535 results about "Recurrent neural network model" patented technology

The invention discloses a video description method based on deep learning and text summarization. The video description method comprises the following steps: through a traditional image data set, training a convolutional neural network model according to an image classification task; extracting a video frame sequence of a video, utilizing the convolutional neural network model to extract convolutional neural network characteristics to form a <video frame sequence, text description sequence> pair which is used as the input of a recurrent neural network model, and training to obtain the recurrent neural network model; describing the video frame sequence of the video to be described through the recurrent neural network model obtained by training to obtain description sequences; and through a method that graph-based vocabulary centrality is used as the significance of the text summarization, sorting the description sequences, and outputting a final description result of the video. An event which happens in one video and object attributes associated with the event are described through a natural language so as to achieve a purpose that video contents are described and summarized.

The invention discloses a method and a system for enhancing features by the aid of bidirectional long-term and short-term memory recurrent neural networks. The method includes steps of training inputted features with noise and corresponding features without noise by the aid of bidirectional long-term and short-term memory recurrent neural network models; enhancing the features by the aid of the bidirectional long-term and short-term memory recurrent neural network models. The method and the system have the advantages that the long-term and short-term memory recurrent neural network models can be built in two directions by the aid of the inputted features, and accordingly models of contexts of current frames can be effectively built; long-term and short-term memory cells and certain control variables are introduced into the neural networks, and accordingly the models can be built in the dependence on the long contexts; excellent enhancement effects can be realized by the models owing to long-term dependence when the features are enhanced by the aid of the models, and accordingly the performance of voice recognition systems and audio event classification systems can be improved.

Embodiments of the invention provide an anomaly detection method and device and electronic equipment, and belong to the technical field of internet. The method comprises the following steps of: obtaining a current real value, at a current moment, of a to-be-monitored index; obtaining a current predicted value, at the current moment, of the to-be-monitored index according to a pre-established timerecurrent neural network model and the current moment, wherein the time recurrent neural network model is established according to a history real value of the to-be-monitored index; taking a difference value between the current real value and the current predicted value as a current predicted error, and judging whether the current predicted error is in a preset threshold value range or not; and ifthe current predicted error is not in the preset threshold value range, determining that the to-be-monitored index is abnormal at the current moment. According to the method, the to-be-monitored index is predicted through the time recurrent neural network model, and the obtained current predicted value at the current moment is compared with the current real value, so that the self-adaptability ofanomaly detection is improved and then the correctness of anomaly detection is improved.

The invention relates to a text key information identification method, an electronic apparatus and a readable storage medium. The method comprises the steps of, after a to-be-identified text is received, performing segmentation processing on the received to-be-identified text by using a predetermined word segmentation model to obtain segmented words of the to-be-identified text, wherein the predetermined word segmentation model is a long and short term memory recurrent neural network model obtained by training a preset quantity of sample statements labeled by adopting a sequence labeling method; and based on word frequencies, positions and word spans of the segmented words in the to-be-identified text, and according to a preset scoring formula, performing calculation to obtain scores of the segmented words, sorting the segmented words in the to-be-identified text according to a sequence of the scores from high to low, extracting the segmented words sorted in front to serve as keywords,and according to the extracted keywords, obtaining key information of the to-be-identified text. A user can be enabled to quickly and accurately obtain the key information in the to-be-identified text.

The invention discloses a behavior identification method based on a recurrent neural network and human skeleton movement sequences. The method comprises the following steps of normalizing node coordinates of extracted human skeleton posture sequences to eliminate influence of absolute space positions, where a human body is located, on an identification process; filtering the skeleton node coordinates through a simple smoothing filter to improve the signal to noise ratio; sending the smoothed data into the hierarchic bidirectional recurrent neural network for deep characteristic extraction and identification. Meanwhile, the invention provides a hierarchic unidirectional recurrent neural network model for coping with practical real-time online analysis requirements. The behavior identification method based on the recurrent neural network and the human skeleton movement sequences has the advantages of designing an end-to-end analyzing mode according to the structural characteristics and the motion relativity of human body, achieving high-precision identification and meanwhile avoiding complex computation, thereby being applicable to practical application. The behavior identification method based on the recurrent neural network and the human skeleton movement sequence is significant to the fields of intelligent video monitoring based on the depth camera technology, intelligent traffic management, smart city and the like.

