129results about How to "Mitigate Vanishing Gradients" patented technology

The invention discloses a collaborative filtering recommendation method and a collaborative filtering recommendation system of a convolution neural network integrating an attention model, which relates to the technical field of data mining recommendation, improves feature extraction efficiency and scoring prediction accuracy, reduces operation and maintenance cost, simplifies cost management mode,and is convenient for joint operation and large-scale promotion and application. The invention relates to a collaborative filtering recommendation method of a convolution neural network integrating an attention model, comprising the following steps: step S1, splicing a user feature vector and an item feature vector into a new vector; S2, sending the new vector as an input vector into the multi-layer perceptron to learn and predict the score; The attention model is fused into the object potential vector to obtain the convolution neural network of the object feature vector or the hidden layer of the multi-layer perceptron.

The invention discloses a cell division detection method based on deep learning, relates to the technical field including cytobiology and the like, and solves the problem of inaccurate cell division detection results. The cell division detection method comprises the following steps: obtaining an image in a cell culture process, and forming a continuous image sequence according to the images obtained at certain time intervals; judging and comparing all cell division processes in the image sequence, and labelling a position where a process that a paternal cell is divided into two daughter cell membranes can be clearly observed in cell division, and forming a labelled dataset; constructing a full convolutional neural network model, determining the parameter of the full convolutional neural network model, loading the labelled dataset into the full convolutional neural network model, using a deep learning method to directly carry out feature learning on an image sequence, and obtaining the trained full convolutional neural network model; and deploying the trained full convolutional neural network model for automatically detecting the cell division of the cell culture data of a similar type. The cell division detection method is used for cell division detection.

The invention provides a cycle nerve network text emotion analysis method by embedding logic rules, and the method comprises the steps of grabbing a text corpus for training and conducting emotion class labeling; then dividing the emotion labeled text corpus into a training set corpus and a test set corpus for dividing the words and removing the stopped words; then conducting training to the training set corpus and the test set corpus whose words are divided and stopped words are removed by using word2vec algorithm so as to obtain a corresponding word vector; inputting the training set corpus and the test set corpus in an existing knowledge base and analyzing in combination with a probability graph model; embedding a first order logic rule in the cycle nerve network through a logic cycle nerve network structure (Logic-RNN and Logic-LSTM). According to the invention, the training direction of the cycle nerve network can be controlled, and inclination to human's intuition is realized; besides, the precision of text emotion analysis is improved. Therefore, the method can be used in other fields like natural language processing and machine learning.

Provided is an acoustic model modeling method of a residual error long short-term memory recurrent neural network. The method comprises: directly connecting input of a standard long short-term memory neural network framework with an internal node, splicing an original vector of the node with an input vector, and then projecting to an original dimensionality. The method provides a position of an internal node for direct connection. Experiments show that the position can bring about performance improvement of identification, and further improve system performance combined with a Dropout technology based on frames. The method improves performance of a speech recognition system based on a long short-term memory recurrent neural network, and prevents an overfitting problem combined with the Dropout technology based on frames, and realizes an objective of improving model generalization ability. The method can be widely applied in the fields of various man-machine interaction related to speech recognition.

The invention discloses a non-uniform motion blurred image adaptive restoration method based on an attention model, and belongs to the field of digital image/video signal processing. The invention designs a conditional generative adversarial network combined with an attention mechanism. The generation network is of a coding and decoding structure, in the coding stage, a dense connection network isadopted to extract features, the feature utilization rate is increased, feature propagation is enhanced, and a visual attention mechanism is added, so that the network can adaptively modulate networkparameters and dynamically remove image blurring for different input images. According to the method, the clear image can be effectively restored from the non-uniform motion blurred image. The technology has a wide application prospect in the fields of target tracking, traffic detection, military reconnaissance and the like.

The invention relates to an image classification method based on a multi-scale dense convolutional neural network and a spectral attention mechanism. The multi-scale dense convolutional neural networkis constructed by using a dense connection mechanism, and the dense connection mechanism can effectively alleviate the problem of gradient disappearance, enhance feature propagation, encourage feature multiplexing, greatly reduce the number of parameters, and reduce the demand for training samples in the network training process; In addition, the network is combined with a spectrum attention mechanism, and the characteristic utilization of the spectrum direction is more effective. According to the method, autonomous extraction and high-precision classification of hyperspectral image depth features are realized under the condition of small samples. Compared with an existing hyperspectral image classification method based on deep learning, the hyperspectral image classification method has the advantages that the sample demand is lower, and the precision is higher.

