40results about How to "Realize automatic learning" patented technology

The invention relates to a device for detecting network access abnormality based on data stream behavior analysis, comprising a flow information collection module, an abnormal behavior detection module and an abnormal flow processing module, wherein the flow information collection module is respectively connected with the abnormal behavior detection module and the abnormal flow processing module;and the abnormal behavior detection module is connected with the abnormal flow processing module. The invention also relates to a method for using the device. In the method, obvious abnormal flow data is filtered out firstly, then a network behavior model is used to detect the filtered flow data, and the network behavior model is automatically updated; and finally, the flow is blocked according to detection results. The device and method provided by the invention is utilized to establish a normal network behavior model. The model is compared with real-time data so as to detect whether real-time flow is abnormal; and the network behavior model is dynamically modified, abnormal flow sources are analyzed, and the abnormal flow is blocked, thus identifying the abnormal flow quickly and effectively and improving the accuracy of the detection.

The invention relates to a training-corpus quality evaluation and selection method orienting to statistical-machine translation. The training-corpus quality evaluation and selection method comprises the following steps of: automatic weight acquisition: adopting small-scale corpus to train an automatic weight acquisition model so as to obtain a characteristic weight and a classification critical value; sentence-pair quality evaluation: using the weight and the classification critical value as well as the original large-scale parallel corpuses as input, carrying out classification on the large-scale parallel corpuses by using a linear model for sentence-pair quality evaluation, and generating all corpus subsets; and high-quality corpus subset selection: on the basis of all the corpus subsets, considering the influence of the cover degree, and selecting the high-quality corpuses as training data of a statistical-machine translation system. The training-corpus quality evaluation and selection method has the advantages that richer sequence-pair quality evaluation characteristic is provided, so that the automatic learning of the characteristic weight is realized, and when the scale of the subsets reaches to 30%, the performance can reach 100%, even better; and the class of any input sequence pair can be divided, and help can be provided for tasks such as selection of high-quality corpus data.

The embodiment of the invention discloses a building extraction method based on a gating depth residual error optimization network. The method comprises the steps of obtaining an image feature combination of a high-resolution aerial image and airborne LiDAR point cloud data; the diversity of image samples is enhanced through the modes of random cutting, rotation, overturning and light and shade adjustment; using the improved deep residual convolutional neural network to automatically learn multi-level features of the image to obtain a rough building extraction result; a gating feature markingunit is adopted to screen and fuse effective features, and a high-quality building extraction result is obtained through successive up-sampling. By implementing the embodiment of the invention, a feature information gating transmission mechanism is combined with the deep residual convolutional neural network, and the method is used for building extraction of high-resolution aerial images and airborne LiDAR point cloud data.

The invention discloses a network encrypted traffic recognition method and device. The method comprises a preprocessing stage and a classification stage. In the preprocessing stage, original flow is subjected to flow segmentation, sampling, vectorization and standardization, a sampling scheme in large flow is provided, and the classification problem of the large flow is solved. In the classification stage, spatial feature capture and abstract feature extraction are performed by using a CNN, and then traffic time sequence features are learned by using stacked bidirectional LSTM on the basis ofabstract features so that automatic feature extraction and efficient recognition of encrypted traffic can be realized. The method has universality, can automatically extract the space-time features ofthe encrypted traffic without manual feature design of experts, and can adapt to traffic feature changes caused by different encryption technologies and confusion technologies.

The embodiment of the invention provides a human face recognition method and terminal equipment. Image statistics features are used for comparing target human face image frame sequences with picture sets in a human face sample bank, if the fact that a sample human face matched with a target human face is unique is know by judging and the proportion of picture frames which are successfully matched in the picture frame sequences is larger than a preset first threshold value and smaller than a preset second threshold value, picture frames which are not matched successfully are added into the human face sample bank and a picture set corresponding to a sample human face, wherein the second threshold value is larger than the first threshold value. An automatic learning and amending process of the human face sample bank is achieved, the problem that a user carries out complex amending operation is avoided, along with accumulation of using time, human face recognition accuracy can be higher, and the practicability of a human face recognition system is improved.

The invention relates to the technical field of image segmentation, and provides a segmentation method of an iris region in an iris image based on a Mask R-CNN neural network which comprises the stepsof S100, S100, establishing an improved Mask R-CNN neural network for iris segmentation S200, before training the neural network, marking a sample iris image set for training; S300, respectively inputting the labeled sample iris images into a neural network, and training the neural network until convergence; S400, inputting the iris image that needs to be segmented by the iris region into the improved Mask R-CNN neural network that completes the training, and acquiring the double circle boundary information and the binary mask map of the iris region; and S500, completing iris segmentation according to the double-circle boundary information of the iris area obtained in the step S400 and the binary mask map. The invention realizes automatic learning to obtain a normalized iris image and aniris region binary mask by using the improved Mask R-CNN neural network in iris segmentation and localization. Accurate iris area double-circle boundary information can also be obtained, and subsequent iris recognition operation is facilitated.

