3147results about How to "Improve training efficiency" patented technology

The invention discloses a cascaded residual error neural network-based image denoising method. The method comprises the following steps of building a cascaded residual error neural network model, wherein the cascaded residual error neural network model is formed by connecting a plurality of residual error units in series, and each residual error unit comprises a plurality of convolutional layers, active layers after the convolutional layers and unit jump connection units; selecting a training set, and setting training parameters of the cascaded residual error neural network model; training the cascaded residual error neural network model by taking a minimized loss function as a target according to the cascaded residual error neural network model and the training parameters of the cascaded residual error neural network model to form an image denoising neural network model; and inputting a to-be-processed image to the image denoising neural network model, and outputting a denoised image. According to the cascaded residual error neural network-based image denoising method disclosed by the invention, the learning ability of the neural network is greatly enhanced, accurate mapping from noisy images to clean images is established, and real-time denoising can be realized.

Track and track / vehicle analyzers for determining geometric parameters of tracks, determining the relation of tracks to vehicles and trains, analyzing the parameters in real-time, and communicating corrective measures to various control mechanisms are provided. In one embodiment, the track analyzer includes a track detector and a computing device. In another embodiment, the track / vehicle analyzer includes a track detector, a vehicle detector, and a computing device. In other embodiments, the track / vehicle detector also includes a communications device for communicating with locomotive control computers in lead units, locomotive control computers in helper units, and a centralized control office. Additionally, methods for determining and communicating optimized control, lubrication, and steering strategies are provided. The analyzers improve operational safety and overall efficiency, including fuel efficiency, vehicle wheel wear, and track wear, in railroad systems.

The invention discloses a machine learning model training method and device. The method comprises the following steps: on the basis of initialized first weight and second weight of each sample in a training set, and with features of each sample as granularity, training a machine learning model; on the basis of prediction loss of each sample in the training set, determining a first sample set, where corresponding target variables are predicated inaccurately, and a second sample set, where corresponding target variables are predicated accurately; on the basis of the prediction loss of each sample in the first sample set and the corresponding first weight, determining overall prediction loss of the first sample set; on the basis of the overall prediction loss of the first sample set, improving the first weight and the second weight of each sample in the first sample set; and inputting the updated second weight of each sample in the training set and features of each sample and the target variables into the machine learning model, and with the features of each sample as granularity, training the machine learning model. Through the machine learning model training method and device, prediction accuracy and training efficiency of the machine learning model can be improved.

The invention discloses a sample data processing method and device and a computer readable storage medium. The method comprises the steps that a classification model is acquired; automatic annotationof model retraining sample data is performed through the classification model; the model retraining sample data and an automatic annotation result are displayed, the selection of correcting the displayed automatic annotation result is received, and an annotation expansion result of the model retraining sample data is obtained; and the model retraining sample data and the annotation expansion result are fed back to the model for training till the classification model no longer obtains classification performance improvement and iterative optimization of the annotation result corresponding to themodel retraining sample data is performed, and the error correction and leakage finding process of the corresponding annotation result is completed. In this way, annotation accuracy is improved, thesample data is continuously expanded, the corresponding automatic annotation result is obtained after automatic annotation through the classification model, the more precise annotation expansion result is obtained through correction, and the sample data in a large scale and large-quantity wide-dimensional annotation performed on the sample data are obtained.

The invention discloses a personalized recommendation method based on deep learning. The method comprises the steps of according to the viewing time sequence behavior sequence of the user, predictingthe next movie that the user will watch, including three stages of preprocessing the historical behavior characteristic data of the user watching the movie, modeling a personalized recommendation model, and performing model training and testing by using the user time sequence behavior characteristic sequence; at the historical behavior characteristic data preprocessing stage when the user watchesthe movie, using the implicit feedback of interaction between the user and the movie to sort the interaction data of each user and the movie according to the timestamp, and obtaining a corresponding movie watching time sequence; and then encoding and representing the movie data,wherein the personalized recommendation model modeling comprises the embedded layer design, the one-dimensional convolutional network layer design, a self-attention mechanism, a classification output layer and the loss function design. According to the method, the one-dimensional convolutional neural network technologyand the self-attention mechanism are combined, so that the training efficiency is higher, and the number of parameters is relatively small.

