183results about How to "Reduce model parameters" patented technology

The invention relates to an expressway traffic flow parameter prediction method and a system based on a gated neural network GRU. The method comprises the following steps: according to the high-speedroute information of the collected data and the longitude and latitude information of a toll station of a section, the research section data is initially screened, then the abnormal data is cleaned according to the manifestation of the abnormal data, and then the velocity time series is calculated in a certain time period, and then the missing data is filled in the missing time series, the filledspeed time series data is divided into training data and test data, and the traffic flow prediction model is obtained by training the training data, finally, the error analysis is carried out by usingthe predicted data and test data. The invention utilizes the advantage of GRU long-time memory data characteristics to obtain higher prediction accuracy, relatively less prediction model parameters and good portability, and can provide technical support for traffic guidance and traffic accident management and dispatching of traffic management departments.

The embodiment of the present application discloses a face key point positioning model training method and device, an apparatus and a storage medium. The training method includes: constructing a CNN model for face key point positioning, wherein the number of convolution layers of the CNN model is greater than a first threshold, and the number of channels of the convolution layers is less than a second threshold; using the CNN model to perform face key point positioning on training samples, and obtaining prediction positions of face key points, wherein the face key points include n types, and the n is an integer greater than 1; calculating respective loss function values corresponding to the n types according to prediction positions and real positions of each type of face key points, and calculating a loss function value of the CNN model; and stopping training the CNN model and saving the CNN model if the loss function value of the CNN model is less than a preset threshold. The embodiment of the present application reduces the model size by constructing the elongated CNN model while ensuring that the positioning accuracy is not lost as much as possible.

The invention discloses a computer aided model construction method based on deep learning gastric cancer pathological sections, and belongs to the technical field of artificial intelligence. The method uses a 121-layer dense-connected convolutional neural network to perform image recognition. A dense block structure in a DenseNet allows the high-level part of the network to acquire shallow features, which greatly reduces the over-fitting phenomenon. At the same time, the model has a large number of layers, which can fit more complex and smoother decision functions. Although the number of layers is large, the number of parameters of the model is not large, which saves resource consumption. In order to further avoid over-fitting, a training mechanism for migration learning is adopted. The model will be pre-trained on an ImageNet dataset to give the model a strong image feature extraction capability. The main optimization of the model during formal training can be better focused on how toextract the features of the diseased area, and the utilization rate of the data is greatly improved.

The invention discloses a method for constructing an elastoplasticity-damage coupling mechanical constitutive model of a rock material. The method comprises the following steps of obtaining the rock material on an engineering site, and manufacturing a standard cylinder sample; carrying out conventional triaxial compression mechanical tests under different confining pressures; obtaining a rock yield criterion, a plastic hardening criterion and a non-associated fluidity rule in combination with a test result; calculating a rock damage variable according to the stress-strain curve, and obtaininga rock damage evolution equation according to a damage variable-axial strain evolution rule; deriving a constitutive equation based on an elastic-plastic mechanics theory and an irreversible thermodynamic damage constitutive theory; combining the test data to obtain model parameters; writing the mechanical model into a UMAT subprogram, embedding the UMAT subprogram into ABAQUS finite element software, and carrying out triaxial test numerical simulation, so as to verify and improve the model. The method is clear in mechanical significance, simple in parameter acquisition, wide in application range and relatively high in accuracy.

The invention discloses an English word and case joint prediction method thereof based on neural machine translation, which mainly comprises the following steps of: establishing a training data set and making a vocabulary; converting the training data set into a vector training set according to the vocabulary; training the translation model, adding the predicted word loss and the predicted word case loss as the total predicted loss of the translation model; when the total loss is no longer reduced in the training process, stopping training of the model; translating Chinese using a trained translation model; after the completion of translation, according to the English translation and the corresponding word case attribute information, restoring the word in the translation to its due form, and obtaining the official translation. The method of the invention not only reduces the size of the vocabulary and the model parameters, but also improves the quality of the translation.

