341results about How to "Improve training accuracy" patented technology

The present invention provides a training method and system and a detecting method and system for a key human facial feature point detection model. The training method comprises: acquiring a human face position of an input picture; obtaining an initial position of a key feature point before updating according to an average key feature point of a training set and the face position; obtaining an initial position of a key feature point after updating according to a position of an authentic key feature point; according to a difference value between the initial positions of the key feature points before and after updating and a region feature extracted before updating, training a dynamic initialization regression model; and training a cascade regression model according to a distance difference between the initial position of the key feature point after updating and the position of the authentic key feature point and a region feature extracted after updating. The detecting method comprises: calling the dynamic initialization regression model and the cascade regression model in turn for a to-be-detected picture, and calculating a position of a key human facial feature point; and determining whether the key facial feature point is accurate according to a comparison with a preset point. By using the methods and systems provided by the present invention, the accuracy of detecting the key human facial feature point is improved.

The invention discloses an image classification method based on vector standardization and knowledge distillation, and the method comprises the following steps: (1), constructing and training a teacher model which employs a deep convolutional neural network; (2) constructing a student model smaller than the teacher model, wherein the student model also adopts a deep convolutional neural network; (3) training the student model by using a distillation loss function, and standardizing probability coding vectors output by the student model and the teacher model in the training process; and (4) inputting a to-be-classified image into the trained student model, and carrying out classification prediction. By utilizing the method, the performance of the student network can be improved and the image classification precision can be improved under the condition of not introducing extra parameters and calculation overhead.

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

The present invention discloses a medical image segmentation method based on a dual-path U-shaped convolutional neural network. A dual-path module and a U-shaped network are combined, original features are reused through stacking operation, and features are added to generate new features so as to improve the network training precision; the U-shaped convolutional neural network configuration is employed to perform transverse feature stacking between blocks so as to greatly improve the network training convergence speed and more rapidly obtain an ideal training model. The medical image segmentation method can meet the demands of data size of medical images requiring being processed is large and high processing efficiency and high precision requirements and can achieve a very good performancein the biomedicine detection application.

The invention discloses a model parameter verification method, device and apparatus in transverse federated learning and a medium, which are used for solving the problems of lower model training efficiency and accuracy caused by incapability of identifying invalid model parameters in the prior art. The method specifically comprises the steps that model parameters reported by all model training apparatuses are issued to other model training apparatuses except the model training apparatuses of the model training apparatuses through a server for error evaluation; the server can obtain an error value set of each model parameter; therefore, according to the error value set of each model parameter; screening invalid model parameters, therefore, in the model training process, when the initial model parameters used in the next model training period are determined, the invalid model parameters can be eliminated, so that the problem that the model training efficiency and accuracy are low due tothe fact that the invalid model parameters cannot be recognized is effectively solved, and the model training efficiency and accuracy are improved.

The invention discloses an intelligent traditional Chinese medicine diagnosis method. The method comprises the steps that a server acquires inspection training data, auscultation and olfaction training data, inquiry training data and pulse diagnosis training data of a patient from distributed client clusters; by means of the inspection training data, the auscultation and olfaction training data, the inquiry training data and the pulse diagnosis training data, the server trains a model to be trained built based on a deep neural network algorithm, and a trained model is obtained; by means of the trained model, the server diagnoses disease data of the patient, and diagnosis results of the disease data are obtained. According to the technical scheme, through comprehensive collection of the disease information of the patient, the accuracy of the medical diagnosis results is effectively improved, and moreover, disease data of multiple patients can be rapidly processed at the same time. Besides, the invention further discloses an intelligent traditional Chinese medicine diagnosis system and a traditional Chinese medical system correspondingly.

The invention relates to a photovoltaic string fault diagnosis method based on a kernel function limit learning machine. The method specifically comprises the following steps that 1, voltage-current characteristic scanning is conducted on a photovoltaic string, and curve fitting is conducted to acquire five photovoltaic internal equivalent parameters; 2, the acquired five photovoltaic internal equivalent parameters are integrated and unified; 3, a plurality of coefficients of an optimal KELM algorithm kernel function are calculated by adopting a pattern search algorithm; 4, the calculated coefficients are brought into the KELM, and samples are trained to obtain a training model; 5, fault detection and classification are conducted on the photovoltaic string through the training model. According to the photovoltaic string fault diagnosis method based on the kernel function limit learning machine, the accuracy of fault detection and classification on a photovoltaic power generation array can be effectively improved.