The invention provides a specific target emotion classification method based on attention coding and a graph convolution network, and the method comprises the steps: obtaining a context and a hidden state vector corresponding to a specific target through a preset bidirectional recurrent neural network model, and carrying out the multi-head self-attention coding of the context and the hidden statevector; extracting a syntax vector in a syntax dependency tree corresponding to the context by combining a point-by-point convolution graph convolutional neural network, and performing multi-head self-attention coding on the syntax vector; then, multi-head interaction attention is used for carrying out interaction fusion on syntactic information codes, context semantic information codes, syntacticinformation codes and specific target semantic information codes; and splicing the fused result with the context semantic information code to obtain a final feature representation, and obtaining an emotion classification result of the specific target based on the feature representation. Compared with the prior art, the relation between the context and the syntax information and the relation between the specific target and the syntax information are fully considered, and the accuracy of sentiment classification is improved.

The invention is applicable to the technical field of identity authentication, and provides a voiceprint recognition method and device. The method comprises the steps of preprocessing input voices toobtain a valid voice in each voice; extracting MFCC acoustic features of each voice and outputting first and second feature matrices including the MFCC dimension and the voice framing number; constructing a short- and long-time recurrent neural network model, and taking the first feature matrix as the input; training feature extraction matrices by using training parameters of the neural network model and speaker characteristics of the voices, each feature extraction matrix corresponding to a speaker model; and selecting the speaker model matching the second feature matrix, with the speaker output corresponding to the matched speaker model as a voiceprint recognition result. The method and device in the invention adopt a supervised learning method to train a voiceprint background model andcan excavate more suitable acoustic features from the training voices so as to be able to distinguish the difference characteristics of speakers more accurately and learn a more robust speaker model for a better voiceprint recognition result.

The invention discloses a method and a device for speech recognition by use of a long-short term memory (LSTM) recurrent neural network model. The method comprises the following steps: receiving speech input data at the (t)th time; selecting the LSTM hidden layer states from the (t-1)th time to the (t-n)th time, wherein n is a positive integer; and generating an LSTM result of the (t)th time according to the selected at least one LSTM hidden layer state, the input data at the (t)th time and an LSTM recurrent neural network model. With the method and the device, the 'tail effect' of the deep neural network is well solved, and the accuracy of speech recognition is improved.

The invention discloses a generation method of an image description from a structured text. The generation method comprises the steps of downloading pictures from the internet to form a picture training set; conducting morphological analysis on descriptions which correspond to the pictures in the picture training set to form the structured text; using an existing neural network model to extract convolution neural network characteristics of the pictures in the training set, and using <, picture characteristics and structured text < as inputs to form a multitasking recognition model; using the structured text extracted from the training set and a description which corresponds to the structured text as inputs of a recurrent neural network, and conducting training to obtain a parameter of a recurrent neural network model; inputting the convolution neural network characteristics of an image ready to be described, and obtaining a predicted structured text through the multitasking recognition model; inputting the predicted structured text, and obtaining the image description through the recurrent neural network model. Compared with the prior art, a better image description effect, accuracy and sentence variety can be generated through the method, and the generation method of the image description from the structured text can be effectively popularized in an application of image retrieval.

The invention discloses a deep learning-based text keyword extraction method. The method comprises the following steps of: firstly training a recurrent neural network model, wherein the used training data comprise a large amount of texts and keywords thereof, and the training target is maximizing text-based condition probability of the keywords; converting each text and the keyword thereof into word vectors, inputting the word vectors into the recurrent neural network model and updating network parameters by using a random gradient descent method; and after the model training is finished, converting a section of text, the keyword of which is to be extracted, into a word vector, inputting the word vector into the trained recurrent neural network model so as to generate the keyword of the section of text. According to the method disclosed by the invention, the extraction of text keywords is realized by learning an end-to-end model through data driving; and compared with the traditional statistics and linguistics-based method, the method disclosed by the invention is stronger in adaptability, and can be used for obtaining different models according to different training data so as to extract keywords according to the requirements of specific fields.