The invention provides a remote sensing image thin and weak target segmentation method. Firstly, data enhancement and corresponding preprocessing are carried out on an original remote sensing image, U-net is improved by means of the dense connection thought of DenseNet, and a Dense-Unet network structure is provided. Dense convolution is used in a network structure, the cascade relation between convolution channels is enhanced, through a symmetric structure and a jump connection thought, the connection between features of all layers is further tighter, and thin and weak target features can belearned more effectively. In order to ensure the real-time performance of final network identification and reduce the parameter quantity, a bottleneck layer and a batch normalization layer are introduced behind each dense block. And the objective function is adjusted by using the cost-sensitive vector weight, so that the problem of unbalanced segmentation target categories is solved, and the segmentation precision is further improved. And finally, a plurality of independent models are trained by using an ensemble learning method, the independent models are combined, and target category information is jointly predicted in the picture.

The invention provides a DGA domain name detection method based on a GAN and a Char-CNN, which is used for solving the problem of low detection recall rate of a low-randomness DGA domain name in the prior art, and comprises the implementation steps of obtaining a training sample set and a verification sample set; constructing a generative adversarial network (GAN) and a character-level convolutional neural network (Char-CNN); carrying out iterative training on the GAN; obtaining an augmented training set; carrying out iterative training on a character-level convolutional neural network Char-CNN; and detecting the domain name based on the trained character-level convolutional neural network Char-CNN '. According to the method, the GAN is used for generating the adversarial domain name to augment the data set, the richness of the training sample set is improved, the error rate of the detection model is reduced through the residual block structure, the detection recall rate of the low-randomness DGA domain name is increased. Meanwhile, hyper-parameters needing to be calculated by the Char-CNN are few, and the training time of the detection model is shortened.

The invention discloses a landmark building identification and detection method based on deep learning. The method comprises the following steps of inputting a to-be-identified landmark building imageinto a DenseNet network to obtain a feature block diagram containing the target building feature information, and then sending the feature block diagram into a region suggestion network to predict abinary category of the feature block diagram and the coordinates of a target building in an original image; completely mapping a prediction candidate box to the feature block diagram by using a RoI Agign method; finally, carrying out classification and frame regression on the more accurate feature block diagrams to obtain the prediction probabilities and the coordinate positions of different landmark buildings, removing the redundant candidate frames through a non-maximum suppression method, fusing the diagrams with the wider coverage regions, and finally realizing the identification and detection of the landmark buildings. According to the method, the prediction of the landmark building candidate frames is more accurate, the prediction range is larger, and the method has the better identification capability on the landmark building images in the complex environment.

The invention discloses a traffic signal detection and identification method based on a cascaded network, and relates to the technical field of image processing and sign identification. The method comprises the following steps that S1) an original image is collected; S2) the original image is detected to obtain a candidate area which may include a traffic signal or lamp, the trained three-order cascaded network composed of a 12-net, 24-net and 48-net is used for detection in a detection process; and S3) the candidate area is identified, the candidate area without the traffic sign or lamp is eliminated, and concrete types of traffic signs and lamps are obtained. The traffic signs and lamps can be detected at the same time, and the problem, in a present traffic sign detection and identification technology, that missing detection is caused by influence of illumination on a color space, small difference between the traffic signal and the background environment and connection of multiple traffic signs can be solved.

The present invention discloses a newborn pain expression recognition method and system based on a deep 3D residual network. The method comprises: establishing a newborn expression video library containing pain expression category tags, and dividing samples in the newborn expression video library into a training set and a verification set; constructing a deep 3D residual network for newborn pain expression recognition, pre-training the network by using a publicized large-scale video database with category tags to obtain initial weight parameter values, and then performing fine-tune on the network by using the samples in the training set and in the verification set in the newborn expression video library to obtain a trained network model; and inputting a to-be-tested newborn expression video segment into the trained network model for expression classification recognition, and obtaining a pain expression recognition result. According to the technical scheme of the present invention, a deep 3D residual network is used to extract temporal and spatial dynamic features capable of reflecting time information from the video, and the change of the facial expression can be better characterized, so that the accuracy of the classification recognition can be improved.

