29results about How to "Fast training convergence" patented technology

The invention provides a virtual face generation method and system, electronic equipment and a storage medium, and the method comprises the steps: inputting a to-be-processed face image, and audio features and face key point features corresponding to the to-be-processed face image into a coding module in a face generation model, and obtaining image coding features and lip motion coding features; inputting the image coding features and the lip motion coding features into a feature fusion module in a face generation model to obtain fusion features; inputting the fusion features into a decoding module in a face generation model to obtain a virtual face image; wherein the face generation model is obtained by training an initial image generation model, the parameter scale value of which is smaller than a preset value, based on a sample real face image, a mask face image corresponding to the sample real face image, and sample audio features and sample face key point features corresponding to the mask face image. According to the invention, the virtual face generation quality is guaranteed, and the virtual face generation efficiency is improved.

The invention provides a robot navigation method based on multi-mode inverse reinforcement learning, which comprises the following steps: acquiring robot pose data, acquiring laser radar data, and generating a grid map; the laser radar data, the raster map and the robot pose information generate fusion observation representation through a multi-modal fusion network; initializing network parameters, taking the fusion observation representation as expert demonstration track data, and pre-training a generator and a discriminator in a generative adversarial inverse reinforcement learning framework through an expert demonstration track to obtain a pre-training model; standardizing a reward function through expert trajectory data; and constructing a generative adversarial reinforcement learning overall framework, training to obtain an intelligent control model, and deploying the model to realize intelligent control of the robot. According to the method, multi-modal data fusion and generative adversarial reinforcement learning can be realized, and the performance of the robot in the aspect of navigation is effectively improved.

The invention relates to the technical field of optical fiber sensing, and provides an event positioning method, system and device based on OTDR data and a medium, and the method comprises the steps: obtaining an OTDR data input condition, and generating adversarial network denoising; slicing the de-noised OTDR data and the variance and differential curve thereof, wherein the length is between one and a plurality of pulse widths; respectively inputting the sliced data into three feature extraction modules, wherein each feature extraction module is a series structure of a convolution layer, a batch normalization layer and an activation layer; combining the output features and inputting the combined output features into a high-efficiency channel attention layer to obtain an initial probability sequence; the initial probability sequence is binarized through a threshold value x, a binary sequence L is formed, and a point with the value being 1 in the binary sequence L represents that an event exists. According to the method, the problem that an OTDR event positioning algorithm is sensitive to a threshold value can be solved.

The invention relates to the technical field of voice processing, in particular to an emotion voice conversion method and device from source emotion to target emotion. The method comprises the following steps of: 1, acquiring voice data Voice1 with source emotion and voice data Voice2 with target emotion; step 2, carrying out feature processing on the Voice1 and the Voice2 respectively; step 3, inputting a processing result of the step 2 into a trained emotional speech conversion model for processing to obtain a reconstructed Mel spectrum NewMel3; and step 4, carrying out sound code reduction on the NewMel3 to obtain voice data Voice3 with target emotion. Through the designed emotional voice conversion model, effective decoupling of emotional information and non-emotional information can be realized, dependence on target emotional information of a converted speaker is not needed, and emotional voice with a better conversion effect can be obtained.

This application discloses a method and apparatus for constructing a multi-scale adaptive gating MambaPlus network, belonging to the field of artificial intelligence and machine learning technology. The method includes: initializing the network configuration and constructing the basic structure; preprocessing the input data to generate a standard dataset; mapping the input data to the hidden space via an input mapping layer, and extracting backbone features from the Mamba backbone; constructing at least two parallel scale branches in the hidden space to obtain a multi-scale feature set; inputting the backbone features and multi-scale features into an adaptive gating module, dynamically allocating weights and adaptively fusing them through a hierarchical gating mechanism to generate fused features; further enhancing the features through cross-scale attention and feedforward enhancement, and then superimposing the residuals to generate the final discriminative features; finally, completing category prediction and model training evaluation. This application, while retaining the advantages of Mamba's long-range dependency modeling, addresses the problems of insufficient utilization of multi-scale information, poor adaptive feature fusion, and low robustness in complex scenarios.

