60results about How to "Improve feature learning ability" patented technology

The invention discloses a nasopharyngeal-carcinoma (NPC) lesion automatic-segmentation method and nasopharyngeal-carcinoma lesion automatic-segmentation systems based on deep learning. The method comprises: carrying out registration on a PET (Positron Emission Tomography) image and a CT (Computed Tomography) image of nasopharyngeal carcinoma to obtain a PET image and a CT image after registration;and inputting the PET image and the CT image after registration into a convolutional neural network to carry out feature representation and scores map reconstruction to obtain a nasopharyngeal-carcinoma lesion segmentation result graph. The method carries out registration on the PET image and the CT image of the nasopharyngeal carcinoma, obtains a nasopharyngeal-carcinoma lesion by automatic segmentation through the convolutional neural network, and is more objective and accurate as compared with manual segmentation manners of doctors; and the convolutional neural network in deep learning isadopted, consistency is better, feature learning ability is higher, the problems of dimension disasters, easy falling into a local optimum and the like are solved, lesion segmentation can be carried out on multi-modal images of the PET-CT images, and an application range is wider. The method can be widely applied to the field of medical image processing.

The invention discloses a rotating machinery fault diagnosis method based on a one-dimensional depth residual convolution neural network. Firstly, the network learns more deep and abstract fault characteristics of training samples through a stacked one-dimensional residual module. Then, the Adam optimization algorithm is used to optimize all the super-parameters to complete the deep-level featureextraction and fault classification, and the rotating machinery fault diagnosis system model based on one-dimensional depth residual convolution neural network is obtained. Finally, the test samples are inputted into the trained fault diagnosis model, and the deep-seated features are extracted automatically to diagnose the health status of the rotating machinery.

The invention discloses a PM2.5 concentration prediction method and device and a medium, and relates to the technical field of pollutant prediction, and the method comprises the steps: building a PM2.5 prediction model based on a CNN and a bidirectional GRU neural network and based on a one-dimensional convolutional neural network CNN and a bidirectional GRU neural network; the meteorological training data tensor is sent to a PM2.5 prediction model for training; the one-dimensional convolutional neural network CNN respectively performs local feature learning and dimension reduction on each input variable time sequence, and forms a low-dimensional feature sequence through convolution and pooling operation in sequence; inputting the feature sequence into a bidirectional GRU neural network, and learning the feature sequence from the time positive sequence and the time negative sequence by the bidirectional GRU neural network; the meteorological test data tensor is sent to a trained PM2.5prediction model for prediction, and a PM2.5 prediction concentration value is obtained. According to the model, the speed and lightweight characteristics of the convolutional neural network and the sequential sensitivity of the RNN are effectively utilized, more data volume is allowed to be checked during training, and the prediction accuracy is improved.

The invention relates to an ultrasonic image intelligent segmentation method based on the automatic context and data enhancement. The method comprises steps that firstly, a series of image pre-processing is carried out for an ultrasonic image data set to acquire a data set after pre-processing; secondly, data enhancement for the data set after pre-processing is carried out, and the data set scaleis expanded to acquire an amplified data set; thirdly, the amplified data set is inputted to the convolutional neural network based on the automatic context, a model is trained in an end-to-end mode,and preliminary segmentation of the amplified data set is realized; and lastly, refinement post-processing on the preliminary segmentation result is carried out. The method has advantages of high segmentation accuracy, strong robustness and generalization and good segmentation edge smoothness, and the ideal segmentation effect can be acquired under the condition of limited training data sets.

The invention provides a weak supervision remote sensing target detection method based on hybrid hole convolution. According to the method, various customized designs such as hybrid hole convolution,attention mechanism and multi-layer pooling are adopted, multi-scale feature extraction and fusion are enhanced, and the robustness of objects of different sizes is improved. Besides, asynchronous iteration alternate training between a strong supervision detector and a weak supervision detector is utilized, training and detection can be carried out only through an image-level real label, and the purpose of cooperatively improving the detection performance is achieved.