The invention relates to the field of cloud computing and computer networks, and specifically provides a method for adding public cloud network physical host into VPC. The method comprises: the physical host being connected to a ToR switch through a vlan, the ToR of the physical host supporting the vxlan and multicast, and expanding a plugin and an agent of the neutron; adding a plugin of l2 gateway into a new server, and expanding the function of the generic network switch plugin, so as to obtain the plugin of l2 gateway; and configuring vni mapping of vlan and vxlan of a ToR switch where a physical host is located, newly adding agent extension of l2 gateway in an openvswitch agent, and managing a virtual network bridge and an ovs flow table by a user. Compared with the prior art, the physical host and the VPC internal virtual machine can achieve two-layer and three-layer intercommunication through the VPC internal network address, unified management of the virtual machine and the physical host is achieved, and good popularization value is achieved.

The invention provides a method, a device and a system for processing a loop circuit in an access network. The access network comprises a plurality of nodes, each of which corresponds to an equipment layer DevLayer and an equipment network field identifier DevNetID. The method comprises the following steps of: receiving a detection message from a first port, wherein the detection message carries network field identifiers BaseNetIDs, and the DevNetIDs and the DevLayers of the nodes through which the detection message passes; acquiring the DevLayers of the nodes with the same DevNetIDs and the same BaseNetIDs from the detection message, and selecting the DevLayer having a maximum value from the acquired DevLayers to serve as a reference DevLayer; and comparing the DevLayer of the node with the reference DevLayer, and blocking or closing the first port when the DevLayer of the node is greater than the reference DevLayer.

The invention discloses a credit customer qualification classification method based on WOE conversion through machine learning. A system comprises a data preparation and preprocessing module, a modeltraining and evaluating module, a model deployment module, an inlet data processing module and a client qualification division module. The data preparation and preprocessing module is used for calculating original data I from the application data, the credit investigation data and the call record, calculating original data II through the customer category and the repayment data, and carrying out data preprocessing on the original data I and the original data II. The invention relates to the technical field of qualification classification. The credit customer qualification classification methodbased on WOE conversion through machine learning provides the system for realizing customer classification with different qualifications based on the machine learning method, the workload of manual auditing can be reduced, the approval efficiency is improved, learning is performed in time according to newly added customer information, self-adaption to customer qualification change is realized, the manual auditing efficiency can be improved to a greater extent, and the labor cost is reduced.

The invention provides a primary wiring diagram primitive recognition method based on a convolutional neural network, and the method comprises the following steps: (1) intercepting a primitive diagramfile from a primary wiring diagram training set picture, and constructing a primitive training set after processing; (2) training the primitive training set through a convolutional neural network togenerate a primitive recognition model; and (3) performing primitive identification on the input primary wiring diagram through the primitive identification model. The neural network technology and the image recognition technology are applied to the distribution network and the plant station primary wiring diagram, the primary wiring diagram primitive recognition method based on the convolutionalneural network is provided, automatic learning and recognition of the plant station wiring diagram are achieved, the system construction efficiency is improved, manual work is reduced, and a basis isprovided for follow-up intelligent processing of a topological model and ensuring the consistency of information interaction.

The invention provides a topology learning method, device and system of a one-way serial bus network. The topology learning method of the one-way serial bus network comprises the steps that a master node device sends a topology learning instruction message, wherein the topology learning instruction message is used for instructing a slave node device to read a node number of the slave node device in the topology learning instruction message, update the node number after one node is added, add the updated node number, the address thereof or the ID thereof to the topology learning instruction message, forward the topology learning instruction message after finishing the addition to a next-hop slave node device of the slave node device when the slave node device serves as an intermediate slave node device, and return the topology learning instruction message after finishing the addition to the master node device when the slave node device is the last-hop slave node device; and the master node device receives the returned topology learning instruction message and determines a topology structure of a one-way ring network based on the node number, the slave node device address or ID in the returned topology learning instruction message. According to the embodiment, the network topology structure can be learnt automatically, and the change of the network topology structure can be accordingly perceived.