The embodiment of the invention provides a federated learning method based on a node selection strategy, electronic equipment and a storage medium, and relates to the technical field of federated learning. The method comprises the following steps: determining that client nodes in the federated learning system are not totally selected; obtaining an evaluation value corresponding to the client node, weighting the number of local training samples of the client node by taking a model of the evaluation value as a weight, and obtaining a selected probability of the client node according to the proportion of the number of the weighted local training samples in the sum of the number of the weighted local training samples in the federated learning system; selecting a plurality of client nodes according to the selected probability of each client node to obtain a training set of this round; and issuing model parameters to the client nodes in the training set in the round. According to the embodiment provided by the invention, the overall training time of the federated learning model is shortened in a mode of randomly selecting the nodes.

The invention discloses a rolling bearing fault detection method and system based on mechanism and data fusion, and relates to the technical field of equipment state monitoring and fault diagnosis, and the method comprises the steps: collecting a multi-source operation state signal and an operation condition parameter of a rolling bearing in real time; constructing and solving oscillatory differential equations under different fault types based on the signals and the parameters in combination with a rolling bearing dynamical model, and generating a mechanism feature library; denoising vibration signals in the multi-source operation state signals, and performing data reconstruction on the denoised vibration signals based on a gamma hybrid model; generating a high-dimensional data feature set according to the reconstructed vibration signal, and performing dimension reduction on the high-dimensional data feature set by adopting a matrix t-SNE algorithm to obtain low-dimensional data features; constructing a joint feature set according to the low-dimensional data features and the mechanism features, and calculating a health index; and the severity and position of the fault are judged based on the joint feature set and the health index, so that the precise detection requirement of the early-stage tiny fault of the rolling bearing under the complex working condition can be met.

The present application relates to the technical field of computer vision, deep learning and graphics rendering, in particular to a 3D portrait reconstruction method and system based on cascade diffusion prior and UV space physical constraint, which adopts a parameter efficient adaptation paradigm of freezing a unified DiT backbone and a cascade LoRA adapter, and sequentially completes UV space texture completion, illumination homogenization and intrinsic material decomposition; introduces a cross-intrinsic attention mechanism to realize multi-material pixel-level spatial alignment, and executes differential physical rendering supervision to constrain material physical consistency in the UV domain. The present application realizes 100,000-level field image generalization under less than 100 3D scanning samples, and the generated 4K PBR material assets have no baking light and shadow residues, and the physical behavior is consistent under re-illumination, which is suitable for high-fidelity 3D portrait asset production in the fields of digital human creation, film special effects and virtual reality.

According to the method for training the target model based on reinforcement learning, the target model is used for executing tampering detection on an input image, and the method comprises the steps that a training sample is obtained, the training sample comprises a target image and boundary information of a plurality of truth value boxes, and the truth value boxes are borders of a part of images marked in the target image and comprising tampering content; and inputting the target image into the target model, and outputting boundary information of the plurality of prediction frames. And calculating the overall overlap ratio of the plurality of prediction frames and the plurality of true value frames. Determining a first reward score based on a preset piecewise function according to a numerical range of the overall overlap ratio; the function corresponding to each numerical range of the piecewise function is a constant function, and the value of the corresponding constant function is increased along with the improvement of the overall overlap ratio. And based on a reinforcement learning algorithm, according to the first reward score, updating the target model.

This invention discloses an AI-based low-code runtime anomaly identification system, comprising a data acquisition module, a node runtime anomaly identification module, a process runtime anomaly identification module, and a low-code runtime anomaly management module. This invention relates to the field of data processing technology, specifically to an AI-based low-code runtime anomaly identification system. This solution innovatively proposes performing full-dimensional anomaly prediction before low-code deployment and operation, and conducting three-level progressive runtime anomaly identification, thereby improving the stability of low-code process operation. It introduces TT tensor decomposition technology to construct a Laplace matrix, and uses a dynamic correlation matrix matching the current node component type as the correlation weight for graph convolution operations, improving the accuracy of the low-code node anomaly identification output. It employs a sine and cosine periodic perturbation chaotic mapping strategy and introduces a dual adaptive step-size mechanism to improve the optimization algorithm, thereby enhancing the stability and identification accuracy of the process runtime anomaly identification model.

