116results about How to "Strong feature extraction ability" patented technology

The invention discloses an intelligence relation extraction method based on neural network and attention mechanism, and relates to the field of recurrent neural network, natural language processing and intelligence analysis combined with attention mechanism. The method is used for solving the problem of large workload and low generalization ability in the existing intelligence analysis system based on artificial constructed knowledge base. The implementation of the method includes a training phase and an application phase. In the training phase, firstly a user dictionary and training word vectors are constructed, then a training set is constructed from a historical information database, then corpus is pre-processed, and then neural network model training is conducted; in the application phase, information is obtained, information pre-processing is conducted, intelligence relation extraction task can be automatically completed, at the same time expanding user dictionary and correction judgment are supported, training neural network model with training set is incremented. The intelligence relation extraction method can find the relationship between intelligence, and provide the basis for integrating event context and decision making, and has a wide range of practical value.

The invention discloses a multiple dimensioned convolution neural network-based real time human body abnormal behavior identification method. A convolution neural network is used for replacing a conventional feature extraction algorithm, and the convolution neural network is improved so as to satisfy requirements for human body behavior classification; specifically, three dimensional convolution, three dimensional down-sampling, NIN, three dimensional pyramid structures are added; human body abnormal behavior feature extraction capability of the convolution neural network is enabled to be increased; training operation is performed in a specific video set, features with great classification capacity can be obtained, robustness and accuracy of a whole identification algorithm can be improved, GPU speed is increased so as to satisfy requirements for practical application, and therefore multi-channel videos can be monitored in real time.

The invention discloses a tool wear condition prediction method of a numerical control machine tool based on a parallel deep neural network. A dynamometer, an acceleration sensor and an acoustic sensor are installed on a workbench and a fixture of the numerical control machine tool; a milling experiment is conducted, the cutting force and vibration and acoustic signals of a milling process are collected so as to obtain multisensor data, and the wear capacity of a tool is collected; pretreatment is performed so as to obtain training data and to-be-tested data; a parallel deep neural network model is established; the treated training data and the label of the wear capacity of the tool are input into an offline training model in the parallel deep neural network model; and the to-be-tested multisensor data are introduced into the trained model so as to predict the wear capacity of the tool in real time and on line. According to the method, the implied characteristics during tool processingof the numerical control machine tool are fully mined, and the wear capacity of the tool can be predicted in real time. The method has the advantage of wide applicability and can be widely applied tovarious numerical control machine tools.

The invention discloses a method for constructing a face recognition model based on a ParaSoftMax loss function, which comprises the following steps: selecting a basic convolution neural network modelaccording to an application environment of a task; acquiring a face image marked with human face identity information in a specified number as a training data set; the decision edge parameters are obtained according to the difference of the class center angles between the difficult sample eigenvectors and the simple sample eigenvectors and the class center angles in the basic convolution neural network model. Obtaining a ParaSoftMax loss function according to the decision edge parameter; setting the loss function at the last layer of the basic convolution neural network model to form a face recognition model based on the loss function; input the training data set to the face recognition model, minimizing the loss function iterative training model parameters, and obtaining the optimal facerecognition model. Thus, the face recognition model of the present application can improve the accuracy of face recognition.

The invention discloses a vehicle target detection method and aims to solve the technical problems that a traditional method is complicated in detection process, large in calculated quantity and inaccurate in detection. According to the method, a VOC 2007 vehicle dataset is made according to an ImageNet dataset; a training model is configured by the adoption of a caffe deep learning framework andan improved Faster R-CNN algorithm, and an Inception network structure is introduced to perform feature extraction on an image; a sliding window with the area of 642 is additionally arranged and usedfor detecting a small target; a target detection problem in a scene is converted into a target dichotomy problem, and a vehicle target is detected and recognized; an RPN loss function is used to perform optimization; and a SoftMax algorithm is utilized to classify vehicle image features, and therefore a final result is obtained. The method has the advantages that the calculated quantity of the image features is reduced, the detection process is simplified, the image feature extraction capability is enhanced, network recognition precision is improved, and the detection result is more comprehensive.