The invention discloses an image feature fusion method and device, and a medium. The image feature fusion method comprises the following steps: inputting at least two images with an association relationship into a feature extraction network; respectively extracting feature information corresponding to each of the at least two images through the feature extraction network; providing the extracted feature information as point feature information of each node in the graph convolutional neural network; determining the fused feature information of each node through a convolutional neural network and based on the point feature information of each node and the edge feature information of the edge connected with each node; based on the fused feature information of each node, updating the feature information corresponding to each of the at least two images to obtain fusion feature information corresponding to each image, the edge feature information being a weight value indicating the degree ofdependence between two connected nodes, and the degree of dependence comprising at least three different levels.

The invention discloses a bird image recognition system and method based on big data training, and the system comprises an information processing module, a mobile terminal, an image database, a coarseimage data set, a candidate image set, and a model training module; the mobile terminal is connected with the information processing module, the coarse image data set is connected with the information processing module, the candidate image set is connected with the coarse image data set, the image database is connected with the candidate image set, and the model training module is connected withthe image database and the information processing module. The invention relates to the technical field of image recognition, and particularly provides a bird image recognition system and method basedon big data training, which establish a more accurate recognition model for bird image recognition by using a deep convolutional neural network. Compared with a traditional machine learning method, the deep convolutional network recognition model based on bird image features has the advantages that high-dimensional bird image features can be automatically extracted, and automatic learning of the bird image features is realized.

The invention relates to the technical field of credit default prediction, in particular to a machine learning interpretability-oriented credit default prediction method, which comprises the following steps of S1, acquiring data; s2, preprocessing the data; s3, dividing and training data; s4, verifying the model. The invention further discloses a system of the credit default prediction method oriented to machine learning interpretability, the system comprises a data acquisition module, the data acquisition module is connected with a cleaning and screening module through a signal line, the cleaning and screening module cleans the input data, and if a certain variable of the data is missing, a few non-core data is deleted; if the deletion amount is large, the data is filled by an overall distribution sampling method and a method of performing maximum likelihood estimation according to other information, and the cleaning and screening module is connected with a data division module through a signal, so that default prediction can be performed efficiently and accurately.

The invention relates to a split type robot for automatically collecting a respiratory tract specimen by teleoperation. The robot comprises a robot body, the robot body is provided with a fixed arm and an operation arm. The fixed arm comprises first X-direction guide rails distributed in pairs. A forehead pad height adjusting mechanism and a denture ring height adjusting mechanism are installed onthe first X-direction guide rails respectively. A forehead pad assembly is arranged on the forehead pad height adjusting mechanism. A pillow plate is arranged between the first X-direction guide rails. The operation arm comprises second X-direction guide rails distributed in pairs. A sheath and lifting mechanism and a hose fixing and propelling mechanism or a throat swab and connecting pipe are arranged on the second X-direction guide rails. An endoscope lens is installed on the sheath and lifting mechanism, the throat swab and connecting pipe is installed on the hose fixing and propelling mechanism, and a controller is electrically connected to the robot body. Therefore, under monitoring of the endoscope lenses and cooperation of the controller, local automatic operation or remote control can be achieved, and the exposure risk of medical staff during work is reduced or even avoided.

The invention provides a bill data processing method, which comprises: executing account checking operation on input bill data through an account checking network model so as to generate a difference state, and executing account balancing operation on the difference state through an account balancing network model so as to determine a difference processing result, wherein the account balancing operation is determined based on an account balancing strategy and a difference state generated for the account checking network model; determining a reward strategy based on the difference state and a difference processing result, wherein the reward strategy is positively correlated with the difference processing result; and optimizing the account balancing strategy based on the reward strategy so that the input bill data can be matched with the target bill data. In addition, the invention also provides a bill data processing system, a computer system and a computer readable medium.

The invention provides a video frame optimization method and device, and the method comprises: executing a preset number of times of optimal frame selection actions based on a currently used optimal frame selection strategy in an optimal frame selection strategy training process, and updating the currently used optimal frame selection strategy based on a return value corresponding to the preset number of times of optimal frame selection actions until a preset strategy updating completion condition is reached; and performing video frame optimization on the target video frame sequence to be optimized based on a trained optimal frame selection strategy and the state quantity of each target video frame in the target video frame sequence to be optimized. According to the method, the difficulty in design of a combination mode of comparison logic or comprehensive scoring can be avoided, automatic learning of face optimization is realized, and the face optimization performance is improved.

