279results about How to "Effective study" patented technology

Provided are a training method and device and an identification device for intent recognition. The method includes: acquiring a corpus text vector corresponding to a corpus in a corpus text; constructing a joint loss function formula of a training model; acquiring training data; performing segmentation processing on the training data and then mapping to the corresponding corpus text vector, recorded as a training vector; using the training model to predict a training sample vector, calculating a loss function value of each training model according to a prediction result; calculating the jointloss function value based on the loss function value of each model, judging whether the joint loss function value is less than a set threshold, if yes, ending the training, if not, updating each modelparameter, and continuing iterative training. Therefore, the generalization capability of intent recognition can be improved as much as possible, and semantic ambiguity and fault tolerance problems can be solved.

The invention discloses a human body behavior recognition method and system based on a graph convolution network, and the method comprises the steps: extracting human body skeleton information from animage containing human body behaviors, obtaining a human body joint point position information sequence, and constructing a topological graph sequence with any length of a human body skeleton; performing feature extraction and topological structure adaptive evolution on the topological graph sequence through a topological learnable graph convolution-based space-time graph convolution network to obtain new node features fused with local space-time features and a topological graph sequence with a new topological structure; performing feature extraction through a graph convolution long-term andshort-term memory neural network; global spatio-temporal features are obtained through global pooling operation; and performing human body behavior recognition based on the global spatial-temporal features through a classifier. The features of a whole graph are directly learned, the weight matrix in graph convolution is expanded to the whole topological graph structure, the relation between any two nodes in the graph is learned, limitation of the topological structure is avoided, and the recognition accuracy is high.

The invention discloses a license plate super-resolution processing method and system based on deep learning. The method includes the following steps: acquiring a series of images containing license plate information from an original compressed surveillance video; tracking a target of interest in the images; selecting a plurality of points of interest, intercepting images with the points of interest as the center, and registering the images; and using a deep learning training base to get a corresponding deep network weight, and carrying out super-resolution processing on the multiple frames of images after registration with use of the deep network weight to get a clear high-definition license plate I. The intrinsic relationship between degraded characters and clear characters in a highly compressed surveillance video is learned through joint optimization using a super-resolution deep network model and a gradient guide network model. Super resolution and a de-blocking effect are achieved through a common deep network. The resolution of degraded character is improved, and the block effect is removed. The problem that the license plate characters in a highly compressed surveillance video cannot be seen clearly is solved.

The invention provides an image and video enhancement method based on a multi-branch convolutional neural network, which comprises the following steps: inputting a single image or video sequence of low quality to stably solve the enhanced image or video. A novel multi-branch convolutional neural network structure can effectively solve the problem of image or video quality degradation caused by insufficient illumination, noise and other factors. A novel training loss function can effectively improve the accuracy and stability of the neural network. One application of the invention is unmanned vehicle (aircraft) driving. The principle of the invention is to process and enhance the image quality degradation caused by the surrounding environment change or interference of the video sensor, thereby providing higher quality image and video information for the decision-making system, thereby facilitating the decision-making system to make more accurate and correct decisions. The invention canalso be widely used in the fields of video communication, automatic navigation, video monitoring, short video entertainment, social media, image restoration and the like.

Methods for performing a dynamic analysis of complex systems are described. The dynamic analysis of complex systems is performed by sub-structuring of the system into a plurality of sub-structures. The sub-structures and the interfaces between the sub-structures are modelled using a generalised eigenvector analysis of a discretised model of a full system or a subassembly thereof comprising at least two sub-structures. The sub-structures then are combined at the interfaces into a combined system. The combined system then is solved. Optionally, selected sub-structures may be reduced. Performing such a dynamic analysis allows studying characteristics such as e.g. vibration and/or acoustical effects in a computational efficient way. Such methods can be applied for optimising designs of such complex systems.

The invention discloses a space target image restoration method based on a convolutional auto-encoding convolutional neural network. The method comprises the steps of establishing degraded images withdifferent degradation degree as input data, wherein the input data is used for learning and establishing a relatively robust CAE neural network model; and establishing simulation image sets of different types and different blurring degree through utilization of the prior knowledge with limited space target quantity and known part of models, and the feature that the similarity among the models ishigh, wherein the simulation image sets are used for carrying out training in the convolutional network. The method has the advantages that the images with clear edge structures can be restored; the method has excellent turbulence blurring removing capacity; the edge contrast of the restored images is high; the anti-noise capacity is excellent; the internal structures of the restored images are displayed relatively stably and clearly; and the efficiency is relatively high.