The invention discloses a recurrent neural network-based commodity question answering expert recommendation method. The method comprises the following steps of: preprocessing data; screening candidateexperts for all the users corresponding to a question answering community; constructing an expert user file corresponding to each candidate expert; carrying out training on the basis of word vectorsof specific domain knowledges so as to obtain a word vector search table; carrying out feature representation learning on the basis of a recurrent neural network model and constructing a community question answering expert recommendation model; and determining an expert recommendation sequence corresponding to a new question in a to-be-processed question answering community. According to the method, grammar and semantic information of sentences can be effectively expressed, and high-level features of sentence level can be mined, so that manual intervention is decreased and automatic training and learning can be carried out.

The invention discloses a hyper-spectral image classification method based on recurrent neural network with the object to solving the problems that in prior art, the input characteristic determination ability is weak and that the extraction of local spatial characteristics is not complete. The method comprises the following steps: 1) extracting the spatial texture characteristics and the sparse representation characteristics of a hyper-spectral image and piling and combining them as the low-level characteristics; 2) extracting from the low-level characteristics the sample local spatial sequence characteristics; 3) according to the local spatial sequence characteristics, creating a recurrent neural network model; and utilizing the training sample local spatial sequence characteristics to train the recurrent neural network model parameters; and 4) inputting the testing sample local spatial sequence characteristics into the well-trained recurrent neural network model; obtaining the highly abstract high-level semantic characteristics and obtaining the classification information of the testing sample. According to the deep learning method of the invention, the correct efficiency for hyper-spectral image classification is increased and the method can be used for vegetation investigation, disaster monitoring, map making and intelligence obtaining.

The invention discloses a multiple attention and multiple scale-based image describing method. The method comprises the steps of selecting an image detecting model for extracting image features, dividing into a network training set, a verification set and a test set, extracting the image features, constructing an attention recurrent neural network model, training the attention recurrent neural network model and carrying out image description. According to the method disclosed by the invention, an image description generating network model formed by original image feature extracting, multi-attention multi-scale feature mapping, recurrent neural network residual connecting and recurrent neural network language decoding is constructed, so that the quality of the image description is improvedand the detail of the image description is enriched. By means of the method disclosed by the invention, a high quality image can be generated by adopting the neural network model to carry out description under the circumstance of only having an image.

The invention relates to a knowledge graph construction method based on deep learning. The method comprises the steps that a target text statement is given, and a two-way long short term memory recurrent neural network model and a conditional random field model are used to recognize target entities in the target text statement; a context sensitive two-way long short term memory recurrent neural network model and a feedforward neural network model are used to extract the relation between every two target entities; a vector space model is used to normalize the target entities, and the normalized target entities are mapped to a concept; and a knowledge graph is constructed according to the target entities, the relation between the target entities and the concept. According to the method, the deep learning technology is applied to construction of the knowledge graph, entity recognition models of a two-way recurrent neural network and a conditional random field are adopted to recognize the target entities in the target text statement, feature engineering in the entity recognition process and the relation extraction process is reduced, the burden and trouble brought by manual design and feature adjustment are relieved, and knowledge in a text is mined precisely.

The invention provides an automatic vehicle following method and system for simulating driver characteristics on the basis of LSTM. According to the automatic vehicle following method and system for simulating the driver characteristics on the basis of the LSTM, an LSTM recurrent neural network model is introduced; the model adopts sensor information time sequence data, automatic vehicle running time sequence data and the like, which are collected in the steady state vehicle following process of an excellent driver, so as to learn driver vehicle following driving behavior characteristics and establish a non-linear input and output mapping relation knowledge base; and therefore, longitudinal operation control of a vehicle in the vehicle following running process is predicated to realize automatic adaption of the system to the driver characteristics. According to the automatic vehicle following method and system for simulating the driver characteristics on the basis of the LSTM, the driver vehicle following behavior characteristics are simulated by using a characteristic that an LSTM recurrent neural network is good at processing time sequence characteristic data; the output of a designed controller conforms to driving behavior characteristics of a human being on the premise of meeting safety, accuracy and comfort; and meanwhile, self-learning of the driver operation process characteristics can be effectively realized; the self-adaption of the system to the driver characteristics is realized; and a generally applicable range is realized.

