2116 results about "Training time" patented technology

The most preferred embodiment of the present invention is a computer-based decision support system that includes three main components: a database mining engine (DME); an advertising optimization mechanism; and a customized user interface that provides access to the various features of the invention. The user interface, in conjunction with the DME, provides a unique and innovative way to store, retrieve and manipulate data from existing databases containing media-related audience access data, which describe the access habits and preferences of the media audience. By using a database with a simplified storage and retrieval protocol, the data contained therein can be effectively manipulated in real time. This means that previously complex and lengthy information retrieval and analysis activities can be accomplished in very short periods of time (typically seconds instead of minutes or even hours). Further, by utilizing the advertising optimization mechanism of the present invention, businesses, networks, and advertising agencies can interactively create, score, rank and compare various proposed or actual advertising strategies in a simple and efficient manner. This allows the decision-makers to more effectively tailor their marketing efforts and successfully reach the desired target market while conserving scarce advertising capital. Finally, the user interface for the system provides access to both the DME and the optimization mechanism in a simple and straightforward manner, significantly reducing training time.

The invention discloses a method for personalized television voice wake-up by voiceprint and voice identification, particularly a method for performing identity confirmation on a television user through voiceprint identification and controlling a television to perform personalized voice wake-up through confirmed identity and a voice identification result of user voice, and relates to voiceprint identification and voice identification technologies. A composition system comprises a voice control system (1), an information storage unit (2) and a television main controller (3) which are connected through electric signals. The method has the characteristics of short training time, very high voiceprint and voice identification speed and high identification rate. Voiceprint and voice identification can be finished by only offline training and testing, identification results do not need to be sent to a cloud server, use is convenient, and the safety of family information is guaranteed. The method also can be applied to user-personalized automatic voice channel change of the television, can be transplanted to a common high-speed DSP (digital signal processor) or chip for operation, and can be widely applied to the related fields of smart homes.

The invention relates to a cascaded depth neural network-based face attribute recognition method. The method includes the following steps that: 1) a cascaded depth neural network composed of a plurality of independent convolution depth neural networks is constructed; 2) a large number of face image data are adopted to train networks at all levels in the cascaded depth neural network level by level, and the output of networks of previous levels is adopted as the input of networks of posterior levels, such that a coarse-to-fine neural network structure can be obtained; and 3) the coarse-to-fine neural network structure is adopted to recognize the attributes of an inputted face image, and final recognition results can be outputted. According to the cascaded depth neural network-based face attribute recognition method of the invention, a cascade algorithm system is adopted based on depth learning, and therefore, training time can be accelerated; and a cascaded coarse-to-fine processing process is realized, and the performance of a final network can be improved by networks of each level through utilizing information of networks of upper levels, and therefore, the speed and the accuracy of face attribute recognition can be effectively improved.

The invention discloses a convolution neural network parallel processing method based on a large-scale high-performance cluster. The method comprises the steps that (1) a plurality of copies are constructed for a network model to be trained, model parameters of all the copies are identical, the number of the copies is identical with the number of nodes of the high-performance cluster, each node is provided with one model copy, one node is selected to serve as a main node, and the main node is responsible for broadcasting and collecting the model parameters; (2) a training set is divided into a plurality of subsets, the training subsets are issued to the rest of sub nodes except the main mode each time to conduct parameter gradient calculation together, gradient values are accumulated, the accumulated value is used for updating the model parameters of the main node, and the updated model parameters are broadcast to all the sub nodes until model training is ended. The convolution neural network parallel processing method has the advantages of being capable of achieving parallelization, improving the efficiency of model training, shortening the training time and the like.

The invention discloses a Chinese named entity recognition method based on BERT-BiGRU-CRF. The method comprises three stages: in the first stage, preprocessing mass text corpora, and pre-training a BERT language model; in the second stage, preprocessing the named entity recognition corpus, and encoding the named entity recognition corpus through the trained BERT language model; and at the third stage, inputting the encoded corpus into a BiGRU+CRF model for training, and performing named entity recognition on the to-be-recognized statement by using the trained model. Construction of the Chinesenamed entity recognition method based on BERT-BiGRU-CRF is carried out, semantic representation of characters is enhanced through a BERT pre-training language model, semantic vectors are dynamicallygenerated according to contexts of the characters, and the ambiguity of the characters is effectively represented. Compared with a method based on fine tuning of a language model, the method has the advantages that training parameters are reduced, and the training time is saved.