The invention relates to an intelligent training system for a vehicle driver based on an actual vehicle. The intelligent training system comprises a vehicle-mounted intelligent training subsystem, a data processing subsystem, an operation management and control subsystem and a system-level data communication subsystem, wherein the vehicle-mounted intelligent training subsystem is mounted on the actual vehicle and is used for training interactive driving of the driver and acquiring and reporting data of a driver training process; the data processing subsystem manages basic information and is used for receiving, storing and processing the data of the training process acquired and reported by the vehicle-mounted intelligent training subsystem; the operation management and control subsystem monitors an operated state and a running state of the vehicle in real time and remotely controls the monitored vehicle according to actual conditions; the vehicle-mounted intelligent training subsystem is connected with the data processing subsystem and the operation management and control subsystem through the system-level data communication subsystem and realizes bidirectional information interaction.

A muscular energy state analysis system and method for a swing motion and a computer program product thereof are provided. The system includes: a swing implement, for a user to perform the swing motion, and including an acceleration sensor for sensing acceleration of the swing implement when the swing implement is swung, so as to generate a swing speed data; a plurality of signal detection modules, for detecting electromyographic (EMG) signals generated by a muscles of the user; a database, for storing a muscular energy sample value; a muscular energy analysis module, for analyzing the EMG signals and the swing speed data so as to obtain a plurality of muscular performance values; and a comparison module, for comparing the swing speed data and the muscular performance values with the at least one muscular energy sample value in the database, so as to generate a comparison result data.

The invention relates to a dual-channel convolutional neural network (DC-CNN)-based spectral-spatial cooperative classification method for hyperspectral images. For a characteristic that hyperspectral image data adopts a three-dimensional structure, spectral-spatial features are extracted in a mode of combining a one-dimensional convolutional neural network (1D-CNN) channel with a two-dimensional convolutional neural network (2D-CNN) channel to finish the spectral-spatial cooperative classification of the hyperspectral images. For the problem of relatively little artificial mark data of the hyperspectral images, by adopting a data expansion method suitable for the hyperspectral images, the scale of training samples is increased, the training efficiency of a CNN is improved, and the over-fitting problem is reduced.

The invention discloses a lower limbs rehabilitation training robot, which comprises an exoskeletal mechanical structure and a control system independent of the mechanical structure, wherein the exoskeletal mechanical structure comprises thigh mechanisms and shank mechanisms, which are bilaterally symmetrical and connected with a waist mechanism in turn; imitation human hip joints are arranged at the connection part of the waist mechanism and the thigh mechanisms; imitation knee joints are arranged between the connection parts of the thigh mechanisms and the shank mechanisms; the waist mechanism comprises a waist part connecting plate and flexible connecting plates which are connected with the waist part connecting plate in a bilateral symmetrical manner; the waist mechanism, the left and right thigh mechanisms and the left and right shank mechanisms are provided with flexible connecting bands for connecting the corresponding parts of a human body with the robot; the control system is independent of the mechanism of the robot and is connected with the robot body through data wires for controlling rehabilitation training; and the control system comprises hardware and software, the soft ware comprises a user module, a real-time control module and a rehabilitation effect evaluation module.