The invention discloses a single-channel real-time noise reduction method based on the convolutional recurrent neural network, a single-channel real-time noise reduction device based on the convolutional recurrent neural network, an electronic device and a storage medium, and belongs to the technical field of computers. The method comprises the following steps: extracting acoustic features from received single-channel voice signals, carrying out iterative operation on the acoustic features in a pretrained convolutional recurrent neural network model, thus the specific value film of the acoustic features is calculated, carrying out masking on the acoustic features by adopting the specific value film, and synthesizing the acoustic features after masking and the phase positions of the single-channel voice signals, thus obtaining voice signals. For the single-channel real-time noise reduction method based on the convolutional recurrent neural network and the single-channel real-time noisereduction device based on the convolutional recurrent neural network, the parameter number of the neural network can be reduced, the data storage amount and the demand for the system data bandwidth can be reduced, and the real-time property of the single-channel noise reduction is greatly improved while the good noise reduction property is realized.

The invention discloses an echo elimination method based on deep learning, an echo elimination device based on deep learning, electronic equipment and storage medium, and belongs to the technical field of a computer. The method comprises the following steps of extracting acoustic features from received microphone signals, wherein the microphone signals comprise near-end signals and remote-end signals; performing iterative operation on the acoustic features in a pretrained recursive neural network model with long short-term memory; calculating a ratio membrane of the acoustic features; using the ratio membrane for shielding the acoustic features; and synthesizing the shielded acoustic features with the phase of the microphone signals to obtain near end signals subjected to echo elimination.The echo elimination method based on deep learning and the echo elimination device based on deep learning can realize the echo elimination under the conditions of the background noise, double talking, nonlinear distortion and the like; the echo elimination effect is greatly improved; the application scenes are greatly increased; the use of a post-filter is not needed; the electronic equipment iseffectively simplified; and the electronic equipment cost is reduced.

The invention discloses a crowd abnormity detection and positioning system and a crowd abnormity detection and positioning method based on a time recurrent neural network. The method comprises the steps of carrying out gridding division on a scene on the basis of preprocessing acquired sample data, and dividing a video fragment consisting of n frames into a plurality of time-space blocks; then constructing a multi-scale optical flow histogram and Gabor wavelet texture features in each time-space block; vectoring and combining the features in the multiple time-space blocks, finding out a relation between a space dimension and a time dimension by taking the video fragment as a time sequence through respectively using hidden nodes and feedback nodes of the time recurrent neural network, and training a time recurrent neural network model capable of detecting a long dependency relation; finally detecting and positioning a crowd abnormity by the model. The method is higher in instantaneity and accuracy, and can detect abnormities activated by a small quantity of individuals or a large quantity of individuals.

The invention discloses a heart sound signal classification method based on a convolutional recurrent neural network. The method comprises the following steps: performing noise processing on heart sound data; extracting heart sound characteristics of the heart sound signals; standardizing the data; constructing a convolutional recurrent neural network model; training the constructed neural networkby using the training sample data characteristics, and storing the trained network structure and parameters; and testing the test sample data by using the trained model parameters to obtain a final classification and identification result. According to the invention, the system complexity is reduced; the extracted heart sound characteristics do not need to segment heart sound signals; according to the heart sound signal classification method based on the convolutional neural network, the convolutional neural network and the recurrent neural network are connected in series, a deep learning model with the processing advantages of the convolutional neural network and the recurrent neural network is provided, better expressive force is provided for heart sound signal classification, and an effective and convenient tool is provided for detection of normal and abnormal heart sound signals.

The invention relates to an electromechanical device neural network failure trend prediction method, comprising the following steps: (1) obtain a section continuous vibration signal which is sensitive to the failure and is output by a measuring point sensor; (2) respectively carry out exceptional value elimination and missing data filling to the vibration data by a 3 sigma method and an interpolation method; (3) carry out a normalization process to a vibration data sequence; (4) calculate a vibration data sequence which is entropy-weighted according to the sequence which is carried out the normalization process; (5) carry out a time-weighted calculation to the vibration data sequence which is entropy-weighted by utilizing time weight due to the influence of time factor; (6) build a nonlinear dynamic recurrent neural network prediction model by utilizing the data sequence which is obtained by step (5) and determine a hidden layer optimal node number by utilizing a golden section method; (7) carry out normalization process to a trend prediction result and obtain a actual prediction result. A dynamic recurrent neural network model is adopted to carry out prediction in the invention, therefore, the failure prediction reliability is increased. The electromechanical device neural network failure trend prediction method can be widely applied to the failure prediction and analysis of all kinds of electromechanical devices.

