75results about How to "Few training samples" patented technology

The invention relates to a language recognition method and device based on a language model, a text classification method and device, computer equipment and a storage medium. The method comprises thesteps of obtaining a training word vector corresponding to a training statement, and inputting the training word vector into a to-be-trained first model and a trained second model to obtain a featurematrix output by each first network layer of the first model and a feature matrix output by each second network layer of the second model, wherein the first network layers and the second network layers are in one-to-one correspondence, and the network layer number of the first model is smaller than that of the second model; performing similarity calculation on the feature matrix output by each first network layer and the feature matrix output by the second network layer corresponding to each first network layer; and obtaining each similarity, adjusting model parameters of the first model basedon each similarity until the updated target similarity meets a convergence condition, obtaining a trained first model, and performing language recognition through the first model. The method can improve the model training efficiency.

The invention discloses a spatial spectrum attention hyperspectral image classification method based on Octave convolution, and solves the problems of large class spacing, small different class spacing and low classification accuracy in the prior art. The scheme is as follows: inputting images to be classified and preprocessing data, dividing a training set and a test set, constructing an Octave convolutional neural network, determining a loss function of the Octave convolutional neural network, training and updating the Octave convolutional neural network, testing the data of the test set, and completing hyperspectral image classification. According to the method, Octave convolution operation is used to reinforce feature representation, and a spatial attention mechanism and a spectral attention mechanism are introduced, so that the network can more accurately find an area which is more beneficial to classification and contains more comprehensive and detailed information. The method ishigh in classification precision and strong in robustness, and can be applied to analysis and management of hyperspectral image data.

The invention provides a text processing method and system. The text processing method comprises the steps of S1, establishing a classification hyperplane function; and S2, performing prediction on a newly-input text via the classification hyperplane function. The step S1 specifically comprises the sub-steps of S10, performing work segmentation processing on a text and establishing an entry document matrix ; S20, extracting features from the entry document matrix via the decision tree algorithm; S30, constructing the classification hyperplane function. The method and the system have the advantages that after word segmentation of a stored text, the sentence features of the text are extracted; the features are extracted according to the decision tree algorithm, so that the number of dimensions of model training points in a support vector machine is reduced and the training time is shortened. The feature vectors of texts are extracted according to decision tree training and text classification is performed by using a multi-core support vector machine according to the feature vectors, so that the method and the system have the advantages of accurate calculation, fewer model training samples, short training time and high text classification accuracy.

The invention relates to the field of a printing technology and particularly relates to an offset printing ink color matching method based on a least square support vector machine. The offset printing ink color matching method comprises the following steps of: sampling each foundation ink according to different concentration gradients, mixing foundation inks with m concentration ratios selected within a (1-100)% range of the concentration gradients with a reducer, and then measuring an XYZ value corresponding to each standard color sample obtained by sampling; obtaining an XYZ value of a target sample to be matched, and selecting N inks for color matching, so as to obtain m*N XYZ values and corresponding concentration ratio training sample data; training an LS-SVM (Least Square-Support Vector Machine) function model, and establishing relation of the XYZ values and an ink matching formula, so as to convert the XYZ values into the ink concentration ratios; and inputting the XYZ value of the target sample to be matched into the trained LS-SVM function model, so as to obtain the corresponding ink formula through calculation. With the adoption of the offset printing ink color matching method provided by the invention, the precision requirement on a foundation ink database is small, the color matching is simple and convenient and reliable, and the efficiency is high.

The invention discloses a gearbox fault diagnosis method based on a bird flock algorithm and a hidden Markov model. The gearbox fault diagnosis method comprises four steps of: feature extraction, model parameter initialization, parameter training, and output probability calculation. The step (1) feature extraction is implemented by selecting a wavelet function for performing 3-layer wavelet packetdecomposition and reconstruction on vibration signals, and analyzing wavelet decomposition coefficient signals of each frequency band so as to realize the extraction of feature information of different fault states represented by the vibration signals from each frequency band respectively. The step (2) model parameter initialization is implemented by taking frequency band energy of the vibrationsignals as eigenvectors for modeling. The step (3) parameter training is implemented by adopting the bird flock algorithm for re-estimation according to the parameters initialized in the second step.The steps (4) output probability calculation is implemented by extracting monitored vibration data features after models are constructed in the previous step, substituting the monitored vibration datafeatures into different fault state models, calculating an output probability by adopting a forward-backward algorithm, and regarding the maximum probability as a corresponding fault type.