The present application relates to the field of mobile robot autonomous navigation and artificial intelligence, and particularly relates to a mobile robot navigation method, which comprises the following steps: constructing an SCAG-TD3 deep reinforcement learning network integrating target guidance gating and residual target injection; establishing an automatic course learning system comprising bidirectional adaptive promotion and demotion and forgetting perception review; constructing a social control barrier function based on an anisotropic ellipse social domain, introducing a forward-looking point model to convert it into a linear inequality constraint, and forming a safety filtering layer; and encapsulating and deploying the trained policy network and the safety filtering layer. The present application combines deep reinforcement learning with social perception safety constraints, solves the problems of insufficient perception focus, low training efficiency and no guarantee of social safety in the prior art in a dynamic crowd environment, and realizes efficient, safe and socially compliant autonomous navigation of the robot.

The invention belongs to the field of coal gangue grading recognition, and particularly discloses a coal gangue grading recognition system based on deep learning, which comprises a coal gangue image acquisition module, a coal gangue edge separation module and a coal gangue classification module. According to the scheme, gray level inversion is carried out on the initial image to improve the detail definition, a bounded sum operator is applied to realize migration of low-brightness area pixels to a high-brightness area, adaptive threshold processing is adopted to screen edge pixels, and complete and accurate extraction of coal gangue edges is realized; the method comprises the following steps: extracting texture features by adopting empirical wavelet transform, fusing mean value, contrast and homogeneity features calculated by a gray-level co-occurrence matrix, building a deep convolutional pulse neural network fused with a pulse coding technology, introducing an arithmetic optimization algorithm to optimize key parameters of a coal gangue classification model, capturing spatial features and time sequence pulse features of a coal gangue image, and constructing a deep convolutional pulse neural network fused with a pulse coding technology; and comprehensive mining and efficient fusion of the multidimensional key features of the coal gangue image are realized.

The embodiment of the present invention discloses a prediction method for a single-frame stripe image, including: using a Transformer‑based Channel Fusion (TCF) module to automatically identify objects in a single-frame stripe image, and fusing the overall shape and contour information of the object with the original single-channel image, wherein the TCF module is fine-tuned based on a pre-trained model Dense Prediction Transformer (DPT); extracting features from the fused image through a Res‑Unet module to obtain the depth information of the single-frame stripe image. The proposed TCF‑ResNet model realizes end-to-end prediction that directly maps the stripe image to a depth map. This method significantly simplifies the entire process and improves efficiency, making it more suitable for fast depth estimation and real-time three-dimensional reconstruction. In addition, with the application of neural networks, the anti-noise performance is also enhanced, making the prediction results under different conditions more reliable.

The invention discloses a kidney tumor subtype distinguishing method based on 3D deep learning. The kidney tumor subtype distinguishing method comprises the following steps of S1, obtaining patient data; s2, image processing and segmentation: downloading CT images of all patient data, carrying out format conversion, and then carrying out image segmentation and processing; step S3, image omics feature extraction; s4, depth feature extraction is carried out to obtain depth features, and then model construction is carried out based on radiomics features and the depth features; s5, obtaining an optimal distinguishing model, and predicting the type of the kidney tumor subtype through the optimal distinguishing model, the structure is novel, the conception is ingenious, the operation is simple and convenient, the depth feature can be well extracted, the benign and malignant kidney tumor subtypes can be well distinguished, and the method is suitable for popularization and application. The function of combining the prediction probability with clinical features, genetic markers and histopathology is added, and the method is used for carrying out auxiliary decision making on cases with medium confidence coefficient prediction in clinical practice.

