of a model; taking a component concentration array of the sample to be detected of the calibration set as standard output of the model, so as to establish the best calibration analysis model; by virtue of the model, predicting the concentration values of components to be detected of the predicated set sample; and reducing the modeling operation time by screening the characteristic wavelength, and removing a large amount of noise and redundant variables, thus prediction performance and accuracy of the finally established model are higher.

The invention discloses an unsupervised cross-domain pedestrian re-recognition method based on clustering. The method comprises the steps: firstly inputting a source domain image with a label into a defined network architecture for pre-training, and obtaining a baseline model; then, the label-free target domain image subjected to style conversion being input into a base model for feature extraction and defining a pseudo label, and then the defined pseudo label being used for refining a model pre-trained in the last stage; and finally, loading the trained pedestrian re-recognition model, extracting pedestrian picture features of the to-be-retrieved picture and the target domain, retrieving the most matched pedestrian picture from the target domain, and outputting the most matched pedestrianpicture. According to the method, the practicability of the pedestrian re-recognition model in actual life is effectively improved, the re-recognition performance is improved, the network performanceis good, and the generalization ability is high.

The invention relates to a face living body detection method based on transfer learning, and belongs to the technical field of image processing and computer vision. The method comprises the followingsteps: segmenting video data into an image sequence, detecting faces in the image sequence, and dividing the data into a training set and a test set; Training the 3D convolutional neural network by using the training set of the source domain to obtain a label classifier for distinguishing true and false faces; Adding a gradient inversion layer behind the convolution layer, and extracting common features of a source domain and a target domain; Performing confrontation training on the data of the source domain and the target domain through a gradient inversion layer to obtain a domain classifierfor distinguishing the data of the source domain and the target domain; And sending the test set of the target domain into the trained label neural network, and selecting the maximum probability of network classification as the final detection result. According to the method, the idea of resistance transfer learning is applied to in-vivo detection, so that the generalization ability of in-vivo detection is improved; Through the 3D convolutional neural network, the spatial information and the time information of the video can be utilized, and the living body detection precision can be improved.

The invention discloses a measurement learning method and system for pedestrian re-recognition. Implementation of the method comprises the steps of collecting feature vectors of a pedestrian target under two cameras, building a positive sample pair feature vector set and a negative sample pair feature vector set, calculating the distance of positive sample pair feature vectors and the distance of negative sample pair feature vectors, restraining the negative sample pair feature vectors by adopting dual thresholds, building a measurement matrix based loss function under distance constraint conditions of the positive and negative sample pair feature vectors, and iteratively updating the measurement matrix by taking the minimum loss function value as the target. The measurement learning method can effectively reduce influences imposed on matrix learning by irrelevant variables such as image background and noise, so that occurrence of an overfitting phenomenon is avoided, and the acquired measurement matrix is high in generalization ability.

The invention provides a vehicle target joint cognition method and system based on point cloud and image data. The system comprises a data level combination module, a deep learning target detection wood block and a combination cognition module. The data level combination module obtains three-dimensional point cloud data and image data, and fuses the point cloud data and the image data. Fusion datais collected by the deep learning target detection module to be subjected to feature level detection and recognition, a detection result is output. The joint cognition module judges the feature levelfusion detection result and the data level fusion detection result by adopting an evidence theory method, reliability distribution is obtained to serve as output, and detection stability and robustness are improved.

The invention provides a training method of a translation model. The training method comprises the steps of obtaining a first training sample set; denoising the first training sample set to form a corresponding second training sample set; processing the first training sample set through a translation model to determine initial parameters of the translation model; responding to the initial parameters of the translation model, processing the second training sample set through the translation model, and determining updating parameters of the translation model; and iteratively updating encoder parameters and decoder parameters of the translation model through the first training sample set and the second training sample set according to the updating parameters of the translation model. The invention further provides a text processing method and device and a storage medium. According to the method, the generalization ability of the translation model can be stronger, the training precision and the training speed of the translation model are improved, and meanwhile, the gain of existing noise statements on model training can be effectively and fully utilized, so that the translation modelcan adapt to different use scenes.

