23 results about "Network composition" patented technology

The invention relates to an AMT signal denoising method and device based on self-adaptive multi-level U-Net. According to the method, noise segment classification is performed on a noisy AMT signal, a corresponding position mask of the clean AMT signal is recorded, and noise positioning information is provided for training of a denoising model, so that the training effect of the denoising model is improved, and loss of non-noise segments can be avoided; the method also uses two U-Net networks to form a U-Net cascade structure, the first network performs preliminary denoising, the second network processes fragment data output by the first network, generates a noise intensity probability by using a residual error, uses the noise intensity probability as a driving signal, distributes a corresponding weight value, and outputs the driving signal to the U-Net cascade structure. According to the method and the system, the second network is dynamically controlled to selectively absorb and fuse fragment data output by the first network, an area with remarkable residual noise can be adaptively concerned, complementation and optimization of a feature level are realized, and finally, the denoising effect of the denoising model on the noisy AMT signal is improved.

The invention discloses a self-supervised monocular depth estimation method based on an enhanced multi-scale pose network, and relates to the field of computer vision. The problems that in an existing self-supervision monocular depth estimation method, the pose network structure is simple, the time sequence modeling capacity is insufficient, and geometric constraints are missing are solved. According to the method, a self-supervision joint training framework composed of a depth estimation sub-network and an enhanced pose estimation sub-network is constructed; wherein the pose estimation sub-network extracts multi-scale spatial structure features through a layer-by-layer feature fusion encoder, and adopts a context fusion decoder based on time sequence attention to model a motion dependency relationship between continuous frames; meanwhile, a self-supervised pose consistency loss function is introduced, geometric continuity of a camera track is enhanced through forward and reverse transformation consistency constraint and closed-loop geometric constraint, and collaborative optimization of depth prediction and pose estimation is realized. The method is also suitable for the application fields of automatic driving, robot perception, augmented reality and the like.

The invention belongs to the technical field of communication, and particularly relates to a router system service exception self-healing method based on cloud AI, which comprises the following steps: deploying an exception detection module at a router end, collecting system logs and state data in real time and generating a standardized exception report; uploading the report to a cloud AI server through an encrypted MQTT protocol; the cloud calls a hybrid analysis model composed of a rule matching engine, a machine learning classifier and a reinforcement learning decision network to generate a self-healing strategy instruction packet; the router end receives and executes the strategy, completes service restart, configuration rollback or hotfix loading and other operations, and verifies the self-healing effect; when communication interruption exceeds a threshold value, an embedded loopback self-healing subsystem is automatically activated, and abnormity is independently handled based on a local strategy library. According to the technical scheme, millisecond-level abnormal response and high-success-rate autonomous recovery can be achieved, the network availability and the service continuity are remarkably improved, and the disaster recovery self-healing capacity is still achieved when the cloud end is disconnected.

The application provides a video summarization method based on multi-dimensional features and fine-grained hierarchical modeling, and relates to the technical field of video processing. In practical application, the video summarization technology can facilitate large-scale video retrieval and browsing. The method comprises the following steps: firstly, frame extraction is performed on an input video to obtain a frame sequence, and a multi-dimensional feature extraction network composed of a 2D network and a 3D network is used to extract multi-dimensional features; then, hierarchical temporal modeling is performed to complete the modeling process of the temporal dependence of the entire video sequence; finally, a regression network is used to obtain the importance score of each frame and generate a video summary. The application further explores the influence of the spatiotemporal features extracted by 3D feature extractors with different spatiotemporal complexities on the video summary result. The application shows excellent performance on the video summary datasets SumMe and TVSum. Whether from the application scene or the performance index, the application has strong practical value.

The application relates to a fine-grained traffic classification method based on an improved residual convolutional network in an SDN environment and belongs to the software technical field. In order to provide finer-grained application-aware traffic classification, let network operators better analyze network composition and manage and schedule network resources, fine-grained classification of the specific application programs is very important. Traditional methods tend to classify traffic based on protocols, which is coarse-grained classification. Inspired by the research in computer vision, the method of the residual convolutional network is applied to the identification and classification of network traffic. The method solves the network degradation problem that occurs in the process of fine-grained network traffic identification by traditional deep learning methods with the increase of network depth, can effectively learn deeper network features, and further realizes fine-grained network traffic classification.

The invention belongs to the technical field of image processing, and relates to a self-priori-guided image restoration method and system, and the method comprises the following steps: S1, constructing a damaged image data set; step S2, a step of network construction; step S3, a step of network training and optimization; step S4, a network performance test step; according to the technical scheme of the invention, aiming at the problems of content repetition, color distortion and the like which are easy to occur when a traditional image restoration method is used for processing large-area defect of a continuous region or reference content is insufficient, a two-stage image restoration method consisting of a semantic prior estimation network and a semantic prior guide restoration network is adopted; and residual semantic information in the damaged image is fully mined and utilized, so that stronger robustness is realized while the visual naturalness and detail accuracy of a restoration result are improved.