This invention discloses a method and system for publishing multimodal data in industrial production for artificial intelligence, belonging to the field of industrial intelligence and data processing technology. The method includes: collecting multimodal data and performing hardware-level time synchronization and preprocessing to obtain a synchronized data stream; extracting multimodal features from the synchronized data stream in parallel and using an attention mechanism to fuse networks and generate a unified feature vector; constructing a five-dimensional quality assessment system and intelligently enhancing it; performing structured annotation and version management; adapting and publishing datasets according to AI tasks; implementing multi-level data anonymization and fine-grained access control, embedding invisible digital watermarks, and generating audit logs; and enabling multi-terminal access through a publishing platform. This invention solves the problems of difficult data alignment, poor quality, and low reusability, achieving secure data sharing and providing a high-quality, standardized data infrastructure for industrial AI applications.

The invention discloses a dense crowded pedestrian detection method and system based on a target detection model, and relates to the technical field of image detection, and the method comprises the steps: inputting a to-be-detected crowded crowd image into a crowded pedestrian detection model; the crowded pedestrian detection model comprises a backbone network used for extracting a multi-scale feature map of a crowded crowd image; the feature enhancement and refinement module is used for carrying out refinement and enhancement on low-layer features in the multi-scale feature map to obtain an enhanced multi-scale feature map; the hybrid encoder is used for encoding the enhanced multi-scale feature map and transmitting the encoded multi-scale feature map to each layer of the decoder; all layers of the decoder are connected in sequence, the last layer of the decoder is replaced by a prediction suppression module, and the prediction suppression module is used for screening out clear prediction and non-clear prediction and suppressing the non-clear prediction so as to output a prediction result of a target pedestrian; according to the densely-crowded pedestrian detection method, the detection capability of densely-crowded crowds is improved, and the detection speed is high.

A text recognition method comprises the following specific steps that S1, a text recognition model based on an improved CRNN structure is established, and output of the improved CRNN structure comprises a double-row output sequence branch and a single-row output sequence branch; s2, inputting a picture to train a text recognition model, inputting a single-row picture into the text recognition model to train a single-row output sequence branch, inputting a double-row picture into the text recognition model to train a double-row output sequence branch, and taking CTC Loss introduced with an editing distance as a loss function; and S3, inputting the to-be-identified picture into the text identification model, and simultaneously identifying a single-line text and a double-line text. According to the method, the CTC Loss loss function is improved by introducing the editing distance, so that the optimization process of the loss function is smoother. The training convergence speed is obviously improved, and the training precision is also improved to a certain extent. By improving the CRNN structure, both single-line text recognition and double-line text recognition can be realized.

The invention discloses a classroom facial expression recognition method based on an improved RT-DETR model. The method comprises the following steps: 1, obtaining a data set of classroom facial expression images; 2, preprocessing the data set by using a computer; 3, training a data set by using a classroom facial expression recognition detection model based on an improved RT-DETR model; 4, training the classroom facial expression recognition detection model to obtain an optimal detection model; and 5, performing real-time identification detection on the classroom facial expressions of the students in combination with monitoring and an optimal detection model. According to the invention, the CAFI module is designed in the backbone network, so that the classroom facial expression features can be extracted and screened more accurately; an MCFEEConv convolution feature enhancement module is designed in the head network, and the feature expression ability of faces with different distances and different sizes is improved; a RepC3LWGA module is designed in a head network, and joint modeling of local convolution features and global context features is achieved.

The invention relates to the technical field of electronic digital data processing, and discloses an industrial wastewater quality prediction method which comprises the following steps: acquiring historical time sequence data of key water quality parameters including an organic pollutant index, a nutritive salt index and a physical and chemical index, and preprocessing; performing periodic feature coding and linear trend feature coding on an input timestamp, and splicing all features into an enhanced feature sequence; inputting the enhanced feature sequence into a recurrent neural network for time sequence coding, and extracting a long-term dependency relationship of historical data; based on time sequence coding output, introducing an adaptive attention mechanism capable of learning a scaling factor to calculate the attention weight of each time step, and carrying out adaptive weighted fusion on the features; and processing the weighted and fused features, and outputting a plurality of key water quality parameter predicted values of a plurality of time steps in the future. The problems of insufficient utilization of time sequence features and weak multi-parameter correlation modeling are solved, and the purposes of multi-parameter joint prediction and adaptive attention adjustment are achieved.