The invention discloses a method for predicting the remaining service life of complex equipment based on a combined depth neural network. The main steps are as follows: acquiring multi-sensor data ofcomplex equipment; Obtaining effective measurement data by feature selection; obtaining A plurality of slice samples by preprocessing; Establishing the neural network regression model which combines the attention mechanism and depth neural network; The slice samples and their corresponding labels are inputted into the neural network regression model to train the neural network regression model offline. inputting The slice samples of multi-sensor data to be predicted into the trained neural network regression model, and the remaining service life of complex equipment is obtained. Considering the data characteristics of the multi-sensor signal, the invention fully excavates the local characteristics and the time sequence information in the data, has high prediction accuracy and wide applicability, and can be widely applied to various pieces of complex equipment.

The invention discloses a remote sensing image scene classification method based on a scale attention network, and the method comprises the steps: randomly dividing a scene data set into a training set and a test set in proportion; then, preprocessing the data set, wherein preprocessing includes image scaling and normalization; inputting the data set into an attention module for significance detection, and generating an attention graph; secondly, initializing scale attention network parameters by using the pre-training model, finely adjusting the scale attention network by using the training set and the attention map, and storing the trained network model; and finally, predicting the category of the to-be-classified image scene by using the fine-tuned scale attention network. According tothe remote sensing image scene classification method based on the scale attention network, the multi-scale attention graph is used for weighting the feature graph for multiple times, the multi-scale image features are extracted and fused, feature representation with enhanced discrimination is generated, and a better effect is achieved in remote sensing image scene classification.

Embodiments of the invention provide a neural network model compression method and apparatus, a storage medium and an electronic device. The method comprises the steps of obtaining a first neural network model; keeping or increasing the depth of the first neural network model and compressing at least one network parameter of at least one network layer of the first neural network model to obtain asecond neural network model; and based on a sample data set and at least according to an output of the first neural network model, training the second neural network model. Therefore, the compressed neural network model having good feature extraction capability and performance equivalent to that of the uncompressed neural network model can be obtained by training; and the method has universality and is suitable for the neural network model realizing any function.

The invention discloses a complaint short text classification method based on deep integrated learning, which comprises the following steps: preprocessing a client complaint text set to obtain a preprocessed complaint text set; Designing complaint classification labels according to the theme classification of the preset complaint text, and marking corresponding complaint classification labels on the preprocessed complaint text set to obtain a training sample set; Performing text feature extraction on the training sample set by adopting a BTM topic model to obtain text feature vectors; Carryingout text feature extraction on the training sample set by adopting a convolutional neural network to obtain a convolutional semantic feature vector; Performing normalization and fusion on the text feature vector and the convolutional semantic feature vector by adopting a normalization combination strategy to obtain a combined text feature vector; And inputting the combined text feature vectors into a random forest model for training, combining classification results of a plurality of decision trees by adopting a weighting method according to the difference of different decision trees, and obtaining the category with the maximum probability as a text classification result of the training sample set.

The invention discloses an automatic millimeter wave image target identification method and a device. The method comprises steps that (1), sub image blocks of a to-be-identified target millimeter wave image are acquired; (2), based on a trained convolutional neural network, target containing probability values of the sub image blocks are acquired; (3), based on the probability values, a probability cumulative graph of the to-be-identified target millimeter wave image is acquired; and (4), based on the probability cumulative graph, target marking is carried out so as to accomplish target identification of the to-be-identified target millimeter wave image. The invention further discloses an automatic millimeter wave image target identification device. The device and the method are advantaged in that the device and the method are suitable for automatic millimeter wave image target identification, the excellent target identification effect is realized, a problem that employing traditional manual design characteristics for the millimeter wave image can not realize the excellent detection effect in the prior art is solved, precise target positioning is realized, false alarms are reduced, safety check efficiency is improved, and manpower cost is reduced.