This invention relates to a robust image tampering localization method based on consistency-guided learning, belonging to the fields of image content tampering forensics and deep learning technology. The method employs an alternating training strategy based on epochs, constructing a consistency-guided learning framework consisting of a target network and a source network: the target network takes the unprocessed image as input and performs joint optimization through localization loss and contrast loss; the source network takes the post-processed image as input and learns by fitting the response of the target network. This invention improves the model's robustness to post-processing operations while minimizing interference with the model's learning of tampering traces.

The present disclosure relates to the field of computer vision image dehazing and discloses a single image dehazing method based on detail recovery. The method includes step 1: constructing a training dataset; step 2: constructing a backbone dehazing network to achieve preliminary image dehazing: based on the U-Net network model, introduce a residual module based on pixel attention mechanism in the encoding region, and an feature enhancement module in the decoding area; step 3: constructing a detail recovery network for image detail recovery, and introducing residual shrinkage module and spatial attention mechanism; step 4: training an overall network model composed of the backbone dehazing network and the detail recovery network; step 5: Testing. The present disclosure can effectively remove fog while improving the problem of edge detail information loss, reducing the blurring of the edges of the dehazing image, and generating a higher-quality dehazing image.

The application belongs to the technical field of decoding, and discloses an LDPC decoding system based on dynamic hypernetwork and two-stage learning, which is composed of a main decoding network and a dynamic hypernetwork, the dynamic hypernetwork is composed of a plurality of hypernetwork units corresponding to each layer of the main decoding network one by one, and the dynamic hypernetwork and the main decoding network are a controller-executor relationship; the dynamic hypernetwork dynamically outputs key parameters of the main decoding network for each iteration, and the main decoding network uses the key parameters to complete message passing and LLR updating; the input of the first hypernetwork unit is a log-likelihood ratio (LLR) value vector received by a channel, and the vector carries key information of a received signal; starting from the second hypernetwork unit, the input of the tth hypernetwork unit is the output of the t-1th layer of the main decoding network. The application significantly improves decoding performance, greatly improves decoding efficiency, enhances environmental adaptability, and reduces calculation and hardware costs.

This invention belongs to the field of remote sensing image processing and deep learning technology, specifically involving a refined training method for YOLO deep learning models. The method includes acquiring a remote sensing image dataset, preprocessing it and initializing the basic network, calculating the detection probability of each target category in the preprocessed remote sensing image, selecting target categories suitable for the current network based on the detection probabilities, and decoupling these target categories. Based on the weights of the current network, the network is dynamically expanded so that the expanded network's feature extraction layers can provide suitable feature learning capabilities for the undecoupled target categories. This process of network expansion and target category decoupling is iteratively performed during training until all target categories match the optimal network, ultimately forming an ensemble model composed of M networks with different number of layers. H =[ H 1 K1 , H 2 K2 ,..., H M Km ],in K m Let m be the feature extraction layer number of the m-th sub-network. This invention allows for flexible adjustment of parameters such as the initial network layer number and the number of extended layers, making it suitable for various target detection tasks.

The application discloses an anti-knowledge forgetting mutual guidance type semi-supervised medical image segmentation method, first, a double network model composed of two initialized different networks is constructed, and the internal difference is used as an implicit disturbance source; secondly, a double-path data fusion strategy is designed, structured fusion of region blocks of labeled and unlabeled images is carried out, a forced model is generated, and mixed images of two data domains are simultaneously processed in a single iteration, so that forgetting of learned labeled knowledge in the semi-supervised learning process is effectively inhibited; then, a mutual guidance learning mechanism is adopted, the prediction result of one network is used to generate high-quality pseudo labels, and a unified mixed supervision signal is constructed for the mixed images; finally, the model is trained by jointly optimizing a loss function customized for the mixed samples. The application is suitable for two-dimensional and three-dimensional image segmentation tasks, can significantly improve segmentation performance and robustness in a small amount of labeled data scene, and can enhance the universality and ease of use of the method.