The invention relates to a non-model self-adaptive controller. The controller consists of two forward neural networks which are connected in series, wherein the front neural network is a network controller MLPc; and the back neural controller is a network simulator MLPo. The output of the network controller is taken as the input of the network simulator and a controlled object; a closed loop reverse feedback is arranged between the output of the controlled object and the input of the network controller; and the relation between the output of the network simulator and the output of the controlled object is set as the subtraction relation. The network simulator consists of a time lag multilayer perceptron (TDMLP) and is used for simulating the industrial process; and the network controller which is also a TDMLP is used for exerting the optimal control on the simulator in the precondition of considering model estimation errors. The non-model self-adaptive controller has the advantages of high control accuracy, high system response speed, and simplified network internal mechanism.

The invention discloses a short-term load predicting method based on a quick fuzzy rough set. The method comprises the following steps that firstly, electrical load data recorded by an electricity meter installed in a power grid are collected, and an initial attribute decision table is constructed; secondly, a fuzzy subordinate function of the condition attribute and the decision attribute is determined; thirdly, the attribute reduction is carried out according to the quick fuzzy rough set method, and the reduction condition attribute is obtained; fourthly, the reduction condition attribute serves as input data of a neural network to train normalized historical load data; fifthly, the neural network obtained through training is utilized for carrying out the short-term load prediction on an electric power system; sixthly, reverse normalization processing is carried out on the obtained normalization value of the maximum load of the prediction day, and a short-term electric power load prediction result is obtained and is the maximum load of the prediction day. The computing amount of the fuzzy rough set attribute reduction is small, the computing time is short, and the computing efficiency is improved.

The invention discloses an artery blood vessel image model training method, a segmentation method, a device and an electronic device, belonging to the technical field of digital image processing. Theartery blood vessel image model training method comprises the following steps: 1, pre-processing the acquired DSA image to construct an artery blood vessel image database; 2, labeling part of the sample images in the arterial blood vessel image library to construct a labeled sample image set; 3, constructing a convolution depth network and setting parameter of that depth network to generate an initial artery blood vessel segmentation model; 4, training an initial artery blood vessel segmentation model by using a label sample image set to generate an artery blood vessel image segmentation model; 5, further labeling the blood vessel target image obtained by using the artery blood vessel image segmentation model to segment other images except part of the sample images in the artery blood vessel image library, so as to carry out iterative training on the artery blood vessel image segmentation model. The embodiment of the invention can extract target blood vessels from DSA images with highaccuracy.

The invention provides a semantic comprehension model training method, which comprises the steps of obtaining a first training sample set, and performing denoising processing on the first training sample set to form a corresponding second training sample set; processing the second training sample set through a semantic comprehension model to determine initial parameters of the semantic comprehension model; responding to the initial parameters of the semantic comprehension model, processing the second training sample set through the semantic comprehension model, and determining updating parameters of the semantic comprehension model; and according to the update parameters of the semantic comprehension model, performing iterative update on the semantic representation layer network parametersand the task-related output layer network parameters of the semantic comprehension model through the second training sample set. The invention further provides a semantic comprehension model processing method and device and a storage medium. According to the method, the training precision and the training speed of the semantic comprehension model can be improved, so that the semantic comprehension model can adapt to different use scenes, and the influence of environmental noise on the semantic comprehension model is avoided.

The invention provides an annotation data generation method and device and a computer readable storage medium. The method comprises following steps of S100: acquiring a data universal set and annotated annotation data sets included in the data universal set; S200: analyzing data characteristics of the annotation data sets and manufacturing pseudo-data sets satisfying the data characteristics according to the data characteristics; S300: based on a GAN neural network, expanding the pseudo-data sets so as to form an expanded data sets; S400: identifying whether data in the expanded data sets needs to be annotated, and screening annotated data so as to form a training data set; S500: carrying out neural network trainings on the training data set so as to form a training model; and S600: basedon the training model, cleaning data in the data universal set except for the data in the annotation data sets, annotating data satisfying the training model and arranging the data in the annotation data sets. Thus, based on little data, a training data set which highly matches the sample data and is highly random can be rapidly and highly efficiently generated, and data quantity of the annotationdata is 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 reservoir optimal operation method based on a neural network model. For the nonlinearity and complexity of a reservoir group operation function, the main branch- neural network technology in computational intelligence is adopted, and a reservoir group is modeled. The method specifically comprises the steps that a target function is determined at first for reservoir operation, then an artificial neural network system is built between a target variable and variables influencing the target variable, actually-measured data are selected as samples, the artificial neural network achieves continuous learning according to a certain learning rule, complex nonlinear mapping is built between input and output, then other data are used as predicted samples to be input into a well-trained network, and therefore target variable prediction or function fitting can be conducted. Due to the facts the neural network directly achieves learning from the actually-measured data samples and time needed by the well-trained neural network during an application is short, requirements for instantaneity of reservoir operation are met, operation efficiency and learning precision and speed are greatly improved.