The invention relates to a digital image denoising method based on NSST and CNN. The digital image denoising method comprises the following steps: 1, performing network training: converting noise approximately complying with Poisson distribution in an image into noise approximately complying with standard Gaussian distribution through Ascombe transformation; through the NSST, namely the non-subsampled shearlet transform, decomposing a noise image and a high-quality image corresponding thereto into multiple levels of sub-band images respectively, and shearing each level of sub-band image into image blocks with certain sizes as a data set; based on the obtained data set, performing convolutional neural network training, namely CNN training; 2, denoising the image based on a network model obtained in the step 1.

The invention discloses a fingerprint image grain direction calculating and image dividing method based on a BP neuronic network and the method comprises the following steps: (1), dividing a fingerprint image I with a width of m and a height of n into blocks W(i, j) which are not mutually overlapped with a size of Omega*Omega(15*15); (2) using a method based on gradient vectors to calculate the grain direction O(W(k, l)) of W(k, l); (3), using the method of the BP neuronic network to confirm the accuracy on the result of the direction O(W(k, l)) of each image block and carrying out the primary cutting on the image; if the direction of the image block W(i, j) is correct, then the image block is set to be a foreground and M(W(i, j)) is equal to 1; if the direction is wrong, then the image block is set to be a background and M(W(i, j)) is equal to 0. The method can better process the fingerprint image with low quality, is high in the accuracy rate of the characteristic extracting result, is low in calculating complexity and is accurate in image cutting.

The invention discloses a CNN multipoint regression prediction model for traffic flow prediction. The CNN multipoint regression prediction model comprises the following steps that a first perception input layer; a second convolutional layer that performs convolution on the input layer data and outputs the data by an activation function; a multi-layer convolutional layer that performs convolution on the output of the last layer as an input and outputs data by an activation function; a fourth full-link layer; a fifth discard layer: a random discard layer discards some redundant neurons and retains 40% to 70% of the full-link nodes of the last layer; a sixth output layer that subjects the effective node output of the discard layer to regression calculation, obtains a regression value as the output of the entire network, sets a total of m output nodes, that is, maps the full-link layer to the output layer to achieve weight combination. Compared with a traditional statistical regression model, the CNN multipoint regression prediction model has a feature extraction ability associated with a data space, has the advantages of local receptive field and weight sharing, and makes a better balance in time complexity and feature selection.

The invention discloses a dynamic gesture recognition method and system based on a two-dimensional convolutional network, and the method comprises the steps: collecting an actual dynamic gesture video, and carrying out the framing processing of the video; carrying out frame sampling on the actual image after framing; encoding the actual image after frame sampling to obtain an actual feature vectorof the actual image; fusing the actual feature vectors to obtain an actual feature matrix; and inputting the actual feature matrix into the trained two-dimensional convolutional neural network, and outputting a gesture recognition result. By processing a source video stream into a frame of image, and sending the frame of image into a two-dimensional convolutional network, a classification resultof a gesture action is obtained; wherein the image generated by the video contains the spatial feature information and the time sequence information of the video. According to the method, the calculation complexity of gesture recognition is effectively reduced.

The invention discloses a gas path fault diagnosis method and system for aeroengine dynamic process. The method includes the steps of establishing a feature extraction model based on a multi-layer kernel extreme learning machine; adopting a hidden Markov model based on time series modeling for fault identification. The method solves the problem that the conventional data-based engine fault diagnosis uses time series measurement data to diagnose the fault with low accuracy in the existing aeroengine gas path fault diagnosis in the dynamic process, is suitable for the engine dynamic fault diagnosis in consideration of the degradation of gas path components and the redundancy of sensor parameters, and has a positive promotion effect on engine health management and maintenance cost reduction.

The invention provides a Laplace spare deep belief network image classification method, and belongs to the field of image processing and deep learning. The method comprises the following steps that: firstly, on the basis of inspiration for primate visual cortex analysis, importing a punishment regular term into an unsupervised stage likelihood function, through a Lapalce sparse constraint, obtaining the sparse distribution of a training set while a CD (Contrastive Divergence) algorithm is used for maximizing a target function, and therefore, enabling unlabeled data to learn visual characteristic representation; secondly, putting forward an improved spare deep belief network, using Laplace distribution to induce the spare state of a hidden layer node, and meanwhile, using the scale parameter in the distribution to control spare strength; and finally, using a stochastic gradient descent method to carry out training learning on the parameters of the LSDBN (Laplace Spare Deep Belief Network). By use of the method which is put forwarded by the invention, even if the amount of each category of samples is small, best identification accuracy can be achieved all the time, and in addition, the method exhibits good spare performance.