The invention discloses a digital economic index prediction system and method based on big data, and the system comprises a multi-source data collection module which is used for obtaining digital economic related original data from a plurality of data sources; the data preprocessing module is used for cleaning, converting and standardizing the original data to obtain preprocessed data; the digital economic feature engineering module is used for extracting feature indexes from the preprocessed data; the MoE prediction model module is composed of a gating network and a plurality of expert networks and is used for performing digital economic index prediction based on the characteristic indexes; the Bayesian optimization module is used for dynamically adjusting hyper-parameters of the MoE prediction model; and the model evaluation and feedback module is used for evaluating prediction performance and optimizing a feature selection strategy. The invention belongs to the technical field of big data processing, and particularly provides a method for solving the problems of single data dimension, insufficient model adaptability, lack of multi-scale analysis and lack of real-time optimization capability in the prior art.

This application pertains to the field of spacecraft on-orbit servicing. It provides a robust model prediction and efficient configuration characterization method for space rigid-flexible coupled systems. The model prediction neural network in this disclosure consists of an LSTM encoder and a fully connected network with three hidden layers; the final output of the neural network is the net depth. By capturing the state variables of MUs in a past time series and the temporal mapping relationship between the net depth and the net depth at the next time step, an affine dynamic equation implicit in the net depth is established. Based on this, a cost function coupling formation and configuration information is constructed. The required control gain matrix is ​​obtained using a system identification method. Adaptive dynamic programming is used to approximate the optimal solution of the Hamiltonian function corresponding to the cost function, iteratively updating the network weight update rate during the process. Finally, the optimal control input is applied to the corresponding MU to achieve coordinated control of the TSNR system configuration and formation.

A new generation of network resource management architecture for coverage is provided, which constructs a structure of “signal coverage+capacity coverage” by incorporating satellite, space-based, and ground access points to achieve large-scale regional signal coverage and local area capacity coverage. In the core network, a global resource orchestrator is added, and in the access network, local resource orchestrator and virtual network management and maintenance unit are added for multiple functional units of resource management and control. In terms of coverage structure, it is divided into three-dimensional dense capacity coverage and three-dimensional global signal coverage. The former consists of ground dense network and aerial ad-hoc network, respectively realizing high traffic density coverage and large connection three-dimensional coverage. The latter consists of ground network and satellite constellation, respectively realizing extended coverage towards low altitude and three-dimensional coverage towards space, sky, ground, and ocean.

The invention discloses a metasurface robust reverse design method for a complex interference scene, and relates to the crossing field of metasurface reverse design and a deep learning technology. The method comprises the following steps: 1) constructing a data set containing metasurface phase distribution and an ideal far-field pattern; 2) constructing a series neural network architecture consisting of a global cross attention Transform reverse network and a pre-trained forward prediction network; 3) training the GCAT by using the training set only containing the non-interference data; and 4) inputting interfered actual far-field data into the trained GCAT network, and outputting target metasurface phase distribution. According to the method provided by the invention, high reconstruction fidelity can be kept in various complex interference scenes without training on interference data, the method has the advantages of high robustness, excellent generalization ability, high design efficiency and the like, and the problems of low reverse design precision and poor generalization ability in a complex electromagnetic environment in the prior art are effectively solved.

This disclosure provides a training method and apparatus for an object detection model, an object detection method and apparatus, a computer-readable storage medium, and an electronic device. The training method for the object detection model includes: using a backbone network composed of at least two sub-networks within the object detection model to be trained, performing step-by-step feature extraction on a sample image to obtain feature data, wherein the at least two sub-networks are arranged in order of data processing volume; using the detection network within the object detection model to be trained, obtaining the detection result of the target object in the sample image; determining a loss value representing the error between the detection result and pre-annotated information for the target object; and adjusting the parameters of the object detection model to be trained based on the loss value. This disclosure helps to more accurately detect target objects from images because the lower-level sub-networks can acquire finer-grained features from the image.

The application designs an optimization method suitable for a deep learning network for a classification task. Existing deep network optimization methods mainly use a large network as a teacher network to migrate knowledge to a student network. However, the Soft target is the Softmax output of the teacher network, and the Hard target is the dataset label, which is not suitable for a distributed student network. The application retains the basic framework of Teacher-Student knowledge distillation, but changes the Soft target to the feature output of the last convolutional layer of the Teacher network, comprehensively evaluates the influence degree of feature points on the final result according to the intra-class and inter-class fluctuations, and then filters and samples the feature points according to the composition of the distributed network. Finally, the spatial fluctuations of the Teacher-Student are counted into the loss function, and the overall training of the network is performed. The application can distribute the deep learning model to an embedded terminal in a distributed manner, and lays a foundation for the landing of deep learning.