The invention relates to a U-Net model-based medical image segmentation method and apparatus, and a storage medium. The method comprises the steps of determining a target segmentation area of a plurality of medical images; respectively carrying out medical scanning on the target segmentation areas of the plurality of medical images to obtain color medical image samples; respectively preprocessingeach color medical image sample to obtain a gray image after the G channel is extracted; respectively carrying out noise removal operation on each gray image, and respectively generating a corresponding segmentation label image according to each gray image after noise removal; performing at least one data enhancement processing operation of rotation, translation and scaling on the medical image sample and the segmented label image to obtain a plurality of bitmap samples; dividing each bitmap sample into a training set and a verification set; inputting each training set into a medical image segmentation model to train the medical image segmentation model; debugging the model parameter by using each verification set to obtain an optimal model parameter; and performing performance testing byusing each verification set pair to obtain the optimal segmentation accuracy.

The invention provides a punctuation mark prediction method based on a self-attention mechanism, which comprises the following steps of performing voice recognition based on an automatic voice recognition technology to obtain a punctuation mark-free text; processing the punctuation-free text to obtain a text sequence; and constructing a punctuation mark prediction model, and importing the text sequence into the model to complete punctuation mark prediction of the text sequence. According to the punctuation prediction method based on the self-attention mechanism provided by the invention, the punctuation prediction of the speech recognition text is realized by constructing the punctuation prediction model, the problem of gradient disappearance is effectively alleviated, the feature transferis enhanced, and the long-term dependency relationship of the text is effectively established. Meanwhile, compared with a previous model, the additional parameters are not needed, the transmitted data size is effectively reduced, and the training difficulty of the parameters is reduced.

The invention relates to the technical field of target attribute identification, and provides a small target vehicle attribute identification method based on feature fusion. The method comprises: firstly, constructing a small target vehicle attribute recognition network based on feature fusion, comprising a feature pyramid network, a regional nomination network and a small-size target cascade network; inputting a traffic image to be detected into the feature pyramid network, generating a feature map containing low-level edge detail information, middle-level stacking fusion scale information and high-level semantic information, and stacking and fusing the feature map to obtain a multi-scale feature map; inputting the multi-scale feature map into a regional naming network to generate a candidate box containing a target; inputting the multi-scale feature map and the candidate box into a small-size target positioning network at the same time, outputting target coordinate information, and cutting a target according to the information; and finally, inputting the sheared target into a small-size target classification network, and identifying and outputting the target and the category thereof. According to the invention, the accuracy of small-size target attribute identification can be improved, and the false identification rate and the missing identification rate are reduced.

The invention provides a stable controllable image generation model training method based on a W distance, and the method comprises the following steps: a, preprocessing image data to acquire sample data of a training set; b, constructing a stable controllable image generation model based on a W distance; c, constructing an overall model according to a loss function of each network; d, alternatelyiterating model training by a gradient descent algorithm to ensure normal convergence of model parameters; and e, after the model parameters converge, respectively separating a generator network G, an encoder network E and a classifier network C as independent service products. The invention is an end-to-end network model, improves a model in the prior art from two aspects of a model structure and a distribution measurement standard, solves the problems of unstable model training gradient and unstable model gradient descending direction, can stably and directionally generate an image sample.Robustness in the model training process and controllability of the generated sample are improved.

The invention discloses an image restoration method based on a generative adversarial neural network, and the method employs four parts: a generative network, a global discrimination network, a localdiscrimination network, and an LSTM neural network. And the two discriminators are mainly used for ensuring that the repaired missing area can keep consistent with the surrounding area. The LSTM neural network is mainly used for repairing damaged images by stages. The algorithm comprises a data preprocessing module, a model training module and an image restoration module, and is mainly used for performing a semantic restoration task on a large-area missing image so as to reconstruct a complete and vivid image conforming to human eye senses.