The invention discloses an integrated navigation system fault diagnosis method based on Gaussian process regression. The integrated navigation system fault diagnosis method comprises five steps: 1, collecting a sample, and setting a fault detection threshold; 2, initializing a Gaussian process regression model; 3, training the Gaussian process regression model; 4, inputting a system measurement quantity to the Gaussian process regression model so as to obtain a predicted value of Kalman filtering information and a variance of the predicted value; 5, constructing a fault detection quantity, comparing the fault detection quantity with a fault detection threshold, and judging whether a fault occurs. The integrated navigation system fault diagnosis method has the advantages of being easy to realize, being capable of giving the variance of a predicted output value, and the like, and provides an assurance to the reliability and accuracy of combined navigation. Model parameters are remarkably reduced, and hyper-parameters can be conveniently calculated through a numerical analysis method, thus the integrated navigation system fault diagnosis method has the advantage of being easy to realize.

The invention discloses a DBT (Digital Breast Tomosynthesis) image lump automatic detection method based on a Faster R-CNN (Faster Region-based Convolution Neural Network). According to method, lump automatic detection is carried out through utilization of an advanced Faster R-CNN; a new feature extraction network structure is designed; deep features are integrated; the deep features of DBT imagetargets are mined; model parameters are shared; the trained model parameters are reduced; the network overfitting degree can be mitigated; a detection effect is relatively good; and the detection average precision is improved.

The invention belongs to the field of electronic reconnaissance identification, in particular to a radar emitter signal modulation identification method combined with multi-dimensional feature migration fusion, comprising the following steps of generating nine kinds of radar signals to form a radar signal set; transforming the radar signal into time-frequency image by time-frequency transform; transforming the time-frequency image so as to meet the input requirements of the pre-trained large-scale network; sending the pre-processed time-frequency image to LeNet 5 network for feature extraction, and using the feature extraction module from input layer to form C5 convolution layer to output the feature extraction module; selecting a dimensionality reduction mode for the data obtained from the extracting feature step and processing the dimensionality reduction mode. The invention adopts the method of time-frequency analysis, maps the one-dimensional time-domain signal to the two-dimensional time-frequency domain, analyzes and processes the radar signal in the time-frequency domain, and has better effect for the non-stationary radar signal. The self-training network adopted by the invention has simple structure, and can improve the reliability of the system under the condition of low signal-to-noise ratio.

The invention discloses a lightweight multi-speaker voice synthesis system and electronic equipment. The system comprises a text feature extraction and normalization module, a speaker feature extraction module, a feature fusion module and a voice generation module. The text feature extraction and normalization module is used for carrying out encoding and feature extraction on to-be-processed textinformation by adopting a lightweight encoder, carrying out duration prediction on each word or phoneme corresponding to text deep features output by the lightweight encoder by adopting a lightweightduration prediction network, and carrying out length normalization processing to obtain regular text features with length equal to that of a target Mel spectrum. The speaker feature extraction moduleis used for generating features capable of representing a target speaker. The feature fusion module is used for fusing the features of the target speaker with the regular text features. The voice generation module is used for carrying out deep feature extraction, dimension mapping and residual error integration on the fused features and generating voice. The system supports multi-speaker voice synthesis and is high in synthesis speed.

The invention discloses a real-time pedestrian detection method and system based on deep learning, and the method comprises the steps: firstly obtaining video data, carrying out the size adjustment ofan inputted video image, and carrying out the feature extraction of the inputted image through depth separable convolution; performing up-sampling operation on deep features by a passthough layer structure in a network, performing feature fusion on the deep features and shallow features, and then outputting a deep feature map with low resolution and a feature map with high resolution, which fusescoarse-grained features and fine-grained features; and finally, carrying out regression and prediction on the two feature maps with different scales, and outputting a bounding box and confidence of each pedestrian detection result. According to the method, in an actual monitoring scene, the real-time pedestrian detection method based on the high-definition video, which meets the requirements of the real scene, is realized, and the detection efficiency is improved under the condition of ensuring the accuracy.

The invention belongs to the technical field of named entity recognition, and particularly relates to a Chinese named entity recognition method and device based on a RoBERTa-BiGRU-LAN model, and the method comprises the steps: converting a to-be-processed Chinese corpus into a word vector sequence; inputting the obtained word vector sequence into a first layer of BiGRU-LAN of a RoBERTa-BiGRU-LAN model, and obtaining a coding sequence fused with local information; inputting the obtained coding sequence into a second layer of BiGRU-LAN of the RoBERTa-BiGRU-LAN model, and obtaining attention distribution fused with global information; and obtaining a named entity identification result according to the obtained attention distribution. According to the improved word embedding method disclosed by the invention, Chinese representation is better carried out, and meanwhile, BiLSTM-CRF is improved into BiGRU-LAN, so that the parameters of the model are reduced, the complexity of the model is reduced, and the training time is saved.