The invention provides a vehicle model recognition model construction method based on depth learning. According to the method, a deep convolution neural network structure is designed for vehicle model recognition; softmax classification loss and sorting loss which is built through a vehicle model type hierarchical structure and is based on tetrad are used together to supervise the deep convolution neural network training; and intra-class differences and inter-class differences of vehicle models are constrained at the same time to learn character representation with rich semantic information and a discriminating classifier. In order to speed up network convergence, an online difficult sample mining policy and an improved gradient backtracking optimization algorithm are used, which greatly reduces the network training time. The invention further provides a vehicle model recognition method based on depth learning. The vehicle model recognition method uses the model constructed by the model construction method to recognize the vehicle type in a vehicle picture, which effectively improves the vehicle model classification accuracy.

The invention discloses an intensive learning based urban intersection passing method for a driverless vehicle. The method includes a step 1 of collecting vehicle continuous running state informationand position information through a photographing method, the vehicle continuous running state information and position information including speed, lateral speed and acceleration value, longitudinal speed and acceleration value, traveling track curvature value, accelerator opening degree and brake pedal pressure; a second step of obtaining characteristic motion track and the velocity quantity of actual data through clustering; a step 3 of processing original data by an exponential weighting moving average method; a step 4 of realizing the interaction passing method by utilizing an NQL algorithm. The NQL algorithm of the invention is obviously superior to a Q learning algorithm in learning ability when handling complex intersection scenes and a better training effect can be achieved in shorter training time with less training data.

The invention discloses a traffic state prediction method for the urban road network based on key road sections, which is characterized by comprising the steps of first, carrying out data preprocessing; second, establishing a spatial weight matrix of the road network; third, establishing a time correlation matrix; fourth, recognizing key road sections by using a time-space correlation matrix; andfifth, establishing a deep convolution neural network, predicting the state of the road network in the future, and carrying out evaluation on a prediction model. The traffic state prediction method predicts the urban traffic flow state from a level of the wide-range road network, thereby being conducive to guiding the traffic flow from a macroscopic perspective, and fully exploring time-space correlation characteristics of the traffic flow. The key road sections in the road network are recognized, so that the training time of the model can be greatly reduced compared with a method of taking historical states of all road sections as input data, and the prediction efficiency is improved; and the convolution neural network is adopted to serve as the prediction model, and the prediction resultis also more accurate.

The invention discloses a traffic sign deep learning mode identification method. The method includes the following steps: pretreatment is performed on a test sample and, a training sample of a trafficsign image, a residual deep learning network based on the multiscale characteristic weight operation fusion of a convolution nerve network is designed, the characteristic is automatically extracted by the network in order to eliminate the artificial trace mainly through training, and identification is performed on the traffic sign test sample by using a deep color characteristic trained classifier. The image characteristic weight operation combining with the multiscale convolution fusion network is applied to the traffic mode identification technology, the training efficiency of the network is substantially improved, and the problem that the accuracy and the real-time performance are not ideal, the network structure is complex, the training time is long, the stability and the robustness are bad encountered in the traffic sign identification method can be effectively solved. The image recognition accuracy of the trained network in 43 kinds of traffic signs reaches 97%.

The invention discloses a diagnosis method for the fault position and the performance degradation degree of a rolling bearing, belonging to the technical field of fault diagnosis for bearings, and solving the problems of low accuracy of diagnosis for fault position and performance degradation degree, and high time consumption of training existing in an intelligent diagnosis method for a rolling bearing in the prior art. A white noise criterion is added in the disclosed integrated empirical mode decomposition method, so that artificial determination for decomposition parameters can be avoided, and the decomposition efficiency can be increased; and via the disclosed nuclear parameter optimization method based on a hypersphere centre distance, the small and effective search region of nuclear parameters in a multi-classification condition can be determined, so that training time is reduced, and the final state hypersphere model of a classifier is given. The intelligent diagnosis method based on parameter-optimized integrated empirical mode decomposition and singular value decomposition, and combined with a nuclear parameter-optimized hypersphere multi-class support vector machine based on the hypersphere centre distance is higher in identification rate compared with the existing diagnosis method. The diagnosis method disclosed by the invention is mainly applied to intelligent diagnosis on the fault position and the performance degradation degree of the rolling bearing.