The invention discloses a deep learning model training method, device and system based on mixed precision, and the method comprises the steps: carrying out the data processing of sample data in a sample data set based on a deep learning model obtained through the last training, and obtaining a set number of first weight gradient data with the data precision being the first data precision; determining the data precision as a scaling coefficient of the second data precision according to the set number of first weight gradient data and the second data precision, the first data precision being higher than the second data precision; based on the sample data set and the scaling coefficient, training the deep learning model to update the weight parameters of the deep learning model, obtaining thedeep learning model trained this time, and the scaling coefficient is used for amplifying the loss value with the data precision being the second data precision in the process of training the deep learning model, so that the training efficiency and the training precision are improved.

The invention discloses a multi-label microblog text classification method based on semi-supervised learning, and relates to the field of natural language processing. The method comprises the following steps: firstly, preprocessing an original microblog text, and labeling a small amount of texts; generating augmented data of the annotation data set by using reverse translation, generating augmented data of the unannotated data set by using synonym replacement and random noise injection, guessing and generating a pseudo tag of the unannotated data by using a classifier, and forming a new training set together with the augmented annotation data set; converting the multi-label classification task into a plurality of dichotomy tasks, training a semi-supervised microblog text classification model, randomly extracting two samples from a new training set each time during training, generating a new sample in a text hidden space by using a sample mixing technology, calculating a loss value, and updating network parameters; and finally, classifying the microblog texts by comprehensively using a plurality of trained classifiers. The invention has an important application value for fine-grained information extraction of the microblog texts.

The invention relates to a GA-BP neural network algorithm-based city ecological construction evaluation method and belongs to the field of smart city ecological construction. The method comprises thefollowing steps of constructing a city ecological evaluation index system; generating neural network training and test samples; determining a topological structure of a neural network; setting relatedparameters of neural network training; optimizing initial weight and threshold of the neural network by utilizing a genetic algorithm; and inputting the training and test samples for training and checking the GA-BP neural network to finally build a GA-BP neural network algorithm-based city ecological evaluation model. The method can realize a city ecological construction evaluation function and has relatively high scientificity, validity and practicality.

The invention discloses a lower limb movement ability evaluation method based on an improved convolutional neural network. The method comprises the following steps: obtaining gait video images and human skeleton joint position information in the gait process of a subject; generating depth data for the gait video image and carrying out binarization processing through a bilateral filtering method soas to obtain a gait contour image; calculating a knee joint angle by using a space vector method; extracting gait contour features of the gait contour image by using an improved convolutional neuralnetwork; connecting the gait contour features and knee joint angles in series, performing normalization, performing feature dimension reduction on the features by using a kernel principal component analysis method, establishing a lower limb athletic ability evaluation index, and performing lower limb athletic ability evaluation on the subject. According to the method, the video image features areautomatically extracted by using the improved convolutional neural network in which the space pyramid pooling layer and the COCOB optimization algorithm are added in the traditional convolutional neural network, so that the complexity is greatly reduced, and the evaluation accuracy is improved.

The invention relates to a photovoltaic power station output power prediction method based on wavelet transformation and an extreme learning machine. Firstly, a prediction data set of a photovoltaic power station is extracted from a historical power generation and meteorological environment parameter monitoring data set of the photovoltaic power station; Secondly, preprocessing the photovoltaic power station prediction data set; then, a PCA algorithm is adopted to extract features from historical power data of the photovoltaic power station. Binary classification is carried out through a K-means algorithm, and the classification is divided into a smooth type and a fluctuation type; and finally, obtaining meteorological characteristic parameters of a to-be-predicted day through the NWP to generate a test set, judging the type of the test set according to the Euclidean distance, and traversing to find an optimal training set. And the output power of the photovoltaic power station is predicted by directly utilizing an extreme learning machine network in a smooth manner. And the fluctuation type needs to perform feature extraction on each object of the data through a WT algorithm, predict one by one and reconstruct predicted values. According to the photovoltaic power station output power prediction method based on the extreme learning machine, the accuracy of photovoltaic power station output power prediction can be effectively improved.