The invention relates to a transformer substation field operation management and control system and method based on three-dimensional dynamic modeling. The method comprises the following steps: performing three-dimensional dynamic modeling on a transformer substation; accurate positioning; based on the three-dimensional digital model, carrying out intelligent management and control on an operation site through intelligent identification and early warning; based on a three-dimensional digital model, performing live-action rehearsal and equipment simulation training on typical production operation and safety accidents through a virtual reality system. According to the invention, visual data support is provided for field operation work and personnel training work; real-time positioning and track checking of personnel and equipment are realized, and visual technical support is provided for remote intelligent management and control; an intelligent management and control means is provided for field operation safety, dynamic simulation of an operation task plan, typical operation, dynamic rehearsal of safety accidents and disassembly and reassembly analogue simulation of typical equipment are achieved, a visual simulation scheme is provided for field operation, operation risks are reduced, immersive training experience is provided, and training efficiency is improved.

The invention provides a vehicle license plate recognition method, and relates to a method for image recognition. The method comprises the following steps of preprocessing an image; partitioning a vehicle region according to color and texture characteristics; extracting a remarkable factor graph of a vehicle region diagram; extracting candidate vehicle license plates by an Adaboost classifier based on expanded Harr-like characteristic; determining the position of a real vehicle license plate from the candidate vehicle license plates; partitioning the marked vehicle license plate from the corresponding vehicle region original diagram; carrying out character segmentation according to structural characteristic; and carrying out character recognition based on the improved template matching method. With the method provided by the invention, the defects that the application scene of the traditional vehicle license plate recognition method is relatively single, some traditional vehicle license plate recognition methods are only suitably used for single vehicle license plate recognition of the single scene and difficultly used for multiple-vehicle license plate recognition of multiple scenes, and the recognition rate is easily affected by strong light, haze and weak light environments are overcome.

The invention provides a multi-task deep learning network-based training method, a multi-task deep learning network-based training system, a multi-task deep learning network-based identification method and a multi-task deep learning network-based identification system. The training method includes the following steps that: the face region of a face image in a training set is obtained; key point detection is performed on the face region, so that key feature point positions are obtained; affine transformation is performed on the face image according to the key feature positions, so that an aligned face image can be obtained; and the aligned face image is inputted into a multi-task deep learning network, so that training can be carried out, and therefore, a multi-task deep learning network model can be obtained. The identification method includes the following steps that: affine transformation is performed on a face image to be identified according to the key feature positions of the face image to be identified, so that an aligned face image can be obtained; the aligned face image is inputted into a trained multi-task deep learning network model, so that feature extraction can be carried out, and feature information can be obtained; and the feature information of the face image to be identified is matched with feature information corresponding to each face image in a registration set, so that identification results can be obtained. With the methods and systems adopted, the training and identification efficiency of the multi-task deep learning network can be improved.

The invention relates to a target detection system and method based on self-adaption combined wave filtering and multilevel detection. The system includes a target detection module combing a moving target detection unit with an obvious target detection unit, a target identification module based on a convolutional neural network and a target tracking module based on combination decision and multi-channel image features; the target detection module, the target identification module and the target tracking module tightly cooperate with one another to constitute the stable and reliable target detection system together. System output includes detected candidate target position information, target classification information and position information of selected targets obtained by target tracking. The target detection system is achieved on a high-performance multi-core DSP chip and used for carrying out targeted optimization on the multi-core DSP chip, real-time target detection and target tracking are achieved, and a function of quickly identifying a target is achieved. The target detection system and method have the advantages of being high in practicability and feasibility and convenient to integrate into various solution schemes with target detection demands, and can achieve intelligent target detection, identification and tracking.

An embodiment of the invention provides an intelligent question-answer method, device and system. The method includes: receiving question keywords sent by a client, and subjecting the query keywords to word segmentation processing to obtain one or more first terms; based on the first terms, acquiring a set of candidate recommended words matched with the query keywords, wherein the candidate recommended words are words containing the first terms in a preset database; calculating weight of the first terms; based on the weight of the first terms, calculating similarity of the query keywords and the candidate keywords; returning the candidate recommended words and corresponding answer information with the similarity conforming to preset rules to the client. By the method, device and system, result recommendation accuracy in the intelligent question and answer process can be increased.