The invention discloses an inter-frame difference and convolutional neural network fusion-based ship video detection method. The method comprises the four parts of preprocessing a video; obtaining anROI of each frame and extracting low layer features; obtaining high layer features of each frame of image by using a modified VGG16 network; and predicting a ship saliency map of the ROI of each frameand extracting a ship target. A relationship between continuous video frames is fully utilized; the interference of a background is reduced; a moving ship is accurately located; a ship moving regionis obtained; and compared with ship image saliency detection only using the low layer features, the method not only can be directly applied to the ship video detection but also reduces the situation of incomplete ship detection, has higher adaptability to a complex inland river moving ship scene, has higher detection precision, solves the problem of inaccurate inland river ship target saliency detection, and has extremely high practical application values.

The invention discloses a medical image classification method based on collaborative deep learning so as to solve the technical problem of poor classification accuracy of an existing medical image classification method. The technical scheme is characterized in that the method adopts a collaborative learning method between two deep convolution neural networks to carry out training in a mode of learning in pairs; each time, a model receives an image pair as input, and one pair of images are transmitted to the corresponding deep convolution neural networks respectively; the deep convolution neural networks are subjected to initialization and training through a pre-training model fine tuning method; a collaborative learning system is designed to allow the two deep networks to realize collaborative learning; and the collaborative learning is used for monitoring different or same attributes of the image pairs, that is, judging whether the image pair belongs to one category, and carrying out back propagation on collaborative errors generated by the two deep convolution neural networks in real time, and collecting network weight, so that the method can further enhance network learning feature representation capability, and can make an accurate judgment for easily-confused samples more effectively.

The invention discloses a bearing fault classification method based on CNN and Adaboost. A bearing signal is collected, the bearing signal is preprocessed, and a time domain signal and a time-frequency domain signal are extracted; a time-domain weak classification module and a time-frequency-domain weak classification module are constructed based on the time domain signal and the time-frequency domain signal; and then the time-domain weak classification module and the time-frequency-domain weak classification module are integrated and a membership probability value of a to-be-detected unmannedaerial vehicle bearing signal is predicted by using the integrated classification model. Therefore, the classification of UAV bearing faults is realized.

The invention discloses a width learning system based on multi-feature extraction, and the system comprises four sub-width learning systems, wherein each sub-width learning system comprises a featurenode, an enhancement node and a sub-node; each sub-width learning system firstly extracts an image feature from the image data set, combines the image features extracted from the image data set to obtain respective feature nodes, and then enhances the respective feature nodes through an enhanced mapping function to form corresponding enhanced nodes; after each sub-width learning system forms an enhancement node, the feature node of each sub-width learning system is combined with the corresponding enhancement node and then is connected to the sub-node of the sub-width learning system, and thenthe output of the sub-node of each sub-width learning system is normalized and then is connected to the final output layer. The method has the advantages of short model training time and high classification accuracy in the aspect of complex data set classification.

The invention discloses a remote sensing image change detection method based on a spatial-spectral feature fusion network. The method comprises the following steps: firstly, carrying out preprocessing operations of geometric correction and image registration on a remote sensing image; then inputting the training set into the DESSN network for training; and finally, inputting a test image into the DESSN network model, and outputting a segmentation result of dual-temporal remote sensing image change detection. According to the method, an asymmetric double-convolution module combined with Ghost is used for replacing an original double-convolution module in a U-Net network to enhance the feature learning capability and reduce the parameter quantity, and a difference enhancement module used for suppressing irrelevant changes caused by noise is added behind a feature extraction layer to enhance the attention on a changing target; and finally, a non-local space spectrum information fusion module is designed in a feature fusion stage for enhancing boundary integrity and internal compactness of a change object, high-precision change detection of the remote sensing image is finally realized, the change detection level of the remote sensing image can be effectively improved, and memory consumption is reduced.

The invention discloses a CNN-Bagging-based fault diagnosis method for a UAV bearing. A bearing signal is collected, the bearing signal is preprocessed, and a time-domain signal and a time-frequency-domain signal are extracted; a time domain weak classification module and a time-frequency-domain weak classification module are constructed based on an integrated learning algorithm according to the atime-domain signal and the time-frequency-domain signal; and then a membership probability value of a to-be-detected unmanned aerial vehicle bearing signal is predicted based on the time domain weakclassification module and the time-frequency-domain weak classification module. Therefore, the fault diagnosis of the UAV bearing is realized.

The invention discloses a method and a system for detecting faults of a point switch based on a convolution neural network algorithm. The system is improvement for a previous point switch fault detection system based on manual characteristic engineering. A convolution neural network is a deep learning model. A point switch power graph is input to the convolution neural network, and using characteristic learning capability of the convolution neural network, characteristics can be automatically extracted, to realize high-precision fault type detection.