The invention relates to the technical field of biological recognition, and provides a body posture recognition method based on LSTM. The body posture recognition method comprises the steps: obtainingan action video of a to-be-recognized main body; extracting action feature information in the acquired action video of the to-be-identified main body through OpenPose, wherein the action feature information at least comprises skeleton key point information; and recognizing an action specification degree corresponding to the action feature information according to the action feature information and a pre-trained and generated body posture recognition model, wherein the body posture recognition model is a target neural network model generated according to a preset standard action, and the target neural network model is generated by training according to standard action feature information arranged according to a time sequence. According to the body posture recognition method, video actionsdo not need to be cut into isolated features to be recognized, and learning recognition is carried out through cooperation with the neural network model, and the body posture recognition process is rapid and accurate, and the user experience is improved.

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 relates to a scoring card model establishment method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring sample data of a plurality of training samples; Wherein the sample data comprises a plurality of sample variables; Carrying out box separation operation on each sample variable; Determining the number of sub-boxes correspondingto each sample variable, and comparing whether the number of sub-boxes exceeds a threshold; If yes, calculating the sub-box proportion of each sub-box corresponding to the sample variable, the bad sample rate and the chi-square value of the adjacent sub-box; Carrying out merging processing on the plurality of sub-boxes of the sample variables according to the sub-box proportion, the bad sample rate and the chi-square value, and returning to the step of determining the sub-box number corresponding to each sample variable; Otherwise, calculating the WOE value of each sample variable, screening the sample variables according to the WOE values, and establishing a scoring card model based on the screened sample variables. By adopting the method, the model training efficiency and precision can be improved.

The invention relates to an indoor fire prediction method based on a radial basis function neural network and a system thereof. The method comprises the following steps: (a) building an indoor fire probability prediction model which is positioned in a host computer; (b) setting needed monitoring nodes in a room, wherein the monitoring nodes collect indoor environmental parameters in real time and transmit the collected indoor environmental parameters to the host computer; (c) the host computer inputs the received indoor environmental parameters into the indoor fire probability prediction model so as to obtain the corresponding fire probability value of the current indoor environment; and (d) when the host computer obtains the corresponding indoor environmental judgment as a flame or a smoldering fire through the indoor fire probability prediction model, the host computer transmits alarm information to the monitoring nodes and carries out alarm prompting through the monitoring nodes. According to the method and the system, the indoor fire hidden trouble can be found timely, and the method and the system have the advantages of good real-time performance, high reliability and strong stability.

The invention discloses a tooth CT image segmentation method based on 3D multi-feature fusion, and belongs to the technical field of image processing. The method comprises the following steps: firstly, performing image conversion processing on a CT image, and converting the CT image into a grey-scale map; then constructing a neural network model for tooth segmentation and training the neural network model, wherein a skeleton of the neural network model adopts a U-net network; and finally, carrying out CT image preprocessing on an image to be segmented to obtain a grayscale image, inputting thegrayscale image into the trained neural network model, and obtaining a segmentation result based on the output of the trained neural network model. According to the invention, the 3D multi-feature fusion tooth segmentation method is provided by combining the similarity of the upper and lower CT images, and the training precision of the neural network is improved; meanwhile, in combination with aCRF algorithm, redundant information generated by the neural network model is removed, so that the segmentation result is more accurate.

The invention provides an information processing method, which comprises the steps of obtaining a first training sample matched with a use environment of a name prediction model, and determining an initial parameter of a first neural network and an initial parameter of a second neural network in the name prediction model; processing the feature set through a name prediction model, and determiningupdate parameters corresponding to different neural networks of the name prediction model; and according to the update parameters corresponding to different neural networks of the name prediction model, performing iterative update on the parameters of the first neural network and the parameters of the second neural network of the name prediction model through the feature set. The invention furtherprovides an information processing device, electronic equipment and a storage medium. According to the invention, the generalization ability of the name prediction model is stronger; the training precision and training speed of the name prediction model are improved; the name prediction model can adapt to different usage scenarios; and the processing efficiency and accuracy of name remarks are improved.