The present application discloses a method for updating a classification model, comprising: obtaining incremental data within a predetermined period of time from a server providing the user behavior data as a training sample set; determining the number of newly added decision trees; According to the training sample set, the random forest algorithm is used to generate the newly added number of decision trees; the decision trees included in the classification model and the newly generated decision trees are sorted according to the classification effect, and the sequence position is selected from them. High-level, the predetermined number of decision trees; summarizing the selected decision trees to obtain an updated classification model. The present application also provides a device for updating a classification model. With the method provided in this application, since it is not necessary to perform training based on the full amount of data, but to use the incremental update method on the basis of the original classification model, it can improve the efficiency of model training and achieve rapid response to business.

The present application discloses a method for updating a classification model. The classification model includes a predetermined number of original decision trees for class prediction according to user behavior data in network applications. The method includes: providing the user Incremental data within a predetermined period of time is obtained from the behavior data server, and a training sample set is extracted from it; according to the training sample set, a new number of decision trees is generated; according to the historical classification accuracy and the training sample set For this classification accuracy rate, select the predetermined number of decision trees from the new decision tree and the original decision tree to form an updated classification model. The present application also provides a device for updating a classification model. The method provided in this application can improve the efficiency of model training and realize rapid response to business; and because the historical classification accuracy is introduced when evaluating the classification effect of the decision tree, it can smooth the short-term fluctuation of data.

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.

A beamforming weight training method, a base station and a terminal, relating to the field of communications. Disclosed is a beamforming weight training method. The method comprises: a base station sending an N-type beam pilot signal to a terminal; the terminal selecting one or more types of beam pilots in the received beam pilot signal, and feeding back beam information corresponding to the selected beam pilots to the base station; and the base station determining an initial training signal vector according to the beam information fed back by the terminal, and sending a channel vector to the terminal to perform beam weight training, N being a positive integer. Also disclosed are another beamforming weight training method, a base station, and a terminal. The technical solutions of the present application improve the efficiency of beamforming weight training.

The invention discloses a shopping behavior prediction method, and the method comprises the steps: selecting different target users at different shopping stages, and obtaining sample data from a user behavior log of the selected target users; respectively extracting a first feature set for marking user behavior from the sample data at each shopping stage; respectively training the first feature sets at different stages through a decision tree model, obtaining a feature combination through multiple iteration, and enabling the feature combination to serve as a second feature set; respectively employing the first and second feature sets at all shopping stages to train a pre-built machine learning model, wherein the machine learning model is used for predicting the shopping demand degree of a user; and determining the shopping stage of a to-be-detected user according to the machine learning model at different shopping stages. The invention also provides a corresponding shopping behavior prediction device.

The invention provides a facial emotion recognition method based on a deep sparse convolutional neural network. The method comprises the following steps: to begin with, carrying out emotion image preprocessing; then, carrying out emotion feature extraction; and finally, carrying out emotion feature identification and classification. The facial emotion recognition method based on the deep sparse convolutional neural network carries out optimization on weight of the deep sparse convolutional neural network through a Nesterov accelerated gradient descent algorithm to enable network structure to be optimal, thereby improving generalization of the face emotion recognition algorithm; since the NAGD has a precognition capability, the algorithm can be prevented from being too fast or too slow foreseeingly; and meanwhile, response capability of the algorithm can be enhanced, and better local optimum value can be obtained.

The invention discloses a medical anatomy auxiliary teaching system based on augmented reality, and relates to the field of biomedical engineering. The system consists of intelligent glasses, an anatomy instrument, a teacher control unit, a wireless router, and a medical image scanning device. A three-dimensional virtual model (vessels, muscle and skeleton) is built through the anatomy information obtained by the medical image scanning device and modeling software. The three-dimensional virtual model and a real anatomic site in an operation vision are integrated and stacked, thereby enabling students to see a subcutaneous internal structure. Meanwhile, the system can obtain the precise spatial information of the anatomy instrument relative to the anatomic site, achieves the augmented display of the operation vision, and has three teaching modes: discovery learning, interactive training, examination and evaluation. The system can assist students carry out theory learning before operation, interactive guide in operation and intelligent estimation after operation, enables the anatomy operation to be standard, improves the operation precision, and reduces the training cost.