The invention belongs to the technical field of digital image processing and pattern recognition, and discloses a gender recognition method and system based on multispectral fusion, a storage medium and a terminal. A camera with multiple wavebands is utilized for acquiring a face image and performing image preprocessing. A convolutional neural network module is used for subsequent face image feature learning; respectively pre-training the visible light and the infrared rays of each sub-band to obtain respective pre-training model parameters; connecting the network modules corresponding to thevisible light and the sub-band infrared rays in parallel, and adding a multispectral feature fusion layer at the tail end of the network; and adding a full connection layer behind the parallel fusiontype neural network for identification and classification, and performing retraining by using multispectral data to obtain a final gender identification result. When the method is implemented, specific fusion sub-bands can be selected and combined from five sub-bands of visible light, near infrared, short-wave infrared, medium-wave infrared and long-wave infrared; and the method has the characteristics of high precision and relatively high robustness.

The invention relates to a double-attention generative adversarial network based on channel enhancement and an image generation method. The network comprises a generator and a discriminator. The generator comprises a convolution block I and a double-attention mechanism module; the discriminator comprises a convolution block II and a double-attention mechanism module; compression activation operation layers used for obtaining channel attention through compression activation operation are arranged in the convolution block I and the convolution block II; the double-attention mechanism module comprises a position attention unit and a channel attention unit which are parallel to each other; the position attention unit establishes inter-position relevance based on a self-attention mechanism to obtain position attention features, and the channel attention unit establishes inter-feature channel dependence based on a channel attention mechanism to obtain channel attention features; and the double attention mechanism module fuses the position attention features and the channel attention features. According to the invention, the generation performance of the generative adversarial network canbe improved, the generated data distribution is closer to the original data distribution, and the generated image quality is better.

The invention discloses an unmanned aerial vehicle bearing fault diagnosis method based on CNN-Stacking. According to the method, first, bearing signals are collected, then the bearing signals are preprocessed, and a time domain signal and a time-frequency domain signal are extracted; second, a time domain weak classification mode and a time-frequency domain weak classification mode are constructed through an integrated learning algorithm based on the time domain signal and the time-frequency domain signal respectively; and last, after cascade fusion of the time domain weak classification modeand the time-frequency domain weak classification mode, a membership probability value of a to-be-detected unmanned aerial vehicle bearing signal is predicted, and therefore unmanned aerial vehicle bearing fault diagnosis is realized.

The invention provides a three-dimensional point cloud model training method and device for three-dimensional model construction. The method comprises the following steps of obtaining a set of training, dividing each group of three-dimensional point cloud data in the training set into a first set and a second set, inputting the point data in the first set into a preset three-dimensional point cloud model to obtain a first prediction set, inputting the prediction point data in the first prediction set into the same model to obtain a second prediction set, obtaining a first loss function valueaccording to a preset first loss function, obtaining a second loss function value according to a preset second loss function, calculating the first loss function value and the second loss function value to obtain a third loss function value corresponding to each group of point cloud data, and training a preset three-dimensional point cloud model according to the plurality of third loss function values corresponding to the plurality of groups of three-dimensional point cloud data, so as to construct the three-dimensional model according to the trained preset three-dimensional point cloud model.Therefore, the trained preset three-dimensional point cloud model improves the point cloud data feature learning effect.

The invention relates to a convolutional-neural-network learning method of multi-scale progressive accumulation. The method can be widely applied to the fields of machine vision and artificial intelligence such as target detection, target classification and target recognition. Firstly, average pooling operations are employed by the method to construct a multi-scale image pyramid on input images; and then images of different scales are progressively sent into a convolutional neural network, the convolutional neural network is enabled to carry out learning on the multiple images of the differentscales with progressive increasing of network depth and carry out progressive feature accumulation, and thus feature learning ability of the convolutional neural network is improved.