In one embodiment, a pressure transmitter is provided which diagnoses the condition of a primary element and / or an impulse line which connects to a pressure sensor. A difference circuit coupled to the pressure sensor has a difference output which represents the sensed pressure minus a moving average. A calculate circuit receives the difference output and calculates a trained output of historical data obtained during an initial training time. The calculate circuit also calculates a monitor output of current data obtained during monitoring or normal operation of the transmitter. A diagnostic circuit receives the trained output and the monitor output and generates a diagnostic output indicating a current condition.

Embodiments of the present invention provide a deep neural network learning method. The method comprises: conducting, by a plurality of processors, forward processing on data distributed to the processors layers in parallel layer by layer from a first layer to a last layer, and acquiring error information when forward processing is finished; and conducting, by the plurality of processors, backward processing on the error information layer by layer from last layer to first layer, wherein each of the plurality of processors immediately transfers a parameter correction value to other processors after backward processing of a current layer of a corresponding deep neural network model generates the parameter correction value. With the method according to the embodiments of the present invention, time consumed by transfer of the parameter correction values is reduced, and efficiency of training the deep neural network models is effectively improved; and particularly under the conditions of a large volume of training data and a great number of layers of each deep neural network model, such manner can greatly reduce used time, and effectively save model training time. Further, the embodiments of the present invention provide a processor, and a deep neural network learning system.

The invention discloses an on-line sequential extreme learning machine-based incremental human behavior recognition method. According to the method, a human body can be captured by a video camera on the basis of an activity range of everyone. The method comprises the following steps of: (1) extracting a spatio-temporal interest point in a video by adopting a third-dimensional (3D) Harris corner point detector; (2) calculating a descriptor of the detected spatio-temporal interest point by utilizing a 3D SIFT descriptor; (3) generating a video dictionary by adopting a K-means clustering algorithm, and establishing a bag-of-words model of a video image; (4) training an on-line sequential extreme learning machine classifier by using the obtained bag-of-words model of the video image; and (5) performing human behavior recognition by utilizing the on-line sequential extreme learning machine classifier, and performing on-line learning. According to the method, an accurate human behavior recognition result can be obtained within a short training time under the condition of a few training samples, and the method is insensitive to environmental scenario changes, environmental lighting changes, detection object changes and human form changes to a certain extent.

A static gesture recognition method based on a multi-scale convolution neural network is firstly proposed. The invention is based on the Caffe frame of depth learning to carry out optimization design,and uses the technical principle of image processing to recognize the static gesture picture. Firstly, the static gesture image data in simple background and complex background are collected and preprocessed. The data are divided into training data and test data. After setting up the experiment and testing environment, the convolution neural network based on multi-scale is designed, that is, determining the number of neural network layers, selecting the appropriate scale features, and so on. The training data are put into the network structure for learning and then the test data samples are input for testing, and the recognition accuracy is obtained. The invention can automatically learn gesture features by using a convolution layer and overcomes the shortcomings of manual feature extraction and the shortcomings that common convolution neural network feature extraction is not precise and comprehensive enough and the stability is not good enough, and the recognition accuracy is higher,and the training time is equal. The method has strong flexibility and wide applicability.

The invention discloses the technical field of wind power prediction, particularly, relates to an ultrashort-term power prediction method. The method comprises the following steps of: firstly, acquiring the wind speed, the wind direction and the wind power of a wind power farm to form a sample set; then, preprocessing the data of the sample set; reducing dimensions of the preprocessed sample set by a depth autocoder network; and finally, training a relevance vector machine regression model by the sample set with reduced dimensions, and predicting the ultrashort-term wind power through the trained relevance vector machine regression model. The method reduces the training time of a prediction model, satisfies the requirements on precision and real-time property in system status estimation, and enables the prediction method to be more accurate.