The invention provides a training method of a posture recognition model and an image recognition method and device. The training method of the posture recognition model comprises: inputting a sample image marked with human body key points into a feature map model included in the posture recognition model, and outputting a feature map corresponding to the sample image; inputting the feature map into a two-dimensional model included in the posture recognition model, and outputting two-dimensional key point parameters for representing a two-dimensional human body posture; inputting a target humanbody feature map cut from the feature map and the two-dimensional key point information into a three-dimensional model included in the posture recognition model, and outputting three-dimensional posture parameters for representing a three-dimensional human body posture; constructing a target loss function by combining the two-dimensional key point parameters and the three-dimensional attitude parameters; and based on the target loss function, updating model parameters of the attitude recognition model.

The invention provides a video classification method, a model training method and device thereof, and electronic equipment. The training method comprises the following steps: extracting initial features of a plurality of video frames through a convolutional neural network; extracting final features of the plurality of video frames from the initial features through a recurrent neural network; inputting the final feature into an output network, and outputting a prediction result of the multi-frame video frame; determining a loss value of the prediction result through a preset loss prediction function; and training the initial model according to the loss value until parameters in the initial model converge to obtain a video classification model. According to the method, the convolutional neural network and the recurrent neural network are combined, so that the operand can be greatly reduced, and the model training and recognition efficiency is improved; and meanwhile, the association information between the video frames can be considered in the feature extraction process, so that the extracted features can accurately represent the video types, and the accuracy of video classificationis improved.

The invention provides a remote sensing image target detection method based on a new frame regression loss function. The method comprises the following steps: training a candidate region generation network through employing a high-resolution remote sensing image as a training sample, and enabling the frame regression loss function of the candidate region generation network to employ the new loss function; obtaining a candidate target frame as a target initial position training region detection network through the trained candidate region generation network, wherein a new frame regression lossfunction is adopted as a frame regression loss function of the region detection network; alternately training a candidate region generation network and a region detection network; and sharing backbonenetworks of the candidate region generation network and the region detection network, combining the trained candidate region generation network and the region detection network to construct a detection model, and obtaining the position and the category of the interested target of the to-be-detected high-resolution remote sensing image. By improving the frame regression loss function of target detection, the target detection precision of the high-resolution remote sensing image can be effectively improved.

The invention provides a seven-degree-of-freedom exoskeletal rehabilitation robot for upper limbs. The seven-degree-of-freedom exoskeletal rehabilitation robot for upper limbs comprises a front arm and wrist joint movement mechanism, an elbow joint movement mechanism, an upper arm movement mechanism, a gravity compensation movement mechanism, a shoulder joint movement mechanisms and a chair mechanism which are sequentially connected. The seven-degree-of-freedom exoskeletal rehabilitation robot for upper limbs is driven and controlled by a motor to help a hemiplegic patient do active and passive training for upper affected limbs, including bending / stretching of wrist joints, adduction / abduction movement of the wrist joints, inward / outward rotating movement of the front arms, bending / stretching of elbow joints, bending / stretching of shoulder joints, inward / outward rotating movement, and outward swinging / adduction movement.