The invention discloses a biological information recognition method based on dynamic sample selection integration, mainly solving the problem of low correct recognition rate of subclass samples caused by data imbalance. The realizing process for solving the problem comprises the following steps: (1) a training set is divided into a series of balanced sub data sets by adopting a training set dividing method; (2) the obtained balanced sub data sets are divided into respective matrix classifiers as initial training sets; (3) on the matrix classifiers, cyclic training is carried out by adopting a dynamic sample selecting method; (4) a testing set is tested by decision functions obtained in each training, thus obtaining decision results; (5) weight of the decision results is calculated by adopting a cost-sensitive idea; and (6) the decision results of each time are weighted and integrated, thus obtaining the final recognition result. Compared with the prior art, the method has the advantages of high accuracy and low calculation complexity, the size relation between a correct ratio and a recall ratio can be regulated as required, and the method is used for recognizing biological information, network intrusion and financial fraud and detecting anti-spam.

The invention discloses a vehicle GNSS/INS integrated navigation method based on a discrete gray neural network model. The method comprises the following steps: S1: solving the attitude, speed and position of a vehicle by using an inertial navigation value update algorithm according to the angular increment and specific force output by a micro inertial device; S2: establishing a discrete gray prediction model based on DGM (1,1); S3: improving a multi-layer neural network MLP; S4: designing a discrete gray neural network-based hybrid intelligent prediction algorithm DGM-MLP; S5: estimating thestate of a combined navigation system by using an inertial navigation error equation as the state equation, the difference between the position calculated by an INS and the position of a GNSS as an observed quantity or the difference between the position calculated by the INS solution and a pseudo position of the GNSS, and a Kalman filter KF; S6: performing output correction on the inertial navigation calculation result based on the position, speed and attitude errors estimated by the Kalman filter KF, and performing feedback correction on the inertial navigation based on a gyro and a meter error. The method can effectively solve the problem that the navigation precision is reduced when the GNSS signal loses efficacy.

The invention discloses an intelligent SAR radar air flight target recognition system, comprising an SAR radar, a database and an upper computer, wherein the SAR radar, the database and the upper computer are sequentially connected, the SAR radar performs real-time monitoring in the air and stores image data obtained by the SAR radar in the database, and the upper computer comprises an image preprocessing module, a feature extraction module, a feature selection module, a classifier training module, an intelligent optimization module and a result display module. The invention provides an air flight target recognition system which realizes on-line recognition and has high accuracy.

The invention relates to a hyperspectral image classification method, and concretely relates to a hyperspectral image classification method based on spectral space attention fusion and a deformable convolutional residual network. The objective of the invention is to solve the problem of low classification accuracy of hyperspectral images caused by insufficient spectrum and spatial feature extraction and overfitting under small samples due to the fact that the hyperspectral images contain abundant information in the existing hyperspectral image classification. The method comprises the following steps: 1, acquiring a hyperspectral image data set and a corresponding label vector data set; 2, establishing an SSAF-DCR network based on spectrum space attention fusion and a deformable convolution residual error; 3, inputting the x1, the x2, the Y1 and the Y2 into a network SSAF-DCR, and performing iterative optimization by adopting an Adam algorithm to obtain an optimal network; and 4, inputting x3 into the optimal network to carry out classification result prediction. The method is applied to the field of hyperspectral image classification.

The invention discloses a classification method of infrared spectrums of tea leaves based on possibility fuzzy identification C-means clustering. Infrared spectrum data of tea samples are collected byusing a Fourier infrared spectrum analyzer; the infrared spectrum data of the tea samples are preprocessed; dimension reduction processing is carried out on the infrared spectrum data of the tea samples after preprocessing by using a principal component analysis method; and identification information of the infrared spectrums of the tea training samples is extracted by using linear identificationanalysis. Possibility fuzzy identification C-means clustering is carried out on the training sample in the step four to obtain a clustering center; and tea class determination is carried out by usingthe possibility fuzzy identification C-means clustering method. According to the invention, on basis of combination of possibility fuzzy C-means clustering and linear discriminant analysis, the method has advantages of fast detection speed, fast classification speed and high classification accuracy and the like and is used for realizing correct classification of tea varieties.