The invention discloses a crowd density estimation method based on deep learning. The crowd density estimation method comprises the following steps: S1, labeling a training set; S2, generating a firstcrowd density map by utilizing the labeled information; S3, constructing a deep neural network model, and extracting features of different layers through the deep neural network model to establish amulti-layer feature pyramid; analyzing the output features of the feature pyramid to obtain a second crowd density map; S4, training the constructed neural network model by using the first crowd density map generated in the S2 to obtain a trained deep neural network model; and S5, inputting a to-be-estimated image into the trained deep neural network model to obtain a final crowd density map of the to-be-estimated image, and integrating the final crowd density map to obtain the estimated number of people of the to-be-estimated image. The crowd density estimation method based on deep learning estimates crowd density by fusing pyramid features of different layers, and has the advantages of being high in robustness and good in performance.

The invention discloses a method for predicting the residual life of complex equipment based on a double-layer long-short-term memory network. The method includes by adopting a depth learn algorithm,preprocessing the historical data of complex equipment, and building a two-tier long-term and short-term (LSTM) network, wherein the number of LSTM cells in a two-layer long-short-term (LSTM) networkis determined by a continuous time period. The current data are preprocessed and transferred to the trained two-layer LSTM network, and the output of the two-layer LSTM network is set as the predictedvalue of the residual life of complex equipment. The residual life prediction model of the complex equipment based on the double-layer LSTM provided by the invention can improve the prediction accuracy of the residual life of the complex equipment. Complex equipment can thus be maintained in a timely and effective manner, while reducing the occurrence of accidents. It is of great significance toensure the safety of complex equipment operation and reduce unnecessary maintenance at the same time.

The invention discloses an arrhythmia prediction method. The prediction method includes the following steps: generating, by an ECG data set, multiple feature vectors respectively through multiple CNNmodels, and stacking the multiple feature vectors to form input vectors; training the input vectors through multiple RNN models to obtain multiple bottom layer classifiers; and adopting a fully connected neural network to fuse the multiple bottom layer classifiers to predict abnormal ECG events. The multiple CNN models are adopted to extract the feature vectors of ECG data so as to form the inputvectors; the RNN is adopted to extract the timing information in the input vectors so as to form the bottom layer classifiers having the abilities of predicting abnormal ECG; and the bottom layer classifiers predict the ECG data again to obtain prediction results, and the prediction results can be fused to form the optimal prediction result by adopting the fully connected neural network so as to further make the prediction results more accurate.

The invention relates to a seal script recognition method and system based on an Incep-CapsNet network. The method comprises the steps that a seal script image dataset is acquired, and classificationtags are produced according to character patterns and character styles; preprocessing operation and image augmentation operation are performed on the image dataset; feature extraction and classification recognition are performed on seal script images obtained after preprocessing and image augmentation through the Incep-CapsNet network; and reconstruction is performed on the seal script images obtained after preprocessing and image augmentation through a reconstruction network to serve as an effective regularization mode. Through the seal script recognition method and system, more and more essential features can be extracted according to an Inception structure, the defects that the orientation of features and relative positions in a space cannot be recognized through a traditional deep convolutional neural network model are overcome according to a capsule network, and high recognition precision and high robustness are both achieved.