The present invention discloses a driving model training method, a driver identification method, a driving model apparatus, a driver identification apparatus, a device, and a medium. The driving modeltraining method includes the following steps that: the training behavior data of a user are acquired, wherein the training behavior data are associated with a user identifier; training driving data associated with the user identifier are obtained on the basis of the training behavior data; positive and negative samples are obtained from the training driving data on the basis of the user identifier, and the positive and negative samples are divided into a training set and a test set; the training set is trained by using a bagging algorithm, so that an original driving model can be obtained; and the test set is adopted to test the original driving model, so that a target driving model can be obtained. With the driving model training method adopted, the generalization of the driving model can be effectively enhanced; the problem of poor recognition results of current driving recognition models can be solved; and the accuracy of identifying the driving of drivers is improved.

The invention provides a method and a system for prediction of advertisement clicking rate to solve a problem of accuracy of the clicking rate prediction affected by seriously imbalanced sample data in the original sample set. The method comprises: extracting sample data to construct an original sample set, wherein the sample data comprises clicked data and unclicked data of users; constructing a training sample set by carrying out sampling on the original sample set; constructing a prediction model by using the sample data in the training sample set as model parameters; and predicting the user clicking rate of each kind of advertisement by utilizing the prediction model to predict the testing sample set. The method and the system in the invention can eliminate the problem of serious imbalance of the proportion between the clicked data and the unclicked data in the original sample set and construct a relatively balanced training sample set. The method and the system improve recognition rate of the prediction model on the clicked rate and improve the accuracy of clicking rate prediction.

The invention discloses a convolutional neural network medical CT image denoising method based on residual error learning. The method comprises the following specific steps: 1) constructing a medicalCT image model; 2) constructing a neural network model; 3) training the network; 4) updating parameters; and 5) denoising the medical CT image: inputting the medical CT image containing the noise intothe constructed network model, and outputting the medical CT image without the noise through the network. The method has the advantages that medical CT image denoising is carried out by combining knowledge in the aspect of the convolutional neural network in deep learning. Noise in the image is approximately learned in a residual error learning mode, good pertinence is achieved, and meanwhile thetraining efficiency of the neural network is improved. The convolutional neural network and the residual error learning method are adopted, so that the feature information in the image can be betterlearned, and more local image information can be reserved in the image denoising process. And meanwhile, the image denoising capability is also improved.

The invention provides a human face identification method and apparatus, computer equipment and a readable storage medium. The method comprises the following steps: any human face image in a trainingset is input into a depth residual error network module group for training operation, a human force identification model is built; human face identification characteristics with specified dimensions are output, corresponding categories of multiple tasks for a corresponding human face image are predicted according to the human face identification characteristics with the specified dimensions basedon a full link layer classifier; a training loss value is calculated according to the predicted corresponding categories of multiple tasks for the corresponding human face image based on a loss function; whether to input any human face image in the training set into the depth residual error network module group for the training operation is determined according to the training loss value, and thehuman force identification model is built. Via the human face identification method and apparatus, the computer equipment and the readable storage medium disclosed in a technical solution of the invention, in-depth of training and construction of the human face identification model can be realized; multiple tasks can be subjected to the training operation, the human face identification model obtained based on the training operation is high in identification effects, multiple tasks of human face identification can be performed, and identification efficiency can be improved.

The invention discloses a rolling bearing probabilistic intelligent fault diagnosis method based on adaptive MRVM. The method comprises the steps that the original fault data of a rolling bearing are measured through an acceleration sensor; a vibration signal is segmented, and wavelet packet energy characteristics are extracted; principal component analysis and dimension reduction are used for normalization simultaneously; a training sample set and a test sample set are processed and divided; an algorithm is used to adaptively select nuclear parameters; the training sample set is used to train and test a multi-class correlation vector machine; and the test result is compared with the actual fault type to acquire the validity of a diagnosis model. According to the invention, the method overcomes the defect that a traditional intelligent fault diagnosis method cannot output the fault probability value; the fault diagnosis accuracy of the rolling bearing is improved; more fault type determining information of the rolling bearing can be provided; through the fault type probability value provided by the invention, the state of the rolling bearing can be further assessed; and method has the advantages of good engineering value and application prospect.