The invention discloses a convolutional neural network automatic segmentation method and system for a mammary molybdenum target data set, which can obviously reduce model parameters and improve the practicability while ensuring the precision of a deep learning model on a mammary molybdenum target small data set. The method comprises the following steps: pre-training a convolutional neural big network on a mammary molybdenum target big data set; performing model compression on the trained convolutional neural large network by adopting attention transfer and knowledge distillation methods to obtain a convolutional neural small network; and carrying out fine tuning on the convolutional neural small network on the mammary molybdenum target small data set.

The application relates to the field of landform segmentation and identification technologies, in particular, to a machine-vision-based land use planning method and system, and an electronic device. The method comprises: collecting landform image data of a target area; constructing a convolutional neural network model based on a regional convolution neural network branch and an object area full-convolution branch; inputting the collected landform image data into the convolutional neural network model based on a regional convolution neural network branch and an object area full-convolution branch, extracting landform features of all landform objects in the landform image data by the convolutional neural network model, carrying out landform object classification and landform region segmentation based on the landform features; and determining landform composition of the target area based on the landform object classification and landform region segmentation results and carrying out land use planning on the target area. Therefore, lots of manual outdoor surveying and mapping work is reduced; the restrictions of application scenes are reduced; the application range is extended; and therecognition accuracy is improved.

The invention discloses a coal mine goaf crack identification method and detection system based on an unmanned aerial vehicle. The detection system comprises a camera, an unmanned aerial vehicle, an unmanned aerial vehicle ground station and a data server. The coal mine goaf crack identification method is characterized in that a deep semantic segmentation model is constructed through data augmentation processing in combination with deep semantic information of an image; a dense deep separable convolution unit is adopted, and image features are fully utilized, and multi-scale feature extractionof cracks is achieved in combination with a spatial pyramid; a loss function is adaptively set according to the weight of the crack in the training sample in the image, thereby accelerating the training process; and dense classification is adopted to finally obtain a pixel-level detection result. The coal mine goaf crack identification method has high crack detection precision and high training speed, can effectively reduce the inspection time and improve the detection reliability, is suitable for coal mine goaf surface crack detection under large-scale complex backgrounds, and can be popularized and applied to geological anomaly detection in other industries.

The invention discloses an establishing method of a coarse-grained-soil nonlinear elastic constitutive model. A large conventional triaxial solidification water-draining shear mechanical test result is used as a basis and relative compaction and damage ratio are used as media to establish the coarse-grained-soil nonlinear elastic constitutive model by combining with a unified disturbance degree function through deviatoric stress-axial strain relation curves, volumetric strain-axial strain relation curves, relative compaction-initial tangent modulus relation curves and relative compaction-peak stress relation curves under different confining pressures under different relative compaction conditions. The model has higher accuracy and wide usability and can be used for research and analysis on coarse-grained soil mass mechanical characteristics in actual geotechnical engineering.

A corn ear damage detection method comprises the following steps: obtaining a training set picture and a label thereof, labeling the training set picture and a real picture with the label, wherein thetraining set picture comprises a synthetic picture and a real picture, and the label comprises a target parameter and a number of a category to which a target belongs; modifying a network structure,re-clustering and calculating a prior frame, training a detection model by using the labeled training set pictures, and finely adjusting parameters of the detection model by using the labeled real pictures; and carrying out target detection, carrying out damage detection on the corn ear picture by using the trained detection model to predict positions and quantity of the damaged corn ears in the picture through an algorithm. The invention relates to a corn ear damage detection method based on deep learning, which is used for solving the problem of detecting corn ear damage in real time in a mechanical harvesting complex scene.