This invention discloses a meta-gated anomaly detection method based on matrix attention segmentation and multi-branch decomposition convolution, belonging to the field of anomaly detection technology. The anomaly detection method of this invention includes: constructing a sample dataset; constructing and training a teacher-student network structure detection model, wherein the teacher-student network structure detection model adopts a three-level network structure composed of an expert network, a teacher network, and a student network, wherein the student network incorporates a rectangular attention-based segmentation network and a channel-parallel convolution decomposition network; and using the trained and optimized detection model to perform anomaly detection and localization on the image to be detected, i.e., outputting the detection result. This invention, by introducing a rectangular attention-based segmentation network and a channel-parallel convolution decomposition module into the student network, can effectively improve the accuracy of feature extraction and prevent information loss, thus contributing to improved anomaly detection accuracy.

The application discloses a kind of power network and communication network collaborative regulation method and device, method includes: obtaining flexible resource data information;According to the flexible resource data information, the regulation data value of flexible resource is calculated;According to equivalent power network composition structure, the output of power and communication coupling network equivalent model is calculated;Power network operation constraint set is constructed;According to the regulation data value of the flexible resource, the output of power and communication coupling network equivalent model and the power network operation constraint set, collaborative regulation target function is constructed, and the target of the collaborative regulation target function is power regulation net income and data communication net income maximum;According to the collaborative regulation target function, coupling network collaborative regulation is carried out.The application realizes network collaborative regulation, and improves collaborative regulation benefit.The application can be widely applied to power flexible resource scheduling technical field.

Described are examples for providing end-to-end intent definition of network functions for network slice management. Intents are defined for each level of network constituent including slices, slice subnets, and management functions. A system of intent based network slice management includes a network slice management function (NSMF) configured to receive a service profile from a communication service management function (CSMF) and derive an intent for each desired network slice subnet for a network slice subnet management function (NSSMF). The NSSMF is configured to derive requirements for a plurality of network functions (NFs) and provide an intent defining the requirements of a respective NF to a network function management function (NFMF). The NFMF is configured to receive the intent for the respective NF via an intent-based interface for management of NFs and derive a network resource model (NRM) for the respective NF based on the intent.

The application provides a passive device network and network group matching generation method, device, equipment, medium and computer product, generates a passive device network according to a specified matching parameter value, and the passive device network is composed of a first passive device network and a second passive device network which are connected in series or in parallel; the first passive device network and the second passive device network are both composed of a plurality of same type passive devices; a corresponding high-density distribution range is determined based on the distribution of the equivalent electrical parameter value; the matching parameter value is decomposed through the high-density distribution range; the parameter part falling into the range is divided into a decimal parameter, and the rest is divided into an integer parameter; the first passive device network sets a unit passive device of the integer parameter number; a series-parallel network matched with the decimal parameter is searched, and a second passive device network is constructed. The problem that the prior art lacks an automatic tool to efficiently and accurately design a passive device network with overall equivalent parameters meeting the expected design requirements is solved.

The invention discloses a high-reliability network-on-chip circuit architecture which is composed of three layers of networks-on-chip, one layer of network-on-chip is used for normal data transmission, the other two layers of network-on-chip are used for timeout retransmission, and when a master device sends a request and does not receive a response at specified time, a timeout retransmission mechanism is triggered. And the data retransmitted overtime can automatically change an on-chip network path for data transmission. Each layer of network of the network-on-chip circuit adopts a 4 * 4 2D Mesh NoC topological structure, has good expansibility, and can well prevent deadlock in combination with a classical XY-dimensional sequence routing algorithm. The network-on-chip is composed of a router and an arbiter, four ports of the router are used for cascade connection, and one port of the router is connected with a local device. The three-layer network-on-chip carries out data interaction through the distribution arbiter, it is guaranteed that data of the three-layer network-on-chip can be arbitrated to one path to be provided for local equipment, data of the local equipment is distributed to the appointed network-on-chip, and it is guaranteed that the three-layer network-on-chip can share the same inter-chip bus.

This invention discloses a method for predicting the remaining service life (RUL) of aero-engines based on a domain knowledge-enhanced dual-flow graph joint learning network, belonging to the field of engine life prediction technology. This invention addresses the problem that existing aero-engine RUL prediction methods do not comprehensively capture the complex degradation features of engines, thus affecting prediction accuracy. The method includes: first, constructing a static correlation graph representing the macroscopic physical topology between components, and simultaneously constructing a dynamic correlation graph of sensor parameters that evolves with the engine state; second, proposing a dual-flow graph joint learning network, consisting of a multi-hop graph attention network for static graph feature extraction and a gated graph-temporal attention network for dynamic graph feature extraction, efficiently extracting aero-engine degradation features from the heterogeneous static and dynamic graphs respectively; finally, employing a self-attention-enhanced feature alignment module to adaptively fuse the degradation features extracted from the two branches. This invention achieves more accurate and robust prediction of engine RUL.