The invention discloses an image retrieval method. The image retrieval method comprises the following steps: carrying out feature description on a queried image and a querying image; carrying out deep learning on the queried image and the querying image; carrying out similarity measurement on the queried image by using features of the querying image to obtain a feedback list sequenced according to similarity; training a sample selection and classification device by using the querying image and images in the feedback list; carrying out classification prediction on front n(minute) images and pseudo negative example images by using the sample selection and classification device, and taking g images which are closest to a classification face; marking the front m images in the feedback list and the g images obtained in the previous step to obtain positive images and negative images; fusing features of the positive images and the querying image, and carrying out similarity measurement again on the queried image by using the fused features to obtain a final sequencing result. According to the image retrieval method disclosed by the invention, a query expansion method in image retrieval is realized by using an active learning method, and more accurate image retrieval can be realized on the premise of marking to a little amount of users.

The invention provides a semantic comprehension model training method. The semantic comprehension model training method comprises the following steps: recalling a training sample matching a vehicle-mounted environment in a data source; carrying out boundary corpus expansion processing on a statement sample with noise matching the vehicle-mounted environment; annotating statement samples with noise, which are subjected to boundary corpus expansion processing and matched with the vehicle-mounted environment, so as to form a first training sample set; processing the second training sample set through a semantic comprehension model; and according to the update parameters of the semantic comprehension model, performing iterative update on the semantic representation layer network parameters andthe task-related output layer network parameters of the semantic comprehension model through the second training sample set. The invention further provides a semantic comprehension method and deviceand a storage medium. According to the semantic comprehension model training method, the training precision and the training speed of the semantic comprehension model can be improved, so that the semantic comprehension model can adapt to a vehicle-mounted environment full duplex use scene, and the influence of environmental noise on the semantic comprehension model is avoided.

The embodiment of the invention provides a video data segmentation model training method, a video data segmenting method, a video data segmentation model training device and a video data segmenting device. The training method comprises the following steps of performing video feature detection on first video data so as to obtain information about one or more first video feature vectors; training by use of the information about the one or more first video feature vectors, so as to obtain a segmentation result; carrying out segmentation on the first video data by adopting a video data segmentation model, so as to obtain a segmentation result; judging whether the video data segmentation model meets preset verification conditions or not according to the segmentation result; if the video data segmentation model meets the preset verification conditions, outputting the video data segmentation model; if the video data segmentation model does not meet the preset verification conditions, training by use of the information about the one or more first video feature vectors again, so as to obtain the video data segmentation model. According to the embodiment of the invention, different video segmentation models are trained, so that automatic video data segmentation is realized, manual intervention operation is greatly reduced, the segmentation time is greatly reduced and the manpower cost is saved.

An embodiment of the invention provides a lip-reading recognition method and a mobile terminal. The lip-reading recognition method is applied to the mobile terminal. The mobile terminal is provided with vocal modes and silent modes by means of setting. Deep neural networks are trained in the vocal modes. In the silent modes, the lip-reading recognition method includes starting the silent modes; acquiring lip images of users; recognizing contents according to the deep neural networks. The contents correspond to the lip images. The deep neutral networks are established in the vocal modes. According to the technical scheme, the lip-reading recognition method and the mobile terminal in the embodiment of the invention have the advantages that the deep neural networks are trained in the vocal modes, the contents corresponding to the lip images are recognized in the silent modes by the aid of the deep neural networks trained in the vocal modes, and accordingly the technical problems of incapability of protecting the privacy and influence on surrounding personnel due to vocal conversation of existing users in the prior art can be solved; the privacy of the users can be protected, influenceon surrounding people can be reduced, the training time further can be saved, and the training accuracy can be improved.

The invention provides a method for constructing a risk evaluation model of green building material product authentication, comprising the steps of obtaining index parameters of green building material product authentication, wherein the index parameters are determined according to an authentication node and a risk point; acquiring sample values corresponding to the index parameters, and establishing a training sample and a test sample; inputting the index parameters into a neural network model and configuring the parameters of the neural network; training the neural network model by trainingsamples, and after the training, optimizing the neural network model by using test samples, wherein the optimized neural network model is a risk evaluation model of green building material product certification. The neural network model can extract the hidden knowledge and rules from the training data, and reduce the subjective factors and accidental factors in the evaluation process, which is more scientific, intelligent and expansible.