The invention discloses a strong convection wind power grade prediction method based on weather radar data. The weather radar monitoring data is used as prediction information source, and the prediction of strong convection wind power is processed as "classification problem under supervised learning". According to the corresponding relation between the monitoring data of the strong convection wind power by the weather radar and the wind speed monitoring data of an automatic weather station, a classifier is constructed by means of serial SWM method, and the three-level prediction of wind power can be realized. Through the down-sampling method, the influence of the imbalance between low wind speed data sample and high wind speed data sample on the module classification can be reduced. The wind power prediction model can conduct quantitative prediction on ground wind brought by strong convection weather, and the deficiency that a conventional wind forecast of the meteorology department can not cover "small-scale, emergent and easily-passing-away" strong convection weather can be made up. The wind power prediction model can be the important support of the early warning of the risk of power transmission line wind deflection discharging under the weather of strong convection.

The invention discloses a visible light positioning method based on a particle swarm optimization extreme learning machine, and the method comprises the steps: comparing an individual fitness value and an individual extreme value Pbest of each newly obtained particle with a group extreme value Gbest, and updating the individual extreme value and the group extreme value if the new individual fitness value is better; when the number of iterations reaches a maximum value or the fitness value reaches a preset minimum value, outputting an optimal particle; taking the position of the optimal particle as a parameter of an input weight value and a hidden layer threshold value of an extreme learning machine; inputting the training sample set into an extreme learning machine neural network, carryingout neural network training, inputting the test sample set into the trained extreme learning machine neural network, and calculating an overall average positioning error. According to the method, theextreme learning machine is optimized by using the particle swarm algorithm, so that the extreme learning machine has the advantages of high learning speed, high generalization ability and the like of the extreme learning machine, the defect that the extreme learning machine randomly generates an input weight value and a hidden layer threshold value is overcome, and the network structure is simpler.

The invention relates to a conveyor belt carrier roller detection method based on YOLOv3. The method belongs to the field of computer vision and deep learning. The method comprises the following steps: replacing a feature extraction network darknet53 of YOLOv3 with a lightweight feature extraction network Mobilenet; and frame loss and center loss in the YOLOv3 loss function are replaced with GIOUloss, and an improved YOLOv3-Mobilenet belt carrier roller detection model is constructed. And training the model on a training set, testing the performance of the model on a test set, and comparing the performance test result with the performance of other models. The target recognition method provided by the invention is strong in generalization ability, realizes effective detection of the carrier roller, provides effective guarantee for subsequent judgment of whether the belt is separated from the track and monitoring of the running state, reduces the parameter calculation amount, and improves the speed and accuracy of an original YOLOv3 target detection model.

The invention provides a text processing model training method. The method comprises the steps of obtaining a first training sample set; denoising the first training sample set to form a correspondingsecond training sample set; processing the second training sample set through a text processing model to determine initial parameters of the text processing model; responding to the initial parameters of the text processing model, processing the second training sample set through the text processing model, and determining updating parameters of the text processing model; and iteratively updatingencoder parameters and decoder parameters of the text processing model through the second training sample set according to the updating parameters of the text processing model. The invention further provides a text processing method and device and a storage medium. According to the invention, the generalization ability of the text processing model is stronger, the training precision and training speed of the text processing model are improved, and the accuracy and readability of text generation are improved.

The invention discloses a method for mining a rare mode of multi-dimensional time sequence data. Massive multi-dimensional time sequence data sets are divided into one-dimensional time sequence data sets, and the minimum rare mode is calculated for the time sequence data of each dimension by utilizing an AMRSPM algorithm; the generated minimum rare modes serve as a pattern string dictionary, and all modes are searched out for each dimension of time sequence data through a character string matching algorithm Brute-Fore or a KMP algorithm; rare sub-sequences in the multi-dimensional time sequence data are mined through clustering (AP) operation.

The invention provides an urban sound event classification method based on dual-feature 2-DenseNet in parallel. The method has the advantages of being more efficient in fusion capability for feature information, higher in classification accuracy and stronger in generalization capability. The method comprises the following steps: S1, acquiring and processing to-be-processed audio data and outputting an audio frame sequence; S2, performing time domain and frequency domain analysis on the audio frame sequence, and respectively outputting a Mayer frequency cepstrum coefficient feature vector sequence and a gammatone cepstrum coefficient feature vector sequence; S3, constructing a classification model, wherein the classification model comprises a network model constructed by combining a two-order Markov model on the basis of a DenseNet model; the classification model taking a two-order DenseNet model as a basis to construct a basis network, arranging the basic network to be two parallel paths, and training the classification model to obtain a well-trained classification model; and S4, after processing the feature vector sequences output in the step S2, dividing the feature vector sequences output in the step S2 into two paths in a dual-feature mode to input into the well-trained classification model for classification and recognition so as to obtain a classification result of soundevents.