The communication interference signal type intelligent identification method disclosed by the invention is high in identification rate, and solves the problem of interference signal identification in a complex spectrum environment. According to the technical scheme, the interference identification process is divided into a preprocessing part and a network identification part, time domain windowing, power normalization processing and Fourier transform are carried out on signals received by a receiver in the preprocessing part, and time domain data and frequency domain data are obtained; in the recognition network part, time domain data are input into a time domain feature extraction branch, and time domain features are obtained through a time domain convolution module and a time domain LSTM module; the frequency domain data is input into a frequency domain feature extraction branch, and frequency domain features is obtained through a frequency domain convolution module and a frequency domain LSTM module; the time domain features and the frequency domain features are sent to a fusion module for feature fusion; and finally, the obtained fusion features are sent to a classification module to carry out classification identification on the interference signals to obtain a type identification result of the interference signals.

The invention relates to a #-shaped LBP and deep learning combination-based human face identification method, and belongs to the technical field of artificial intelligence. For the problems of insufficient feature extraction and classifier fitting of a conventional LBP, a human face identification method based on a #-shaped LBP and deep learning of a local texture feature is proposed. By utilizingan improved LBP algorithm, the local texture feature of a human face image is extracted, and an LBP histogram is established; then a deep belief network-based deep learning framework is established,the LBP histogram is input to a deep belief network, the network is trained by adopting an unsupervised layer-by-layer training method and a supervised BP algorithm, self-learning and self-optimization of the network are realized, and network parameters are obtained; and finally, the human face image is classified and identified by utilizing the DBN. According to the method, the distinctive texture feature of the human face image can be extracted; the superiority of the algorithm in identification rate is verified; and the robustness is relatively strong.

The invention discloses a radar radiation source individual identification system based on a radar pulse sequence. The radar radiation source individual identification system comprises a radar individual pulse sequence database, a data preprocessing module, a different individual weight calculation module, a weighted extreme gradient lifting radar individual modeling module, a radar individual identification module and a radar individual identification final result calculation module. According to the invention, individual identification of the radar radiation source is realized based on the radar pulse sequence. In the radar radiation source individual identification system based on the radar pulse sequence, an original pulse sequence is used as input, wavelet decomposition is performed on the original radar pulse sequence, so that multi-scale input can be obtained. More distinguishable subtle features are mined, a weighted ensemble learning algorithm is adopted to establish a radar radiation source individual recognition model, the problem that the number of radar individual samples in a database is unbalanced can be solved, and meanwhile the method has the advantages of being high in feature extraction capacity, high in accuracy, high in modeling speed and the like.

The invention discloses a scindapsus aureus leaf shape parameter estimation method based on MRE-Point Net and an auto-encoder model, and the method comprises the steps: carrying out the photographingof scindapsus aureus from a single angle through a Kinect V2 camera, obtaining point cloud data, carrying out the preprocessing of the data through straight-through filtering, segmentation and point cloud simplification algorithms, building a scindapsus aureus leaf geometric model through a parameter equation, and calculating the blade length, the blade width and the blade area of the geometric model; and inputting the discrete point cloud data of the geometric model into a multi-resolution point cloud deep learning network to obtain a pre-training model, and taking the discrete point cloud data of the geometric model as input to obtain a pre-training model of an auto-encoder through encoding and decoding operation, performing secondary processing noise reduction is performed on input point cloud data through a pre-training model of an auto-encoder, and then parameter fine adjustment is performed on the pre-training model by using the measured scindapsus aureus leaf shape parameter label so that shape parameter estimation of the input scindapsus aureus leaf point cloud data can be completed.

The invention discloses a voiceprint recognition method based on a TDNN (time delay neural network), and solves the problems that a voiceprint recognition algorithm is complicated and data are complex. The voiceprint recognition method is technically characterized by extremely strong feature extraction capacity of a neural network. The TDNN is used for extracting the feature vector of a voice segment of a speaker, a pooling layer and a softmax layer are used for acquiring the posterior probability of the voice segment of the speaker, a loss function is used for training to obtain a cross entropy, the softmax layer is removed after training, the feature vector for finally training a PLDA (probabilistic linear discriminant analysis) model is acquired, transcription of training data is omitted, and simple calculation and good recognition effects are achieved.