The invention provides an earthquake precursor data abnormity detection method. The method comprises the following steps: obtaining a historical earthquake precursor data set, carrying out data pre-processing and dividing the processed data set into a training set, a verification set and a test set; constructing an LSTM-RNN prediction model; serving the data in the training set as input, and training the prediction model; checking error changes during training with the data in the verification set; setting an abnormal trigger threshold according to a loss function; testing the prediction modelwith the data in the test set, and adjusting the value of hyperparameters according to recognition rate of abnormal data in the test set; and carrying out earthquake precursor data abnormity detection by utilizing the prediction model. The method can identify all kinds of abnormities in the earthquake precursor data and make clear description and accurate distinguishing, and can identify the abnormal information that is not easily perceived by human.

The invention discloses a Chinese sentence level lip language recognition method combining DenseNet and resBi-LSTM. The Chinese sentence level lip language recognition method divides the lip languagerecognition into two parts: Pinyin prediction and language translation, and reduces the lip language recognition difficulty. The Chinese sentence level lip language recognition method extracts visualfeatures by using DenseNet, and fully utilizes shallow features, thus effectively solving the problem of gradient disappearance, and reducing network parameters. The Chinese sentence level lip language recognition method uses 1 * 1 convolution for replacing full connection to achieve a feature dimension reduction function, and maintains space information in features, wherein the space informationplays an important role in lip language recognition technology research. The Chinese sentence level lip language recognition method uses the resBi-LSTM for processing the visual features, and finally,obtains the complex features combining the visual features and the semantic features, so that the loss of effective information is reduced, and the lip language recognition accuracy is improved.

The invention belongs to the technical field of personnel flow prediction, and provides a community personnel flow prediction method and system, a storage medium and equipment. The community personnelflow prediction method comprises the steps of obtaining community personnel flow related data in a historical time period adjacent to a to-be-tested time period, wherein the community personnel flowrelated data comprises flow quantity, flow direction and external influence factor data; preprocessing the community personnel flow related data in the historical time period by adopting a variationalmode decomposition method, and extracting a plurality of intrinsic mode components and residual components; inputting the intrinsic mode component into a time sequence model and outputting an initialcommunity personnel flow prediction result, inputting the residual component into an allowance prediction model and outputting a community personnel flow external influence prediction result, superposing the initial community personnel flow prediction result and the community personnel flow external influence prediction result, and obtaining a final community personnel flow prediction result. Thecommunity personnel flow prediction precision and resource allocation can be improved.

The invention discloses a hyperspectral image classification method based on singular value decomposition and a spatial-spectral domain attention mechanism. The method comprises steps of reading a hyperspectral image from the data set, wherein the hyperspectral image data set comprises three widely used hyperspectral image data sets, that is, an Indian Pines data set, a Pavia University data set and a Salinas Valley data set; and selecting any one of the data sets and correspondingly processing a class label ground true graph which only has a partial region; performing rough processing on thesample, and constructing an unsupervised feature extraction model based on a singular value decomposition convolutional network; selecting a training set, a verification set and a test set according to a ratio of 10%: 10%: 80% of the training set to the verification set to the test set; carrying out fine processing on the sample, and constructing a double-branch classification model based on a spatial-spectral domain attention mechanism network; training the classification model by using the training data set to obtain a trained classification model; and classifying the test data set by usingthe trained classification model to obtain the category of each pixel point in the test data set. According to the invention, the precision and speed of hyperspectral image classification are improved.

The invention provides a lane line detection method based on a semi-supervised generative adversarial network. The lane line detection method comprises the steps of: S1, constructing the generative adversarial network, and establishing a training set, a verification set and a test set of the generative adversarial network; S2, pre-training the generative adversarial network through utilizing labeled data in the training set; S3, performing real training on the generative adversarial network by using the labeled data and the unlabeled data in the training set, and adjusting hyper-parameters ofthe generative adversarial network in a real training process through using the verification set; S4, after the real training is finished, evaluating the generalization ability of the generative adversarial network through using the test set, and if the generalization ability reaches a preset standard, entering S5; and S5, inputting an actual street image into a generator network subjected to realtraining to obtain an actual lane line of the actual street image, and superposing the actual lane line on the actual street image to complete lane line detection. According to the lane line detection method, the lane line identification precision can be improved.