The invention discloses a universal remote control method, which is characterized by comprising the following steps of: identifying infrared remote control signals of each piece of equipment to acquire primary data, compressing infrared codes in the primary data by using a compression method, storing the compressed infrared codes in a primary data sheet, and when the infrared codes are used, performing decompression reduction to control given infrared remote control equipment to work. By the compression method used by the universal remote control method, a data volume can be reduced to be within 1K by compression, and the highest compression ratio of 16:1 can be realized, thereby reducing a storage space required by the learning of a remote controller code to further make possible all remote controller code learning of a remote controller.

The invention discloses a vehicle automatic classification method and a related device. Image appearance expansion is performed according to the vehicle image sample and/or the vehicle model to obtainthe vehicle sample set, so that the diversity and integrity of training samples are enhanced, the accuracy of subsequent vehicle target detection and classification is further improved, and the technical problem of low accuracy of a vehicle detection method in the prior art is solved.

The invention discloses a real-time information pushing method based on moving learning environment. The method comprises that to-be-pushed information is edited through a terminal by a teacher, and the information is clicked to send after editing; the to-be-pushed information finds a corresponding student moving terminal through an information pushing server, the terminal analyzes storage according to a curriculum name after receiving the information; and the information is judged that whether the information contains an attachment or not, and if the information contains the attachment, the attachment is automatically downloaded and stored in a file under the curriculum name. The method has the advantages of providing solution for to-be-arisen moving study and avoiding inconvenience of e-mail sending and limitation of capacity of text messages. The pushed information can carry information with large capacity, so that the teacher can push important information such as courseware, video or test questions to the students. The pushed information is managed and displayed in classification by the moving terminal, and automatic download of the attachment and storage in classification can be achieved by the moving terminal.

The invention discloses a compression artifacts removing method of an image based on deep learning. Artifacts generated in highly compression of the image can be effectively removed. The method is characterized in that firstly, by use of the newest deep learning technology, a deep residual error network is used as a basic module to be applied to a network model, so a gradient diffusion problem of the depth network model is effectively relieved, and meanwhile, base layer characteristics and high layer characteristics obtained in network learning are fused through skipping connection, so quite enriched characteristic information is provided for reconstruction of an artifacts removing image and the artifacts removing performance of the model is further improved; and secondly, the invention further provides a model selection scheme, so artifacts removing operation can be properly selected for proper models with compression artifacts in different degrees. According to the invention, two sets of public data sets are tested, so the performance of the provided method is remarkably improved than that of the current best artifacts removing algorithm.

The invention belongs to the technical field of natural language treatment and entity leaning and relates to an entity instance leading method based on the machine learning. The method comprises the following steps of: carrying out linguistic data marking after document pretreatment, selecting various feathers including word feathers, word class feathers and combination feathers of words and word class feathers, and converting the linguistic data and texts to be identified into feature vector modes; carrying out the maximum entropy model training, and obtaining the maximum entropy classifier through the marked linguistic data maximum entropy model parameters; and carrying out instance extraction by the maximum entropy classifier. The method has the advantage that the entity instance can be fast and effectively leaned from a large number of texts.

The invention relates to English learning equipment which comprises a processor, a database, a display unit and a loudspeaker. The database is used for storing pairs of multiple audio files and multiple English word files. The processor is used for processing the audio files and the English word files stored in the database. The display unit is used for displaying the English word files stored in the database. The loudspeaker is used for playing the audio files stored in the database. An audio acquisition unit is used for acquiring audio data when the processor receives an audio acquisition instruction and transmitting the acquired audio data to an audio processing unit. The audio processing unit is used for carrying out voice recognition on the audio data transmitted from the audio acquisition unit so as to generate text data and transmitting the text data to the processor. The processor is used for comparing the text data with the pairs of currently processed audio files and English word files when receiving the text data and controlling the display unit to display according to a comparison result.