The invention provides a semantic matching method and system. The method includes the steps of conducting vector transformation treatment on any one of a first statement, a second statement, and a third statement in at least one domain, and obtaining a first statement vector, a second statement vector and a third statement vector; inputting the first statement vector and the second statement vector as positive samples, the first statement vector and the third statement vector as negative samples to a recurrent neural network model for training, and obtaining a semantic judgment model used for at least one domain; inputting the sentence vector of a to-be-identified statement of any one target domain in at least one domain and the sentence vectors of candidate statements in the target domain to the semantic judgment model; calculating the similarity degree between the sentence vector of the to-be-identified statement and the sentence vector of each candidate statement in the target domain successively; determining the candidate statement with the highest similarity degree as a target statement which is similar to the semantics of the to-be-identified statement.

The invention discloses a traffic flow prediction method and system for an urban road network, and relates to the technical field of intelligent transportation, and the method comprises the steps: building a training data set comprising an adjacent matrix and a time sequence vector matrix; establishing a graph wavelet attention gating recurrent neural network model; wherein the model comprises a graph convolutional neural network based on wavelet transform and a gating recurrent neural network connected with the graph convolutional neural network; the gating recurrent neural network comprisesgating recurrent units which are in one-to-one correspondence with the time periods and are integrated into the attention mechanism; training and optimizing the graph wavelet attention gating recurrent neural network model by using the training data set; and inputting the adjacency matrix of the urban road network to be predicted and the time sequence vector matrix into the trained and optimized graph wavelet attention gating recurrent neural network model to obtain the traffic flow of each road section in a future time period. The method and the system disclosed by the invention can accurately predict the traffic flow.

The invention discloses a video description method based on a dual-path fractal network and an LSTM. According to the method, key frame sampling is carried out on a to-be-describe video and an optical flow characteristic between two adjacent frames of an original video is extracted; learning is carried out respectively by using two fractal networks and high-level feature expressions of the video frame and the optical flow characteristic are obtained; the high-level feature expressions are inputted into two LSTM-unit-based recurrent neural network models; and then weighted averaging is carried out on output values of two independent modules at all times to obtain a description statement corresponding to the video. According to the method provided by the invention, the original video frames and the optical flow information are used respectively for the to-be-describe video; the added optical flow characteristic compensates dynamic information that is lost by the sampling frame inevitably; and the changes of the video in the spatial dimension and the time dimension are considered. Besides, the abstract visual feature expression is carried out on the bottom feature by using the novel fractal networks, so that the person, object and behavior and space position relation that are involved in the video can be analyzed and dug out precisely.

The invention relates to a segmentation-free off-line handwritten Chinese character text recognition method. The method comprises the steps that (S1) an off-line handwritten Chinese character text image is preprocessed; (S2) a spatial transformation network model is constructed; (S3) a deep convolutional neural network model is constructed; (S4) a recurrent neural network model is constructed through depth features extracted by the deep convolutional neural network model; (S5) probability distribution of sequence tags is output through a classifier CTC; and (S6) greedy search and search basedon dictionary rules are adopted to obtain a final text recognition result. According to the method, by the adoption of a model combining a spatial transformation network, a deep convolutional neural network and a recurrent neural network, correction processing and segmentation-free recognition can be performed on text lines with large offset, and the accuracy and robustness of recognition of complicated text lines are improved; the whole model framework is solved based on an iterative algorithm without the need for complicated excessive segmentation preprocessing, therefore, losses brought byan excessive segmentation method can be well reduced, entire model parameters can be optimized in a united mode, and recognition accuracy is improved.

Embodiments of the invention disclose a target tracking method and apparatus. Two layers of convolutional neural network models are combined with a time recurrent neural network model, so that the problem of low detection rate of a small target is solved. Moreover, information associated with the target in a background is extracted to perform target detection, so that the speed and accuracy of a target tracking model in video target detection are improved.