The invention discloses a multiple features fused bidirectional recurrent neural network fine granularity opinion mining method. The method comprises the following steps of: capturing comment data of a specific website through internet and carrying out labelling and preprocessing on the comment data to obtain a training sample set; carrying out training by using a Word2Vec or Glove model algorithm to obtain word vectors of the comment data; carrying out vectorization after carrying out part of speech labeling, dependence relationship labeling and the like; and inputting the vectors into a bidirectional concurrent neural network to construct a bidirectional recurrent neural network fine granularity opinion mining model. According to the method, attribute words in fine granularity opinion mining is extracted and emotional polarity judgement is carried out through the training of a model, so that plenty of model training time is further saved and the training efficiency is improved; no professionals are required to carry out manual extraction on the attribute words, so that a lot of manpower cost is saved; and moreover, the model can be trained by using a plurality of data sources, so that cross-field fine granularity opinion analysis can be completed, thereby solving the problem of long-distance emotional element dependency.

The invention relates to a robust ear detection method which improves an AdaBoost arithmetic and belongs to the technical field of image mode identifying. The invention is characterized by proposing an ear detection method with excellent performances under a complex underground. The invention proposes four anisomerous Haar-like corner characteristics which are used for describing the grayscale changes on the partial areas of the ears; a policy of subsection selection is adopted for selecting the best sorting threshold of the Haar-like characteristics, thus reducing the sample training time; the weight of a weak sorter is modified for reducing the mistaken detection rate of the sorter; the threshold HW is set and eliminated according to the distribution change of the sample weight in the training, thereby preventing an over-studying phenomenon from being generated and leading the miss-detection rate and the mistaken direction rate of the ear detection to be reduced; besides, the invention also provides a single-ear detection policy for leading both the defection efficiency and the detection effect to be improved. The excellent performances of the robust ear detection method are shown on a PC machine and a DSP.

The invention discloses an OCR recognition method. The method comprises the steps of obtaining a to-be-recognized image of business party data; Inputting the to-be-identified image into a general OCRtemplate for identification to obtain text information recorded in the to-be-identified image and position information corresponding to the text information, Wherein the universal OCR template comprises a detection model and a universal identification model, and the universal identification model is obtained by training field image samples of various service types of a service party; And synthesizing the text information and the position information corresponding to the text information into structured identification data. The invention further provides an OCR electronic device. According to the OCR identification method and the electronic equipment thereof, the image of the to-be-identified object (such as a contract, an invoice, a bill, a certificate and the like) can be efficiently andrapidly identified through the general OCR template, the structured identification data is generated, and the identification from the optical character to the text information is completed. The universal OCR template adopted in the method is short in training time, high in adaptability, capable of adapting to various different to-be-identified objects, high in identification accuracy and high in overall efficiency in the identification process.

A method for identifying diseases and pests of fruit trees under complex background based on cavity convolution is disclosed, At first, that original train set with sample labels is obtain by collecting images of diseases and insect pests of fruit tree leaves, and then the expanded train sample set is obtained by data augmentation operation, and then the average value is obtained, all the originaltraining set images are averaged, and then randomly scrambled to form a training data set; Then, the cavity convolution neural network model is established, and the disease classification training iscarried out by using the training data set. The characteristic image of the input image is extracted by the cavity convolution layer, and the nonlinear characteristic image of the input image is obtained by using the nonlinear activation function. Pooling layer is used to reduce the convolution layer weight parameters. Using multi-scale convolution kernel to obtain different features of the inputimage, and finally using Soft Max classifier to classify the disease types of the input image; The invention shortens the training time of the model, accelerates the recognition speed and improves the recognition accuracy.

The invention provides a human falling detection method based on machine vision. Under the conditions that all detected objects are people in indoor environment and the activities of all people can be captured by a depth camera, the method comprises the following steps of: (1) acquiring an image by using the depth camera; (2) subtracting a background from the image which is acquired, and extracting a foreground from the image; (3) acquiring a foreground contour by using a binary image contour extraction algorithm; (4) performing convolution computation on the foreground contour and Gaussian functions in different scales, mapping a contour image to a curvature scale space (CSS) to form a CSS image, extracting peak points of the CSS image, and thus obtaining a video word bag based on curvature scale characteristics by the peak points in different scales; and (5) training an extreme learning machine classifier by using the video word bag which is obtained. By adoption of the method, under the condition of few training samples, an accurate falling detection result can be obtained within a short training time.