The invention discloses an attention mechanism text recognition method based on deep learning, wherein the method comprises the following steps: S1, acquiring a model training data set; S2, preprocessing the model training data set; S3, constructing a model convolution layer; inputting the model training data set for feature extraction to obtain a feature map of an image, outputting to a subsequent recurrent neural network structure, wherein the visual features of the method are extracted through a multilayer convolutional neural network in which a convolutional layer and a maximum pool layerare staggered, the CNN accepts raw input, a feature grid V with the size of D * H * W is generated, wherein D represents the number of channels, and H and W represent the height and width of the result feature map. The method overcomes the problem of low text recognition accuracy of a neural OCR technology using a CTC-based model, can significantly reduce the amount of calculation of the network,and can ensure that the prediction precision of the model to the formula is not greatly reduced.

The invention discloses a breast nodule detection method and system based on ultrasonic medicine. The method comprises the steps of obtaining breast ultrasonic video data, carrying out the preprocessing, carrying out training on the preprocessed data through a transfer learning method, building a breast nodule detection model, inputting to-be-tested video data into the breast nodule detection model, and obtaining a result. The breast nodule detection method and system based on ultrasonic medicine aim to establish a precise method for judging benign and malignant breast tumors and carrying outpathological typing by taking ultrasonic breast video image examination as a main means, and solve the difficult clinical problem that dynamic identification and identification of breast tumors are difficult. The establishment of the highly accurate breast tumor qualitative system is of great clinical significance to the diagnosis and treatment of breast tumor patients.

The invention discloses a cloth defect detection method and system based on a deep neural network, and belongs to the technical field of pattern recognition. The method comprises the steps that a defect cloth image training set is used for training a deep neural network model, labels are defect types and real frame position information, the deep neural network model is composed of a backbone network and a detection network, and the backbone network is used for extracting three feature maps with different scales from defect cloth images; the detection network includesthree detection sub-networks and the detection result fusion module, wherein the three detection sub-networks are the same in structure. Each detection sub-network is used for detecting a defect type and prediction frame position information from the feature map, and consists of three dense connecting blocks, and the feature channel connection between the dense blocks is used for enhancing feature transfer, and the detection result fusion module is used for performing non-maximum suppression on the prediction result to obtain a final prediction frame and a defect type, and inputting to-be-detected cloth into the traineddeep neural network model to obtain a detection result, so that the type and the position of the defect in the cloth can be detected more quickly and accurately.

The invention discloses a coronary artery stenosis estimation method, system and device. The method comprises the steps: carrying out the preprocessing of an obtained medical digital imaging and communication DICOM image, and obtaining a DICOM image meeting the requirements; inputting the preprocessed DICOM image into a pre-trained neural network, and outputting a calcified area and a coronary artery area through the neural network; determining a calcified area where the plaque with the maximum diameter is located and a corresponding coronary artery area; calculating the maximum diameter of the calcified area and the corresponding coronary artery diameter, calculating the coronary artery stenosis degree according to the maximum diameter of the calcified area and the corresponding coronaryartery diameter, obtaining the lesion level corresponding to the coronary artery stenosis degree, and outputting the lesion level and the coronary artery stenosis degree.

The invention relates to a fruit picking robot target detection method based on deep learning in an unstructured environment, and belongs to the technical field of intelligent agricultural production.According to the method, a Mask R-CNN is used as a target detection framework, ResNet-101 is used as a backbone network and is combined with an FPN architecture to perform target feature extraction,then a feature map output by the backbone network is sent to an RPN to generate RoI, and then the RoI output from the RPN is mapped to extract corresponding target features in a shared feature map; and finally, the features are respectively output to an FC layer and an FCN layer to carry out target detection, frame regression and instance segmentation. According to the method, the problem of low detection precision caused by illumination condition change, branch and leaf shielding, fruit clustering overlapping and the like of a traditional digital image processing technology in an unstructuredenvironment is solved, and the defects of complex structure, slow gradient disappearance, large calculation training amount, slow model convergence and the like of a common target detection neural network are also overcome.