The application relates to a progressive fine-tuning method and device for power scene model migration. The method comprises the following steps: obtaining a plurality of model parameters of a target model to be trained, determining the attention weights of each model parameter to each power knowledge entity in a pre-constructed power knowledge graph; mapping the attention weights corresponding to each model parameter into the activation scores of each model parameter; classifying each model parameter into a corresponding parameter layer according to the activation scores; obtaining a plurality of power corpora, and determining the first information entropy of each power corpus; classifying each power corpus into a corresponding corpus gradient according to the first information entropy; activating the parameter layer matched with each corpus gradient in the ascending order of the corpus gradient, training the activated parameter layer by using the power corpus of the corresponding corpus gradient, and obtaining the trained target model until the training stopping condition is reached. The method can improve the training efficiency and performance of model migration learning in the power scene.

The invention discloses an on-duty personnel fatigue behavior detection method and device and electronic equipment, and belongs to the technical field of image detection, and the method comprises the steps: inputting a current frame image into an action detection model, and extracting feature embedded information and first label information; inputting a subject position label in the feature embedding information and the first label information into a target tracking model to obtain a subject identity label and a tracking position label; and inputting the current timestamp and second label information corresponding to the fatigue action label, the main body identity label and the tracking position label of the operator on duty into a post-processing model so as to enable the post-processing model to judge whether the second label information meets the fatigue behavior definition in a preset time period, and if so, outputting third label information and a fatigue alarm signal. According to the method, three-section detection of the action detection model, the target tracking model and the post-processing model is provided, and the time sequence modeling capability is only added into the post-processing model, so that the whole model is lighter, and the operation difficulty of fatigue behavior detection is further reduced.

The invention relates to the technical field of gas measurement, and discloses a methane concentration measurement method and system based on a direct absorption spectrum, and the method comprises the following steps: collecting N groups of methane samples with different concentrations to obtain corresponding real methane transmission spectrum data; adding Gaussian noise into the simulated methane transmission spectrum to obtain simulated methane transmission spectrum data; improving the deep residual shrinkage network model to obtain an improved deep residual shrinkage network model, and constructing a methane concentration measurement model; training the methane concentration measurement model by using the training set data and the verification set data to obtain a trained methane concentration measurement model; acquiring actual methane transmission spectrum data in the target scene, and inputting the actual methane transmission spectrum data into the trained methane concentration measurement model to obtain methane concentration data corresponding to the target scene; according to the invention, the problems of large noise interference and low measurement precision of the existing methane concentration measurement method based on the direct absorption spectrum are solved.

The application discloses a soft-hard interbedded rock mechanical parameter prediction method and system based on a residual attention network, relates to the technical field of rock mechanical parameter prediction, and has the advantages that the traditional neural network is prone to gradient disappearance when processing a deep network, which influences the training effect; the existing method lacks an attention mechanism and cannot effectively identify and strengthen key features; and the modeling capability for interlayer interaction is insufficient; the application provides a soft-hard interbedded rock mechanical parameter prediction method based on a residual attention network, which comprises the following steps: obtaining structure parameters and target mechanical parameters of a soft-hard interbedded rock sample; converting the structure parameters into an enhanced feature vector; constructing a residual attention network model; inputting the enhanced feature vector into the residual attention network model; and outputting a mechanical parameter prediction result and reliability evaluation information.

The application discloses a flame and smoke identification method, and relates to the technical field of computer vision, and the method comprises the following steps: generating a boundary box of a flame and smoke target through a YOLOv target detection framework; extracting multi-scale visual features by using an image encoder of a SegmentAnything Model (SAM); encoding boundary box coordinates into a prompt embedding vector; fusing the visual features and the prompt embedding by using a mask decoder to generate a segmentation mask and a confidence score; confirming the target when the score exceeds a threshold value, and superimposing the boundary box and the segmentation mask and outputting to a fire-fighting early warning system.