The invention belongs to the technical field of the computer network security and machine learning, and specifically relates to a training method of a network security situation model, a recognition method and a recognition device. The training method of the network security situation model comprises the following steps: S101, acquiring network data; S102, extracting a network data sample from the network data; S103, processing the extracted network data sample, analyzing the network data to obtain an attribute value of the network data, and judging whether the attribute value of the network data is safely corresponding to the network data; S104, analyzing and processing the attribute value of the network data to obtain a feature vector of the attribute value of the network data; and S105, determining a parameter of linear classifier through the feature vector of the attribute value of the network data.

The invention provides a photovoltaic cell panel deformation prediction method based on image processing and multi-dimensional perception, and the method comprises the steps: carrying out the extraction of key points of a photovoltaic cell panel image collected by an unmanned plane, obtaining a plurality of grid regions according to the key points, counting the number of pixel points in each gridregion, obtaining a reconstructed image based on the number of the plurality of pixel points, analyzing the reconstructed image, judging whether the cell panel is abnormal or not, and judging a corresponding deformation level if the cell panel is abnormal; acquiring a plate type and a deformation grade of the abnormal battery plate and historical wind direction and wind speed data corresponding tothe abnormal battery plate, and performing calculating to obtain wind power data according to the wind direction and wind speed data; training a time convolution network by taking the wind power dataon the plate type and the time sequence as a training data set and the deformation grade as a training label, and inputting the wind power data on the plate type and the time sequence into the time convolution network after the training is completed, thereby obtaining a predicted deformation grade. The method is high in deformation detection speed, time-saving and labor-saving.

The invention provides a novel support vector machine kernel function selection method in which a sparse representation theory is applied to support vector machine kernel function selection and application. The method includes the steps that (1) specific sample data are given and pre-processed, (2) an SVM kernel function sparse dictionary meeting Mercer conditions is selected and built, (3) a sparse code is solved, (4) an SVM kernel function type is selected according to the solved sparse code, (5) corresponding SVM parameters are optimized and a support vector machine model is determined, and (6) a prediction result is output. Attribution representation and modeling ability of sample data are achieved by means of the sparse theory, sample data prior information of actual problems is effectively utilized, metric characteristics of different kernel functions are taken into consideration for SVM modeling, generalization ability is high, and the defect that in a traditional SVM model selection method, the kernel function type is manually assigned and accordingly models can not have the optimal performance is overcome.

The invention discloses a position error compensation method of a robot. The method includes the following steps that (1) geometric parameter errors and non-geometric parameter errors which cause endeffector position errors of the robot are equivalently regarded as virtual pseudo errors; (2) an end effector position error curved surface caused by the pseudo errors is fitted by the aid of an artificial neural network, and a mapping model of the end effector position errors and robot joint parameters is built; (3) an optimization algorithm of the artificial neural network is designed on the basis of the multi-species co-evolutionary network method, the weight, the threshold value and the network structure of the artificial neural network are optimized, and the fitting precision and the generalization ability of the neural network are further improved; and (4) the end position errors are predicted by the aid of the optimized neural network, the robot joint parameters which are required to be compensated are obtained, and compensation is performed in the robot joint space. According to the method, the end effector position errors of the robot are compensated, so that the end positionprecision of the robot is improved.

The invention relates to a robot collision detection system and method based on a neural network. The method comprises the steps that the actual position of each joint motor is detected and recorded in each instruction period of a robot; the instruction position, the instruction speed and the actual position of each joint motor are taken as input characteristic data of the neural network to predict the position of the joint motor at the current moment; the deviation between the measured actual position at the current moment and the predicted joint motor position is compared, and if the deviation exceeds a set range, it is judged that collision occurs; and otherwise, it is judged that no collision occurs. According to the system and method, no additional hardware equipment is needed, and collision detection can be realized only by a position sensor of the robot. Meanwhile, kinetic parameters of the robot do not need to be known. Collision can still be accurately detected under the working condition that the load of the robot changes, and misjudgment cannot be caused in the acceleration and deceleration process of the robot. The detection method has the advantages of being high in collision detection precision, concise in algorithm, simple in calculation and good in universality.

The invention discloses a brain image segmentation method based on deep learning, and relates to the technical field of medical image processing and computer vision. The method comprises the steps: S1, obtaining an original brain image data set; s2, preprocessing the obtained original brain image data set; S3, importing the preprocessed brain image data set into a brain image segmentation model for training, and segmenting the brain image by using the brain image segmentation model which is maturely trained to finally obtain a segmentation result. The method improves the capability of obtaining image detail information of a network model, reduces the operation complexity of model training, solves the problem of edge blurring of segmented images, and improves the segmentation precision.