The invention discloses a monitoring method of the state of a gearbox bearing of a wind turbine generator system. The method comprises the steps of selecting variables by adopting a ReliefF feature selection algorithm, establishing an improved noise reduction self-encoding network to establish a relation model between the temperature of the gearbox bearing and influence variables thereof, reconstructing modeling variables in a monitoring stage by using the model, and predicting the temperature of the gearbox bearing; performing calculation according to a modeling variable reconstruction errorof normal operation data of the wind turbine generator system to obtain an exponentially weighted moving average control chart threshold; obtaining the fact that the unit operates normally if an EWMAcontrol chart statistic of the monitored unit is less than a threshold; and giving an alarm that the temperature of the gearbox bearing is abnormal if the temperature exceeds the threshold. The methodis used for analyzing the temperature data of the gearbox bearing, the goals of artificial intelligence monitoring and fault early warning of the temperature of the gearbox bearing of the wind turbine generator system are efficiently and accurately achieved, and the example analysis verifies the practicability and the universality of the method.

The invention discloses an image segmentation method for a kidney tumor, and the method comprises the following steps: S1, obtaining an abdomen scanning image, and dividing the abdomen scanning imageinto a data set and a training set; S2, performing down-sampling preprocessing on the acquired abdomen scanning image to obtain a scaled image; S3, determining an area of interest of the preprocessedimage in the step S2 by using global position information of the abdominal space, performing image segmentation, and performing training and prediction by a U-shaped kidney tumor segmentation network;S4, outwards expanding the abdomen scanning image in the step S1 for a certain range, segmenting images of the left kidney and the right kidney, interpolating all segmented images, and unifying the interpolated images into the same data distribution to obtain left and right kidney VOI images; S5, performing tumor segmentation prediction on the left and right kidney VOI images by a U-shaped kidneytumor segmentation network. Interference of other organs and tissues is effectively avoided, the accuracy of kidney tumor identification and image segmentation is improved, and efficiency is higher.

The invention discloses a remote sensing image ship target fine-grained classification method based on dynamic convolution, and the method comprises the steps: inputting a collected feature map into an attention module, and enabling the attention module to generate K normalized attention weight parameters; in the convolution processing of the collected feature map, using K parallel convolution forreplacing independent convolution in a parallel convolution kernel module, combining attention weight parameters and convolution kernels of the K parallel convolution to form a dynamic convolution layer, and finally connecting the dynamic convolution layer to a classification network for classification; fusing an attention mechanism into dynamic convolution, achieving multi-core integration, adding and fusing normalized attention weight parameters calculated through the attention module in the front in a non-linear mode, thus the feature learning capacity of the model is improved, and then the accuracy of ship target fine-grained classification is improved.

The invention belongs to the field of deep learning, and particularly relates to a face mask wearing condition detection method based on deep learning. The method comprises the following steps: obtaining the data of a to-be-detected image in real time, inputting the to-be-detected image into a trained mask detection network model, and obtaining a detection result; marking the to-be-detected image according to a detection result; the mask detection network model comprises a backbone feature extraction network model, a Neck network module and a Prediction network; the CSPDarkNet-X module is used in the backbone feature extraction network in the mask detection model, so that the feature extraction capability of the model can be enhanced, the parameter quantity of the model can be reduced, the structure of the backbone network of the model is simplified, and the feature learning capability of the model is improved.

The invention discloses a brain function network classification method based on a deep forest, and belongs to the field of non-neural network deep learning theory and application research. The methodspecifically comprises the following steps: initializing parameters, performing multi-granularity scanning to generate a multi-granularity cascade feature vector, generating a cascade forest structure, extracting advanced features, and finally calculating a prediction result. According to the method, deep learning and integrated learning are combined, so that the method has a strong feature learning capability of a deep model and a strong generalization capability of the integrated learning; the brain function network classification method has the advantages that when brain network data of high-dimensional small samples are oriented, rapid and accurate brain function network classification is achieved, hyper-parameters are few, the training time is short, the model generalization capacityis high, and the over-fitting problem of previous brain function network classification can be effectively solved.

The invention discloses a voice sound source positioning method using a microphone array. The voice sound source positioning method comprises the following steps: (1) generating a training sample to obtain a time-frequency domain signal, and acquiring a power envelope; (2) judging whether each time-frequency point of the time-frequency domain signal is a voice direct sound signal or not; (3) training a neural network of a UNET structure by using the sample generated in the step (1); (4) predicting a time-frequency point corresponding to the voice direct sound of a to-be-detected noisy signal by using a trained neural network of the UNET structure; and (5) when time-frequency point is judged to be the time-frequency point of the voice direct sound, applying a positioning method to obtain apositioning result. According to the voice sound source positioning method disclosed by the invention, the influence of interference and reverberation can be effectively removed in a high reverberation and high interference environment, and a result with high accuracy and robustness is obtained.