The invention provides a neural network equalization method used for an indoor visible light communication system, and belongs to the visible light wireless communication technology field. The method includes the steps: utilizing a ceiling bounce model to calculate a VLC channel impulse response, carrying out photoelectric conversion for a visible light power signal received by a receiving end, and sending a sequence to a neutral network channel equalizer after amplification sampling; utilizing a heredity algorithm to optimize initialization weights and thresholds among neurons, establishing a neural network for training, and minimizing an error function; and judging an output, restoring the sent sequence, and finally achieving an equalization purpose. According to the scheme, interference among codes is obvious minimized, an error rate is reduced, the communication quality is further improved, a transmission rate that the system can reach is increased, the training time is shortened, and the system complexity is reduced.

The invention discloses a short-term load forecasting method, and belongs to the technical field of power systems. Aiming at the problems that mass data have too much noise, are long in training time and are easily trapped in local minimum or over-fitting and the like in the prior art, the method comprises: performing cluster analysis on historical load data to generate a typical load curve, and digging the generality of mass historical load data to achieve the effect of screening and training data for later load forecasting, thus eliminating the noise influence of the mass data; performing strong fitting of a complex nonlinear function by deep learning to solve the problems of over-fitting and local minimum of a conventional neural network, thus realizing accurate and quick short-term load forecasting; and further constructing a forecasting model by using a distributed memory computation framework Spark, thus improving the efficiency and the instantaneity of the whole short-term load forecasting flow. Compared with the prior art, the method can realize accurate and quick short-term load forecasting.

The invention relates to the technical field of exercise fitness equipment, in particular to an interactive fitness system and method based on virtual scenes. The system comprises fitness equipment, aKinect sensor, a large-screen displayer, a computer host, a loudspeaker and a fitness management system. Fitness users can select different instruments to be combined with the virtual scenes, and byintroducing action evaluation and correction and feedback of multiple aspects of vision, hearing and tactile sense, fitness actions are corrected and standardized in time, so that the effect of quickly developing standard fitness is achieved. The fitness actions of the users are evaluated in real time, action trigger thresholds or program types are self-adaptively changed according to completion effects so as to adjust the training difficulty coefficient, and meanwhile, the training difficulty can be adjusted by adjusting the training frequency and training time and selecting different instruments. When the system detects that the users complete a certain fitness action completely correctly, the number of training is automatically increased by one until the users start a next training program after the users complete training in a targeted number set by the training program, and then the training rhythm is actively controlled.

The invention discloses a method and device for recognizing text in an image, electronic equipment, and a computer readable storage medium. The scheme performs end-to-end recognition of text in an image through a network model of multi-layer superposition. The scheme includes: a space separable convolution operation of the image is carried out layer by layer through a multi-layer mode; convolutionfeatures extracted from the space separable convolution operation are fused to low layers mapped by layer superposition; the low layers are mapped to high layers of output convolution features; Obtaining global features from the bottom layer of the performing spatially separable convolution operation; Candidate region detection and region filtering parameter prediction of text in the image are carried out through global features, and pooling features corresponding to the detected text regions are obtained. The pooling feature is propagated backward to the recognition branch network layer thatperforms the character recognition operation, and the character sequence marked by the text area is output by the recognition branch network layer. The scheme saves the training time of the model andimproves the recognition accuracy.

The invention, which belongs to the technical field of mobile robot navigation, provides a mobile robot obstacle avoidance method based on a DoubleDQN network and deep reinforcement learning so that problems of long response delay, long needed training time, and low success rate of obstacle avoidance based on the existing deep reinforcement learning obstacle avoidance method can be solved. Specialdecision action space and a reward function are designed; mobile robot trajectory data collection and Double DQN network training are performed in parallel at two threads, so that the training efficiency is improved effectively and a problem of long training time needed by the existing deep reinforcement learning obstacle avoidance method is solved. According to the invention, unbiased estimationof an action value is carried out by using the Double DQN network, so that a problem of falling into local optimum is solved and problems of low success rate and high response delay of the existing deep reinforcement learning obstacle avoidance method are solved. Compared with the prior art, the mobile robot obstacle avoidance method has the following advantages: the network training time is shortened to be below 20% of the time in the prior art; and the 100% of obstacle avoidance success rate is kept. The mobile robot obstacle avoidance method can be applied to the technical field of mobilerobot navigation.