The present invention relates to the technical field of image segmentation, and proposes a method for segmenting iris regions in iris images based on Mask R-CNN neural network, S100, establishing an improved Mask R-CNN neural network for iris segmentation; S200, training neural networks Before the network, mark the sample iris image set used for training; S300, input the marked sample iris images into the neural network, and train the neural network until convergence; S400, input the iris images that need to be divided into iris regions into the In the improved Mask R-CNN neural network that has completed the training, the double-circle boundary information and binary mask map of the iris area are obtained; S500 completes iris segmentation according to the double-circle boundary information and binary mask map of the iris area obtained by S400 . The present invention uses the improved Mask R-CNN neural network in iris segmentation and positioning to realize automatic learning to obtain a normalized iris image and a binary mask map of the iris region, and can also obtain accurate double-circle boundary information of the iris region , to facilitate subsequent iris recognition operations.

The invention discloses intelligent casting grinding equipment with a learning function. The intelligent casting grinding equipment comprises a box body, an abrasive belt grinding mechanism, a laser grinding mechanism, a scrap collecting mechanism and a control mechanism, wherein the laser grinding mechanism is used for finely grinding a casting, and the laser grinding mechanism comprises a first support, a second support, a third support and two laser sensors. When the intelligent casting grinding equipment is used, an adjustable speed motor is started to drive a transmission wheel to rotate, the servo motor is started to drive a first transmission roller to rotate, the transmission wheel and the first transmission roller drive the polishing belt to conduct transmission, an operator attaches the surface of a casting to a grinding belt for rough grinding, and after rough grinding, the casting is placed in a containing table at the top of the support; and the positions of a first micro motor and a second micro motor are controlled through the laser sensor, the casting is finely ground through a first grinding rod and a second grinding rod, and the grinding quality of the casting can be improved by combining rough grinding with fine grinding.

The invention provides a resource generation model training and service resource generation method and device, and relates to the technical field of artificial intelligence such as natural language processing and deep learning. The training method of the resource generation model comprises the following steps: acquiring training data, wherein the training data comprises a plurality of state features and feedback tags of the plurality of state features; training the first neural network model by using the plurality of state features and the feedback labels of the plurality of state features to obtain an environment model; and training a second neural network model in a reinforcement learning mode according to the plurality of state features and the environment model to obtain a resource generation model. The service resource generation method comprises the following steps: acquiring a to-be-processed state feature; and inputting the to-be-processed state feature into a resource generation model, and taking an output result of the resource generation model as a service resource corresponding to the to-be-processed state feature.

The invention belongs to the technical field of disc shear tool detection, and discloses a disc shear tool changing opportunity prediction system and method based on shear blade wear detection, and the method comprises the steps: a detection system carries out the collection of a disc shear blade part image, carries out the feature extraction, and calculates the wear loss of a current disc shear tool; and the expert system calculates a tool abrasion loss-steel passing amount curve of the disc shear under different process parameters based on the collected data, and predicts the tool changing time of the tool under different process parameters. The detection system comprises a high-precision and high-speed industrial camera for acquiring an image of the cutting edge part of the circle shear, an image storage platform for storing the acquired image of the cutting edge part of the circle shear, and an image processing platform, according to the method, the workload of technicians is reduced, the accuracy of the disc shear tool changing time is improved, the time point is set to dynamically monitor the disc shear tool, and an accurate basis is provided for tool changing.

Disclosed are an image feature fusion method, device and medium. The image feature fusion method includes: inputting at least two associated images into a feature extraction network; extracting feature information corresponding to each of the at least two images through the feature extraction network; providing the extracted feature information as point feature information of each node in the graph convolutional neural network; and through the graph convolutional neural network, based on the point feature information of each node and the edge feature information of the edges connecting each node, Determining the fused feature information of each node, and updating the feature information corresponding to each of the at least two images based on the fused feature information of each node, so as to obtain the fused feature corresponding to each image Information, wherein the edge feature information is a weight value indicating the strength of the dependency relationship between the two connected nodes, and the strength of the dependency relationship includes at least three different levels.

The present invention proposes a method, device and terminal for controlling online social network information dissemination, including: constructing an information dissemination network, calculating offline features of nodes, extracting training samples from offline features, and training pre-built nodes through machine learning to disseminate influence Force prediction model, calculate the online features of nodes, provide offline features and online features as input to the node propagation influence prediction model, the node propagation influence prediction model makes predictions and output the node propagation influence prediction value; combined with the node propagation of the current node Influence prediction value and information of other nodes within the specified time window size determine whether to control the current node. The invention can adapt to a dynamically changing information dissemination network, and solves the problem of inaccurate node selection for information dissemination control in an online social network.