The invention discloses a method for recognizing typical defect local discharge signals of a power cable. The method includes the steps that local discharge data of a known type is acquired, characteristic parameters are extracted as input parameters, a discharge type label is set for each discharge type characteristic, and the discharge type labels are stored in an information library recognizedby the local discharge types; a neutral network model for recognizing the discharge types is built, the weight and the threshold value of the neutral network model are corrected through a bee colony algorithm, optimum model parameters of the input weight, the hidden layer threshold value and the output weight are acquired, and the optimum model parameters are saved; the local discharge signals ofthe to-be-recognized power cable are acquired based on the optimum model parameters, discharge pulse characteristic parameters are extracted, the to-be-recognized discharge characteristic parameters are input in the built neutral network model, recognition is carried out, and the discharge types are obtained. By means of the artificial bee colony algorithm, the weight and the threshold value of anextreme learning machine are optimized, the output weight is calculated through the obtained optimum weight and the obtained optimum threshold value, and the generalization capacity and the recognition precision of the extreme learning machine are improved.

The invention relates to an oil pumping well semi-supervised fault diagnosis method based on curvelet transformation and kernel sparsity. The oil pumping well semi-supervised fault diagnosis method comprises the steps that data of a plurality of indicator diagrams are obtained to serve as training samples; the indicator diagrams are converted into downhole pump indicator diagrams, and all the pump indicator diagrams are converted into grey images; all the pump indicator diagrams are subjected to curvelet transformation to obtain a coefficient matrix; feature vectors of all the pump indicator diagrams with labels are used as a dictionary, and sparse coefficients of all unlabelled pump indicator diagram feature vectors are evaluated; virtual tags of all the pump indicator diagrams with no label are calculated by means of the sparse coefficients; the feature vectors of all the pump indicator diagrams in the training samples are used as a dictionary; feature vectors of all to-be-diagnosed test samples are calculated to evaluate sparse coefficients; and fault types are diagnosed by calculating virtual tags of the to-be-diagnosed samples by means of the sparse coefficients. The oil pumping well semi-supervised fault diagnosis method can precisely describe features of the pump indicator diagrams, and a semi-supervised sparse expression classifier based on the method can effectively utilize information of unlabeled data and has a low requirement for the number of labeled samples.

The invention discloses a BP neural network-based pneumatic micro-droplet jetting state prediction method, and belongs to the field of micro-droplet jetting. According to the method, firstly, a BP neural network-based micro-droplet state prediction model is established, the prediction model takes an air pressure oscillation signal P(t) as an input, P(t) is collected by an intracavity high-speed pressure sensor, and a micro-droplet state is taken as an output. It is verified that the established model can accurately predict the micro-droplet jetting state. Common jetting state parameters comprise the number Nd of micro-droplets and the distance Hd of the micro-droplets relative to a nozzle at a certain time delay (with the rising edge of a high-speed electromagnetic valve driving signal asreference time). In an application example, the prediction accuracy of the number of micro-droplets is higher than 99%. Compared with a statistical average position of the micro-droplets obtained based on machine vision and image processing, the prediction precision of Hd through a prediction model of P(t) and BP neural networks can be improved by more than three times.

The invention discloses a tenant priority management method for a multi-tenant big data platform, and the method comprises the following steps: a tenant priority manager which consists of a resource request waiting queue and a radial basis neural network is initialized; When the cluster load is light, the resource request of each tenant can be met, and at the moment, the tenant priority manager isnot started; When the resource request of the tenant cannot be met, the resource request of the tenant is collected through a tenant priority manager and corresponding information is put into a resource request waiting queue; If the tenant needs to improve the resource request priority, an application is sent to the tenant priority manager, the tenant priority manager calls the radial basis neural network to calculate the tenant resource request priority, and then a judgment document is generated and sent to the system administrator for auxiliary judgment. Compared with a traditional resourceallocation strategy of the multi-tenant big data platform, the method can effectively overcome the defect of tenant priority processing of the multi-tenant big data platform.