The invention discloses a substation cable lightning strike interference detecting device and method and belongs to the field of power system substation electromagnetic compatibility. The substation cable lightning strike interference detecting device and method aim at recognition of interference of lightning impulse to substation cables and interference caused by other interference sources, and utilizes the H-S conversion technology to extract energy and the maximum of various frequency bands of voltage interfering signals and corresponding characteristic frequency. Compared with the prior art in which small wave packages are used for extracting characteristic values, the H-S conversion can effectively filter interference such as noise. The detecting device and method automatically recognize the type of interference sources and interference modes through a support vector machine. Compared with a neural network, the detecting device and method are small in number of needed samples, short in training time, free of overall optimal solution problems, and good in stypticity so as to be high in recognition rate. The detecting device and method provide a basis for taking effective measures to control and remove disturbance signals in secondary cables, and avoid error operation of secondary equipment connected with a cable or equipment damage.

The invention discloses an intelligent SAR radar marine ship target identification system. The system comprises a SAR radar, a database and an upper computer. The SAR radar, the database and the uppercomputer are successively connected. The SAR radar carries out real-time monitoring on a sea area and image data acquired by the SAR radar is stored in the database. The upper computer comprises an image preprocessing module, a characteristic extraction module, a characteristic selection module, a classifier training module, an intelligent optimization searching module, and a result display module. The invention provides the high-precision marine ship target identification system capable of realizing on-line identification.

The embodiment of the invention provides a training method of an image generation model, a generation method and device and equipment, and relates to the technical field of machine learning and image processing. The training method comprises the steps that a first transformation model is obtained through training, the first transformation model is used for generating a first training image based on a first noise sample, and the first training image is an image of a first type of style; training is performed to obtain a reconstruction model based on the first transformation model; a second transformation model is obtained through training, the second transformation model is used for generating a second training image based on the second noise sample, and the second training image is an image of a second type of style; the first transformation model and the second transformation model are grafted to generate a grafted transformation model; and based on the reconstruction model and the grafted transformation model, an image generation model is generated, and the image generation model is used for transforming the to-be-transformed image of the first type of style into a target image of a second type of style. By adopting the technical scheme provided by the embodiment of the invention, the time cost of model training can be reduced.

The invention discloses an indirect urban high-resolution impervious surface extraction method, comprising the following steps: S1, constructing a conceptual calculation model for indirectly extracting a high-resolution impervious surface of a city, wherein the principle of the conceptual calculation model is to obtain the impervious surface indirectly by removing the background surface other thanthe impervious surface in an urban area; S2, obtaining a high-resolution remote sensing image of the urban area; S3, selecting the training sample and the accuracy verification data set; S4, extracting the background surface of the impervious surface of the urban area by using the high-resolution remote sensing image and the training sample according to the conceptual calculation model proposed in the step S1; S5, synthesizing the background surface of the impervious surface according to the result of the step S4, using the background surface as a mask in the entire urban area, and the remaining part as the impervious surface, so as to indirectly obtain the high-resolution urban impervious surface extraction result; and S6, performing the accuracy verification of the impervious surface indirect extraction result in the accuracy verification data set.

The invention discloses a training method of a text information processing model, a text information processing method, system and device and a storage medium, which can be applied to the field of artificial intelligence; the training method comprises the following steps: coding a first training text to obtain first semantic information; encoding the first candidate word to obtain second semantic information; predicting classification of the first training text according to the first semantic information to obtain a first classification result; predicting the weight of the first candidate word according to the first semantic information and the second semantic information; determining a first loss value according to the first classification result and the type label of the first training text; determining a second loss value according to the weight of the first candidate word and the weight label of the first candidate word; and training the text information processing model according to the first loss value and the second loss value. According to the method, the training cost can be reduced or higher model precision can be obtained under the condition of the same training cost.

The invention relates to a Weiqi referee system based on an MLP neural network and computer vision. The Weiqi referee system comprises an image normalization processing module, an MLP neural network module and a Weiqi referee algorithm module. The image normalization processing module is used for preprocessing an image by means of channel transformation, image cutting, light equalization processing, corner detection and the like, so that subsequent recognition is facilitated. The MLP neural network module comprises a chessboard recognition model and a chess piece recognition model and is usedfor recognizing the positions of the chessboard and the black and white chess pieces and storing the information of the chessboard and the black and white chess pieces in a TXT file. The later-stage Weiqi referee algorithm module is used for judging the winning or losing of the chess game, obtaining the winning or losing state of the black and white chess according to an algorithm by reading the state and position information of the black and white chess in the TXT file, converting a result into an SGF (universal go chess manual) picture and displaying the SGF picture to a user.