The invention discloses a universal circuit breaker accessory fault diagnosis method based on deep learning. The method is used for fault diagnosis of a low-voltage universal circuit breaker switch-on-off accessory. By considering the feature of the switch-on-off coil current signal, a self-adaptive one-dimensional deep convolution neural network is adopted, and a receptive field region is enlarged by setting a convolution kernel of the first layer of convolution layer of the model as the wide convolution; and then the self-adaptive feature extraction is performed on the current signal by utilizing a feature extraction layer; and finally, a Softmax classifier is used for outputting a fault diagnosis result. The fault diagnosis of the switch-on-off accessory shows that the same fault is effectively identified under different switch-on phase angles, and high fault recognition rate still can be maintained in a generalization experiment, and the influence on the fault diagnosis result by the switch-on phase angle change can be effectively overcome.

The invention discloses a sample class classification method of atom Laplacian regularization-based semi-supervised dictionary learning. The method comprises: S1, constructing an atom Laplacian regularization-based semi-supervised dictionary learning model according to training samples; S2, using a block coordinate descent algorithm to optimize various variables in the semi-supervised dictionary learning model until convergence occurs; and S3, linearly reconstructing label vectors of test samples according to dictionary atom labels of solving and sparse codes of the unlabeled samples, and selecting dimensions of largest elements in the label vectors to use the same as classes to which the same belong. According to the method, dictionary atoms are considered as anchor data of reconstructinga training sample set to construct a similarity matrix among the dictionary atoms, thus graph structure information which is more robust for anomalous samples can be obtained, thus the unlabeled samples are forced to more effectively participate in a dictionary learning process, and a learned dictionary is enabled to have better sparse representation ability and classification discriminating ability.

The present invention discloses a dialect language identification technology, a virtual reality teaching method and a virtual reality teaching system. The technology comprises the following steps: performing real-time extraction of accent features through an accent feature extraction module; identifying areas where collected persons' accent belongs to and the collected persons' genders through an accent feature classification module according to the extracted accent features, and performing classification annotation of the collected persons' accent; and selecting a model for accent identification from an accent model base through an accent identification module according to the collected persons' accent features and the areas where the collected persons' accent belongs to and the collected persons' genders, and converting the collected speech data to standard mandarin in real time. The present invention further provides a virtual reality teaching method and a virtual reality teaching system. The identification precision and the identification system response speed are improved while solving the dialect problem.

The invention provides a method of extracting a feature vector from a system log, a method of constructing a system-log classification model and a system-log classification method. The method of extracting the feature vector includes: for each category, calculating semantic similarity degrees of keywords of the category and the system log, and selecting certain similarity degrees to use the same as a feature vector of the system log under the category; and combining the feature vectors of the system log under all the categories to obtain the feature vector of the system log. According to the method of constructing the model, the method of extracting the feature vector is utilized to extract feature vectors of a training data set, and the same are used as input of a convolutional neural network (CNN) to train the model. According to the system-log classification method, the method of extracting the feature vector is utilized to extract a feature vector of a system log, and a classification result is obtained through the model. The methods can realize system-log classification of high accuracy and high timeliness.

This invention is to automatically carry out the retracement of the knowledge. Therefore, this method is executed by a computer having a retracement data storage storing a target type of a past project, data concerning a scale of the past project, a specific phase of the past project, data concerning a problem in the specific phase of the past project, and data concerning an action against a problem in the specific phase of the past project. Then, this method comprises: obtaining project data including a target type of a project, data concerning a scale of the project, and a pertinent phase of the project; calculating an overall similarity for the retracement data of each past project, which is stored in the retracement data storage, by using a first similarity against the target type of the project, a second similarity against the data concerning the scale of the project, and a third similarity against the phase of the project; reading out, based on the overall similarity, from the retracement data storage, the data concerning the problem in the specific phase of the past project or the like.

The invention provides an increment study classification method under limited storage resources, which belongs to the technical field of pattern recognition. The method adopts a minimum distance classifier and comprises the following concrete steps: firstly, pre-classifying new samples, adding the new samples which are correctly pre-classified to the corresponding subset of the classifier, adding the new samples which are wrongly pre-classified to the set of wrong samples, and carrying out K-mean clustering on the samples in the set of wrong samples; then respectively selecting representative samples for each subset in the classifier and the set of wrong samples, adding the subsets in the set of wrong samples to the classifier after selection, and updating the classifier; and finally, adopting the updated classifier to classify the new samples. In the invention, by selecting the representative samples, not only the studied knowledge is saved, but also the new knowledge is acquired, and higher sample recognizing accuracy is achieved on the basis of reducing the storage cost and calculating the cost.