The invention relates to a high-resolution remote sensing image water body extraction method based on low-level feature fusion, which is improved by using a DeepLabv < 3 + > deep learning model by taking the collection of a water body region of a high-resolution remote sensing image as a target, and is used for the extraction work of a high-resolution remote sensing image water body. The loss of feature information in the convolutional neural network is reduced by introducing the work of fusing low-level features and high-level feature maps, and detail feature information is supplemented; training is carried out through a remote sensing image data set with real water body marks, and information supplement of low-layer features is combined, so that the purpose of further distinguishing darkareas such as vegetation and building shadows from water body areas is achieved.

The invention discloses a road center line and double-line extraction method based on convolutional neural network regression, and the method comprises the following steps: predicting a road center line distance map and a road width map of a to-be-extracted high-resolution remote sensing image through employing a trained convolutional neural network; extracting a road center line by using a non-minimum suppression algorithm in combination with the road center line distance map; according to the extracted road center line, extracting road double lines in combination with a road width map; and selecting pixel points on the road center line as initial road seed points, calculating the road direction of the initial road seed points, reconstructing the topological structure of the road networkby using a road tracking algorithm, and outputting a road network extraction result. According to the method, through end-to-end training, the features easy to classify are directly learned from the training data, no post-processing is needed to extract the road centerline and sideline, the generalization ability is stronger, the road extraction precision is high, and the fine road extraction effect is better.

The invention discloses a face recognition system and method. The face recognition system comprises a face acquisition module, a face feature extraction module, a face recognition module and a face index module. The face acquisition module is used for acquiring a face image of a person to be recognized; the face feature extraction module is used for extracting face feature data of a person to be recognized from the face image; the face recognition module is used for carrying out similarity matching on the face feature data of the person to be recognized and the sample data of the known identity stored in the face index module; the face feature extraction is used for replacing a deep neural network formed by a global average pooling layer of the lightweight neural network with the GDConv Block based on a lightweight neural network; the neural network is deeper, stronger in generalization ability and better in robustness, the face feature data of the person to be recognized can be rapidly extracted on the basis of ensuring the accuracy, and then the identity of the person to be recognized can be rapidly and accurately recognized.

The invention belongs to the technical field of radars, and particularly relates to a radar human body motion state classification algorithm and system based on model fusion, and the method comprisesthe steps: obtaining a training set; constructing a support vector machine model according to the training set; obtaining a predicted value of the support vector machine model according to the supportvector machine model; and constructing an extreme gradient boosting tree model according to the prediction value of the support vector machine model. According to the method, a support vector machinemodel and an extreme gradient boosting tree model are fused through a staging model fusion method; the support vector machine model is suitable for processing small samples at high latitude, the extreme gradient boosting tree model has the advantage of strong fitting capability, and the fused model has the advantages of both the support vector machine model and the extreme gradient boosting treemodel, so that the model generalization capability is stronger, the recognition precision is higher, and the model training time in deep learning is reduced.

The invention discloses a robust image watermark embedding and extracting method and system based on deep learning. The method comprises: 1, collecting image data and dividing the image data into a training set and a test set; 2, obtaining a carrier image vector; 3, obtaining a watermark image by using the watermark embedding network, and calculating the distortion loss of the watermark image; 4, converting the watermark image into a lossy watermark image; 5, inputting the lossy watermark image into the watermark extraction network, extracting watermark information and calculating information extraction loss; 6, inputting the carrier image vector and the watermark image into a discriminator, and calculating the difference between the carrier image vector and the watermark image; 7, repeating the steps 2-5 by using the test set to calculate the robustness and imperceptibility of the watermark of the test set; 8, adjusting corresponding parameters according to the overall loss, repeating the steps 3-8 until the imperceptibility of the robustness watermark image of the test set watermark reaches a threshold value, and completing the training; and 9, carrying out watermark embedding and extraction by using the trained network. The invention further discloses a system using the method.

The invention discloses a synchronous tracking method of an actually measured revolving speed and a calculated revolving speed by a generator excitation device. The method comprises the following contents: preprocessing historical data of a calculated revolving speed and an actually measured revolving speed of a generator rotor as a training sample; utilizing the training sample to train a support vector machine model; calculating a predicted revolving speed through the support vector machine model data to be detected; using the predicted revolving speed as the actually measured revolving speed to perform synchronous tracking. The synchronous tracking method disclosed by the invention solves the problem that the computed result is inaccurate caused by the fact that when the revolving speed of the existing rotor is calculated, the transient process (or the transient state process) cannot be completely taken into consideration, and uses the calculated predicted revolving speed as the actually measured revolving speed to perform synchronous tracking, so as to prevent phenomenon of phase step or even generator tripping caused by loss of actual revolving speed measurement signals.