The invention discloses an SAR radar aerial flight target recognition system. The system comprises an SAR radar, a database and a host computer; the SAR radar, the database and the host computer are orderly connected with each other; the SAR radar performs real-time monitoring on the air, and stores image data obtained by the SAR radar into the database; and the host computer comprises an image pre-processing module, a feature extraction module, a feature selection module, a classifier training module and a result display module. The aerial flight target recognition system provided by the invention achieves on-line recognition and is high in accuracy.

The invention discloses a carbon price prediction method and device, a computer device and a storage medium. The method comprises the steps that the current energy future price of specified energy isacquired; and the current energy future price is input into a preset carbon future price prediction model based on a least squares support vector machine to carry out price prediction, wherein the kernel function of the carbon future price prediction model is an RBF kernel function. According to the invention, based on the least squares support vector machine, the RBF kernel function is selected to establish the carbon future price prediction model, and then the price of carbon futures is predicted; an inequality constraint in a standard SVM algorithm is replaced by an equality constraint; solving quadratic programming problems is transformed into directly solving linear equations; the method is suitable for price prediction in long, medium and short periods, and has the advantages of non-overlapping reflected information and small training sample; and local minimum points which affect the final prediction result are avoided.

A method for quickly estimate voltage of distribution network includes unify line type by load moment normalization coefficient, proposing nodal load moment concept, establishing mapping relation between nodal load moment and nodal voltage, and utilize Levenberg- The BP neural network based on Marquardt algorithm achieves the full coverage of voltage estimation under the condition of a small number of training samples. The convergence speed and accuracy are better than the traditional gradient descent algorithm, which has a strong practicability. The invention can provide more accurate voltagecalculation results, and provides powerful support for the staff to master the voltage distribution and the low voltage management of the table area.

The invention discloses a method for accurately recovering an original image of a digital image after color correction, and the method comprises the steps: generating a color thumbnail matrix based on K-means clustering according to the original image and the image after color correction, training a nonlinear model of a least square support vector machine, and estimating an error compensation matrix for an image saturation region; obtaining an image recovery model and embedding the image recovery model into the image; and when the original image needs to be restored, extracting model parameters from the image embedded with the restoration model to perform transformation processing, and accurately restoring to obtain the original image. According to the method provided by the invention, after the digital image is subjected to color correction or some nonlinear processing, even if partial pixel color component values exceed a gray scale range to cause truncation and saturation, the original image color can still be effectively and accurately recovered. The method provided by the invention is beneficial to good color traceability of the digital image in the transmission and copying process, and can be effectively applied to accurate readjustment of image color display and reediting after image color correction in the printing and copying process.

The invention discloses an intelligent SAR radar land tank target recognition system. The system includes SAR radar, a database and a host computer. The SAR radar, the database and the host computer are connected in turn. The SAR radar carries out real-time monitoring on land, and image data obtained by the SAR radar are stored into the database. The host computer includes an image preprocessing module, a feature extraction module, a feature selection module, a classifier training module, an intelligent optimization module and a result display module. According to the land tank target recognition system provided by the invention, online recognition is realized, and precision is high.

The invention provides a remote sensing image sample intelligent collecting method. According to the method, a sample set required by image classification can be effectively selected, and sampling collecting time and money cost can be saved. The method comprises the steps that as for a remote sensing image to be classified, a few samples are randomly marked by a user; image classification is performed on images by means of the samples; classification results are converted into various types of probabilities; a sample set which is not marked and has the largest information amount is selected; the user performs type marking on the sample set which is not marked; the sample set just marked and the existing sample set constitute a new sample set; the images are trained again by using the new sample set; the process is performed in an iterated mode; iterating is stopped to obtain a set of samples when a certain condition is met.

The invention provides an abstract generation method and device, electronic equipment and a medium. The abstract generation method includes the steps: obtaining at least one announcement abstract of at least one enterprise and performing duplicate removal processing; preprocessing each announcement abstract subjected to duplicate removal processing to obtain at least one segmented word; inputtingat least one segmented word of each announcement abstract into a pre-trained parameter extraction model, generating at least one abstract template, and fusing at least one abstract template to obtainan abstract template library; and when an abstract generation instruction is received, extracting a target text from the abstract generation instruction, determining a text type to which the target text belongs, determining an enterprise type to which an enterprise corresponding to the target text belongs, determining a target abstract template matched with the text type and the enterprise type atthe same time, extracting information required by the target abstract template from the target text, generating an abstract corresponding to the target text, and obtaining the abstract template by analyzing the published announcement abstract, so that the abstract generation accuracy can be improved.