Provided are a method, device and system for processing loops in an access network. The access network includes a plurality of nodes, with each node corresponding to a device layer (DevLayer) and a device network domain identifier (DevNetID). The method includes the steps of: receiving a detection message from a first port, with the detection message carrying a network domain identifier (BaseNetID) and the DevNetID and DevLayer of the node that the detection message passes; acquiring the DevLayer of the nodes with the same DevNetID and BaseNetID from the detection message and selecting the maximum DevLayer as the reference DevLayer from the acquired DevLayers; comparing the size of the DevLayer of the present node with that of the reference DevLayer, and blocking or closing the first port when the DevLayer of the present node is greater than the reference DevLayer.

The present invention provides a topology learning method, device and system for a unidirectional serial bus network. A master node device sends a topology learning command message to instruct the slave node device to read the information of the slave node device in the topology learning command message. Node number, update the node number after adding 1 to the node number, add the updated node number, its own address or its own device ID to the topology learning instruction message, and when it is an intermediate slave node device, add The completed topology learning instruction message is forwarded to its own next-hop slave node device; when it is the last hop slave node device, it returns the added topology learning instruction message to the master node device; the master node device receives In the returned topology learning instruction message, the topology structure of the unidirectional ring network is determined according to the number of nodes, slave node device addresses or IDs in the returned topology learning instruction message. The embodiment of the present invention can automatically learn the network topology, so as to sense the change of the network topology.

The invention discloses a method for realizing credit customer qualification classification based on WOE conversion. The method comprises the following steps of data preparation and preprocessing, model training, model evaluation, model deployment, incoming data processing and a customer qualification division module. The invention relates to the technical field of credit. According to the methodfor realizing credit customer qualification classification based on WOE conversion, the noise influence is reduced, meanwhile, the dimensionality of non-numerical data conversion is smaller than thatof ONE _ HOT conversion, and the purposes of automatic model learning, more sensitive customer data change and higher prediction accuracy are realized.

The invention relates to the field of cloud computing and computer network, and specifically provides a method for a public cloud network physical host to join a VPC. The method connects the physical host to a ToR switch through a vlan, and the ToR of the physical host supports vxlan and multicast, and extends neutron Add the plugin and agent of the l2 gateway in the neutron server and expand the function of the generic network switch plugin, and configure the vlan and vni mapping of the ToR switch where the physical host is located, and add the agent extension of the l2 gateway in the openvswitch agent , the user manages virtual bridges and ovs flow tables. Compared with the prior art, the physical host and the virtual machine in the VPC can use the VPC intranet address to realize Layer 2 and Layer 3 interworking, realize the unified management of the virtual machine and the physical host, and have good promotion value.

The invention relates to a training-corpus quality evaluation and selection method orienting to statistical-machine translation. The training-corpus quality evaluation and selection method comprises the following steps of: automatic weight acquisition: adopting small-scale corpus to train an automatic weight acquisition model so as to obtain a characteristic weight and a classification critical value; sentence-pair quality evaluation: using the weight and the classification critical value as well as the original large-scale parallel corpuses as input, carrying out classification on the large-scale parallel corpuses by using a linear model for sentence-pair quality evaluation, and generating all corpus subsets; and high-quality corpus subset selection: on the basis of all the corpus subsets, considering the influence of the cover degree, and selecting the high-quality corpuses as training data of a statistical-machine translation system. The training-corpus quality evaluation and selection method has the advantages that richer sequence-pair quality evaluation characteristic is provided, so that the automatic learning of the characteristic weight is realized, and when the scale of the subsets reaches to 30%, the performance can reach 100%, even better; and the class of any input sequence pair can be divided, and help can be provided for tasks such as selection of high-quality corpus data.

The invention relates to a pepper leaf disease detection method based on improved AlexNet, and the method comprises the steps: building a disease data set, carrying out the data enhancement of the disease data set, carrying out the disease type classification of the image data of the disease data set, marking a corresponding label, building a model data set, and carrying out the detection of a pepper leaf disease. Dividing image data of the model data set into a training set, a verification set and a test set in proportion; secondly, constructing a convolutional neural network model, performing feature extraction on an AlexNet model in the convolutional neural network model, setting a multi-scale convolution kernel for a first convolutional layer, removing a full connection layer, replacing the full connection layer with a global average pooling layer, adding a BN layer into each convolutional layer, then setting hyper-parameters, and obtaining a multi-scale convolution kernel; and training the AlexNet model by using the training set. According to the improved AlexNet-based pepper leaf disease detection method provided by the invention, the model can be reduced, the identification precision can be improved, and the detection speed can be improved.