The invention discloses a quantum-behaved particle swarm optimization (QPSO) recurrent predictor neural network (RPNN) method for financial time series prediction, which relates to analysis and prediction on a time sequence. Chaos and phase space reconstruction theories are firstly applied, through a saturation correlation dimension (G-P) method, a chaotic financial time series attractor dimension is calculated, and the structure of the RPNN is determined; then, the RPNN is trained by the QPSO algorithm; and finally, the dynamic optimal weight and the threshold of the network are determined, and thus, the simulation prediction value of the RPNN and the actual value can reach the minimum error precision. The problem that optimization of the RPNN based on a gradient algorithm is easy to fall into local minimum can be solved, and the built QPSO-RPNN optimization prediction method can be widely applied in financial investment and social economy, and has the advantages that the convergence rate is quick, the searching is global, the programming is simple and efficient, and the prediction precision is high.

The invention provides a medical image segmentation method based on a lightweight full convolutional neural network. The method comprises the following steps: carrying out preprocessing such as graying, normalization, contrast limited adaptive histogram equalization (CLAHE) and gamma correction on a data set; randomly extracting patches from the training set and sequentially extracting patch graphs from the test set to complete data enhancement; building a full convolutional neural network architecture composed of a contraction path (left side) and an expansion path (right side), and designinga left-one-out training method for a data set with a small number of images; and finally, completing BN channel model cutting through channel sparse regularization training, cutting channels of whichscaling factors are smaller than a set threshold, finely adjusting the cut network to obtain a lightweight full convolutional neural network, and inputting test data into the network for rapid test to complete image segmentation. The lightweight full convolutional neural network not only ensures the advantage of high segmentation precision of the deep network, but also improves the test speed ofthe image segmentation network.

The invention discloses a reservoir incoming water quantity early warning and forecasting method and system based on medium and small basin flood forecasting, and the method comprises the steps: carrying out the interpolation of rainfall data in space, organically combining the grid rainfall data with the site rainfall data, and carrying out the model rainfall calculation; then, converting the potential evapotranspiration of the soil into the total evapotranspiration of the watershed; constructing a runoff production method library, and performing runoff production calculation by adopting different runoff production methods according to different climate types; combining the time area relation unit line of the drainage basin with the confluence speed, and obtaining the confluence flow process through a convolution formula by means of the surface net rainfall; fitting river section characteristic parameters, performing river flood calculation; and processing non-standard rainfall process data into a standard time period for parameter calibration. According to the method, the spatial heterogeneity of underlying surface attributes such as terrain and vegetation soil in the drainage basin is fully considered, the rainstorm flood process of the medium and small drainage basins is well simulated, and reliable technical support is provided for early warning of the water inflow of the reservoir.

The invention discloses a point cloud up-sampling method based on a GAN network. A GAN network composed of a Generator and a Display discriminator is mainly used for carrying out the point cloud up-sampling. The Generator is composed of an annular arrangement module, a multi-frequency pooling module and a GRU network module, and can effectively learn the geometrical characteristics and global characteristics of each point in the point cloud, thereby better mastering the geometrical information of the input point cloud. The discriminator Discriminator is composed of two layers of feature extractors and a deconvolution network module, can discriminate whether the input point cloud is true or false, and helps to better optimize the GAN network. Through annular arrangement, disordered point neighborhoods can be arranged into an ordered annular structure through orthogonal projection and counterclockwise arrangement, and geometrical characteristics of different groups of neighborhoods of each point can be accurately extracted through multi-frequency pooling. The method can be applied to the preprocessing step of three-dimensional point cloud reconstruction.

The invention discloses an infrared image conversion method, a living body detection method, a device, and a readable storage medium. The method comprises the steps: obtaining a visible light image and a near-infrared image; carrying out CycleGAN model training according to the visible light image and the near-infrared image; wherein generators of the CycleGAN model are two functions which are approximately reversible to each other, and the two generators share parameters in the training process; and inputting a target visible light image to the trained CycleGAN model to obtain a converted near-infrared image, and preferably inputting the near-infrared image to the living body detection model to obtain a judgment result. According to the technical scheme, the visible light image is directly converted into the near-infrared image, and the living body detection is carried out, so that the living body detection accuracy is effectively improved, and the attack of the prosthesis can be effectively resisted. A visible light image is converted into a near-infrared image by using a reversible network structure, parameter sharing is performed on a forward generator and a reverse generator by using an additive coupling technology, and the quality of the generated near-infrared image is superior to that of the near-infrared image generated by using a traditional CycleGAN method.