The invention discloses a neural network model training method, device and system and belongs to the image processing field. The method comprises the following steps that: medical samples sent by a plurality of clients are received, wherein the medical sample sent by the first client comprises a plurality of CT images and first label images corresponding to each CT image, wherein the first label images are used for identifying specified organs contained in the CT images, the first client adopts a local neural network model to segment the plurality of CT images so as to form the first label images, and the first client is any one client in the plurality of clients; and a first neural network model in a server is trained according to the medical samples sent by the plurality of clients, wherein the first neural network model is a neural network model of the latest version in the server. With the neural network model training method, device and system of the invention adopted, the training time of the neural network model is shortened, and the accuracy of the training of the neural network model is effectively improved. The neural network model training method, device and system of the present invention are used for training neural network models.

The invention discloses linguistic model training method and system based on distributed neural networks. The method comprises the following steps: splitting a large vocabulary into a plurality of small vocabularies; corresponding each small vocabulary to a neural network linguistic model, each neural network linguistic model having the same number of input dimensions and being subjected to the first training independently; merging output vectors of each neural network linguistic model and performing the second training; obtaining a normalized neural network linguistic model. The system comprises an input module, a first training module, a second training model and an output model. According to the method, a plurality of neural networks are applied to training and learning different vocabularies, in this way, learning ability of the neural networks is fully used, learning and training time of the large vocabularies is greatly reduced; besides, outputs of the large vocabularies are normalized to realize normalization and sharing of the plurality of neural networks, so that NNLM can learn information as much as possible, and the accuracy of relevant application services, such as large-scale voice identification and machine translation, is improved.

The embodiment of the invention provides an image foreground and background segmentation network model training method, an image foreground and background segmentation method, an image processing method and device, and a terminal device. The image foreground and background segmentation network model training method comprises the following steps: obtaining the eigenvectors of a sample image to be trained; performing convolution on the eigenvectors to obtain the convolution results of the eigenvectors; magnifying the convolution results of the eigenvectors; determining whether the magnified convolution results of the eigenvectors satisfy the convergence conditions or not; if so, completing the training of the convolutional neural network model used for segmenting the foreground and background of the image; and if not, adjusting the parameters of the convolutional neural network model according to the convolution results of the amplified eigenvectors and performing iteration training on the convolutional neural network model according to adjusted parameters of the convolutional neural network model until the convolution results satisfy the convergence conditions. By means of the image foreground and background segmentation network model training method, the training efficiency of the convolutional neural network model is improved, and the training time is shortened.

The invention discloses a method for identifying and locating fruit to be picked based on depth target detection, comprising the following steps: image acquisition, image annotation. data Set Preparation. Feature extraction includes using depth convolution neural network CNN to extract the fruit features in the image first; VGG16 is used as a convolution neural network for feature extraction. Model training, For all the training samples, the feature maps on each convolution layer are extracted firstly, and the prediction boxes are generated for the extracted feature maps, the categories of thetargets in the prediction boxes are calculated, the distance between the prediction boxes and the real annotation boxes is calculated, and the joint category loss and the prediction box offset loss are used as the target loss function of the training during the training process. A training is completed based on the complete one-time calculation process described above, and the training ends whenthe number of training times reaches a predetermined threshold or the loss is less than the predetermined threshold.

A computer-implemented system and method for assessing the editorial quality of a textual unit (document, paragraph or sentence) is provided. The method includes generating a plurality of training-time feature vectors by automatically extracting features from first and last versions of training documents. The method also includes training a machine-learned classifier based on the plurality of training-time feature vectors. A run-time feature vector is generated for the textual unit to be assessed by automatically extracting features from the textual unit. The run-time feature vector is evaluated using the machine-learned classifier to provide an assessment of the editorial quality of the textual unit.