The embodiment of the invention provides an intelligent push processing method and system based on a block chain and big data mining, and the method comprises the steps: carrying out the deep learning optimization of an interest decision of a target interest mining model based on a cloud business linkage event; deep learning optimization stage allocation is performed on the cloud service linkage event log in the cloud service linkage event according to the feature learning cost, so that the feature learning effect can be improved and the interest mining precision can be improved when deep learning optimization of interest decision is performed on the target interest mining model; and thus, the precision of subsequently pushing the information to the related target intelligent virtual service user group is improved.

A CT image lung lobe recognition method based on a neural network is a difficult task in the field of traditional image processing due to the fact that lung fissure and lung parenchyma are low in CT value discrimination. By means of the CT image lung lobe recognition method based on the neural network, the network can learn more complex and high-dimensional features by means of the enough deep dimension and enough parameter quantity of the CT image lung lobe recognition method. Due to the introduction of the residual module, the complexity of the model is increased, and the model is endowed with higher feature learning capability. On the basis of the high feature extraction capability of the original U-Net, the subsequent multi-scale fusion channel can enable the model to learn the features of different lung lobe boundaries from coarse to fine. Therefore, the high-accuracy lung lobe recognition model is realized.

The invention discloses an industrial control intrusion detection method based on a multi-classification GoogLeNet-LSTM model. The industrial control intrusion detection method comprises the steps: firstly carrying out the classification of network packages for an industrial control communication process employing a Modbus protocol; then, detecting the network packets without information by usinga feature template comparison method; for a network packet carrying information, constructing a time sequence detection sequence by using original network packets, carrying out one-hot coding on eachnetwork packet, carrying out feature extraction by using GoogLeNet, and inputting an obtained feature vector sequence into an LSTM network based on an attention mechanism to carry out time sequence detection to obtain a detection result; and designing a detection result multi-classification method, and outputting specific intrusion categories by using two detection methods. The industrial controlintrusion detection method has universality, and has the characteristics of high detection precision and strong real-time performance for different types of invasion.

The invention discloses a bank customer data processing method and device. The method comprises the following steps: obtaining bank customer data which comprises one or any combination of personal data, transaction data and behavior data; and classifying the bank customer data according to the bank customer data and a pre-established classification model, wherein the classification model is pre-established according to a plurality of trained machine learning models, the plurality of trained machine learning models are selected from a machine learning model set by using a genetic algorithm, each trained machine learning model in the machine learning model set is provided with different hyper-parameters, and each machine learning model is trained according to historical data of bank customers. According to the invention, bank customer data can be processed conveniently, and high-accuracy and high-reliability customer data classification is realized.

The invention provides a face identification method based on a variable-speed learning deep auto-encoder network, and belongs to the mode identification field of the deep neural network. The method comprises the following contents: images comprise a training image and a to-be-identified image; the steps are as follows: firstly preprocessing the training image to obtain normalized data; secondly inputting the preprocessed training data into the deep auto-encoder network, guiding the layer-by-layer pre-training of the deep auto-encoder network through a variable-speed learning policy, and addinga classifier on the top layer of the network, further optimizing the network through fine adjustment to acquire an identification model; identifying the preprocessed to-be-identified face image, outputting an identification result, and counting an identification rate. The capacity of discovering the data substantive characteristics of the deep auto-encoder network is sufficiently utilized by themodel, and the characteristics learning speed and the network convergence speed are accelerated at the same time, thereby obtaining more optimal identification performance.

The invention relates to the technical field of image learning, particularly provides an image feature extraction method and device, and aims to solve the technical problem of how to improve the image feature learning effect. For this purpose, according to the method of the embodiment of the invention, image random transformation can be performed on each image sample in a training set to obtain one or more random transformation images corresponding to each image sample; classifying each random transformation image to form a first image set and a second image set; obtaining a quality evaluation value of each random transformation image according to each image sample by adopting an image quality evaluation model; performing model training on an image feature extraction model according to the first image set, the second image set and the quality evaluation value of each random transformation image; and carrying out image feature extraction on the target image by using the trained image feature extraction model. Through the steps, the model image feature learning effect can be improved.