The invention provides a bridged convolutional neural network-based emotion electroencephalogram signal classification method, which comprises the following steps of: firstly, extracting electroencephalogram signal bottom-layer features by using a first convolutional layer of V3, taking the electroencephalogram signal bottom-layer features as the input of V1, and inputting the V1 into a third convolutional layer to extract middle-layer features after the V1 is down-sampled by a second pooling layer; and the middle-layer feature is used as the input of V2, is down-sampled by the fourth poolinglayer of V3 and then is input to the fifth convolution layer of V3 to extract the high-layer feature. And then, the three layers of features are respectively subjected to dimensionality reduction andthen are input into an eighth full-connection layer of the V3 for fusion, and finally, the three layers of features enter a Softmax layer for classification. And comparing the classification result with the actual label, calculating a loss value, and then updating the convolution kernel and the connection weight by using a back propagation algorithm. According to the method, the electroencephalogram signal classification accuracy can be high, and the recognition result is superior to that of a traditional machine learning method and a traditional CNN model.

The invention discloses a video anomaly detection method for an ST-Unet network, and relates to the field of computer vision and pattern recognition. According to the invention, a new ST-Unet networkis provided for solving the processing problem of such data with strong space-time correlation as videos; and the the good modeling capability of the Unet network is not only utilized in spatial characteristics, the modeling capability of the ConvLSTM is but also combined in time. In order to improve the accuracy and the blooming capability of the algorithm, damage preprocessing is carried out oninput video data through a Dropout layer in the algorithm training process. The network obtained by training the damaged training data not only has a better anomaly detection effect on complete test data, but also can detect whether the noise-containing data in the test process is abnormal or not. According to the invention, a reconstruction algorithm result and a prediction algorithm result are combined and discriminated, so that high-precision ST-based is realized; the invention discloses a video anomaly detection algorithm of a Unet network.

The invention discloses a diaphragm flaw detection method based on a reverse bottleneck structure deep convolutional network, and the method achieves the detection and calibration of flaws in a diaphragm through employing the reverse bottleneck structure deep convolutional network. The method comprises the steps of image acquisition, image segmentation, data annotation, network training, flaw detection, image splicing and the like. According to the method provided by the invention, the effectiveness of the deep convolutional network for image feature extraction is fully utilized, and the number of parameters can be greatly reduced by the reverse bottleneck structure under the condition that the detection precision is ensured, so that the purpose of quickly and accurately detecting the defect points in the diaphragm is achieved.

The invention discloses a target tracking algorithm based on a dense connection convolutional neural network, and the algorithm comprises the following steps of S101, extracting the characteristics ofan input image, employing the dense connection convolutional neural network as a characteristic extraction network in advance, and obtaining the characteristics of an extracted output image; S103, carrying out output feature map processing, performing bilinear interpolation on the obtained extracted output image features to obtain a bilinear interpolation feature map; and S105, calculating to obtain the feature map. The algorithm has the beneficial effects that the densely connected convolutional neural network is used for replacing a feature extraction network of a traditional twin convolutional network, so that the network obtains the higher feature extraction capability; the bilinear interpolation is carried out on a feature graph outputteed by a feature extraction network, so that theresolution of the feature graph is improved, the positioning precision of a tracking algorithm is improved; and in addition, an RPN layer is added on the basis of a traditional twinning convolutionalnetwork, so that the distinguishing capability of the tracking algorithm on a target and a background is enhanced.

The invention provides an image classification method based on a structure optimization sparse convolutional neural network and a medium. For a connection structure of a convolutional layer of the convolutional neural network to an input feature map channel, a genetic algorithm is used to sparsify the convolution layer, and the sparsified convolution model I used to carry out image classification. Firstly, a convolution model is pre-trained and a pre-training weight is stored; then, according to a certain convolution layer except the model input layer, binary coding is carried out on connection of an input characteristic channel, and a plurality of binary sequences are generated and serve as an initial population; secondly, selecting, crossing and mutating of binary codes are realized byusing a genetic algorithm; and finally, after a plurality of iterations, the obtained optimal binary sequence is decoded to obtain a sparse feature channel connection structure, and the classificationaccuracy of the model is recovered through weight fine adjustment.