The invention discloses a recurrent neural network short-term power load prediction method for improving a whale algorithm, and relates to the technical field of short-term power load prediction. A recurrent neural network is used for short-term power load prediction, similar daily load data of a day to be predicted is used as input data of the recurrent neural network, and the number of input neurons, the number of output neurons, the number of hidden layers, the learning rate and the gradient descent algorithm of the recurrent neural network are determined. And a prediction model of the recurrent neural network is constructed. And the whale optimization algorithm is improved by using a differential evolution algorithm, so that the high-dimensional global optimization capability of a common whale algorithm is improved. An improved whale algorithm is adopted to pre-train the weight in the recurrent neural network, after pre-training is finished, the trained weight is put into a recurrent neural network model, then a gradient descent algorithm is adopted to train the recurrent neural network model, and after training is finished, a neural network model with the fixed weight is obtained, and then load prediction is carried out.

Embodiments of the invention provide an emotion recognition method and apparatus, which is applied to the technical field of multimedia. The method comprises the steps of recognizing an appearance characteristic of a human face in a frame sequence of a to-be-recognized video and a time characteristic of the frame sequence through a pre-built recurrent neural network model to obtain a first emotion result; recognizing the appearance characteristic of the human face in the frame sequence of the to-be-recognized video and the time characteristic of the frame sequence through a pre-built three-dimensional convolutional neural network model to obtain a second emotion result; performing fusion calculation on the first emotion result and the second emotion result to obtain a fusion result of the to-be-recognized video; and according to a maximum value in the fusion result, obtaining an emotion type of the to-be-recognized video through a pre-established corresponding relationship between the fusion result and the emotion type. According to the method and the apparatus, the complementarity of the recurrent neural network model and the three-dimensional convolutional neural network model improves the accuracy of recognizing the emotion type of the video.

The invention discloses a face emotion recognition method, a device, a computer device and a storage medium. Firstly, video data to be detected is obtained, and at least one face image to be detectedis extracted from the video data to be detected. A face detection algorithm is adopted to detect the face image to be detected, and N target face image regions of the video data to be detected are obtained, wherein N is a positive integer; N face video data to be recognized are obtained by video intercepting according to N target face image regions to be detected; N face video data to be recognized are input into a face emotion recognition model for recognition to obtain N face emotion information in the video to be detected, wherein the face emotion recognition model is based on a convolutionneural network. The length of recurrent neural network model training. The method directly recognizes the video data, avoids the lack of information represented by a single frame, and ensures the accuracy of face emotion recognition.

The present invention provides a recommendation method and system based on a recurrent neural network. The method comprises: using a recurrent neural network model to model a user behavior and using the model for a recommendation system; constructing a recurrent neural network model learning method for the recommendation system by combining a BP algorithm principle with a neural network structure of the present invention; training the recurrent neural network according to a feature of the recommendation system, and establishing a unique neural network structure, so as to generate a recommendation list for different users corresponding to interests thereof according to the new neural network structure. According to the recommendation method and system, user preferences are learned from history behaviors of the user by using the recurrent neural network, and a recommendation service is provided for the user on this basis; user behaviors of different types can be represented uniformly based on a time sequence according to a unique recursive structure of the recurrent neural network; and a deep neural network structure is formed if the recurrent neural network expands in time, and the user behaviors with relatively large randomness can be represented more accurately.

The invention provides a flight passenger flow rate prediction method based on a multi-granularity time attention mechanism. The method comprises the following steps of: constructing a recurrent neural network model based on a multi-granularity time attention mechanism; taking the historical flight passenger rate time sequences of all take-off moments on the air route as input sequences of the encoders, enabling the encoders to encode the input sequences, enabling the decoders to decode the coded information output by the encoders, and obtaining a flight passenger rate time sequence of the target flight. According to the method, the time sequence dependence of flight passenger rates at different take-off moments and the influence of the flight passenger rates at other take-off moments on the target flight passenger rate in the air route where the target flight is located are captured through the take-off moment attention mechanism, and meanwhile the trend and periodicity of the passenger rate sequence of the target flight are captured through the take-off day attention mechanism; and the influence of external factors such as flight attributes and holidays and festivals is considered, so that the model has a good effect on flight passenger flow rate prediction.