The invention provides a tomato shelf life prediction method based on a hyperspectral and RGB imaging technology. The tomato shelf life prediction method comprises the following steps: S1, obtaining a hyperspectral image and an RGB image of a tomato sample; s2, preprocessing the hyperspectral image to obtain hyperspectral feature data; s3, extracting color feature data and texture feature data of the RGB image; s4, splicing the data obtained in the steps S2 and S3, performing standardization processing, and dividing the data into a training set, a test set and a verification set; s5, constructing an MLP model; screening the target feature set; s6, training the MLP model to obtain a tomato shelf life identification model; s6, obtaining a hyperspectral image and an RGB image of the detected tomato, performing preprocessing, and screening data according to the target feature set; and S7, splicing the screened color feature data, texture feature data and hyperspectral feature data of the detected tomatoes, inputting the spliced data into the tomato shelf life recognition model for recognition, and outputting the shelf life of the detected tomatoes. According to the invention, the accuracy of tomato shelf life identification is improved.

The application discloses a point cloud target detection model training method and device based on asymmetric feature consistency, relates to the technical field of automatic driving perception and computer vision, and comprises the following steps: acquiring source domain point cloud data, inputting a model to obtain a target detection loss, and respectively extracting geometric and intensity features; generating enhanced geometric features by disturbing the geometric features, and respectively mapping the geometric features, the enhanced geometric features and the intensity features into high-dimensional original geometric feature representation, enhanced geometric feature representation and intensity feature representation; determining geometric self-consistency loss based on the original geometric feature representation and the enhanced geometric feature representation; performing a stop gradient operation on the original geometric feature representation, and determining asymmetric feature alignment loss in combination with the intensity feature representation; and finally weighting and integrating the above losses into a total loss to update model parameters. By establishing one-way gradient blocking, the application not only utilizes intensity semantic auxiliary geometric feature learning, but also completely blocks noise pollution caused by cross-domain intensity deviation, and significantly improves the single-source domain generalization capability of the model.

The present invention discloses a target detection acceleration method based on structural re-parameterization. For a convolutional neural network target detection model, a re-parameterizable detection head structure is constructed. After training, the constructed multi-branch decoupled detection head uses channel segmentation to achieve feature alignment, that is, the channel segmentation layer divides the input feature into two parts, corresponding to classification and regression tasks respectively. Experiments show that the improved network training convergence speed and accuracy are improved; the constructed multi-branch decoupled detection head structure can be equivalently converted into a single-channel structure during actual deployment and inference, and the inference speed is improved in terms of parameter quantity, calculation amount and calculation density; finally, the target detection head realizes group convolution fusion by filling the convolution kernel, and there is only one 3×3 convolution layer after fusion, and the calculation density and parallelism have significant advantages over the existing design.

Embodiments of the present application provide a node selection strategy-based federated learning method, an electronic device and a storage medium, and relate to the technical field of federated learning. The method comprises: determining that a client node in a federated learning system will not be selected; obtaining an evaluation value corresponding to the client node, weighting the number of local training samples of the client node with the modulus of the evaluation value as the weight, and obtaining the selection probability of the client node according to the proportion of the weighted number of local training samples in the total sum of the weighted number of local training samples in the federated learning system; selecting a number of client nodes according to the selection probability of each client node to obtain a current round training set; and issuing model parameters to the client nodes in the current round training set. The embodiments provided by the present application reduce the overall training time of the federated learning model in the random node selection manner.

The present application relates to the technical field of robot motion control, and especially relates to a kind of quadruped robot reinforcement learning motion controller training, simulation, deployment method based on elevation map, the method constructs the asymmetric Actor-Critic deep reinforcement learning model based on double-estimator hybrid explicit and implicit feature state estimation, and the body state estimator and the terrain state estimator are respectively extracted body perception implicit feature and foot end terrain perception implicit feature;Build multiple simulation training terrains and design course teaching mechanism and multiple types of reward functions for training;Respectively build simulation verification environment based on MuJoCo and Gazebo, construct local elevation map by ray detection or laser radar point cloud fusion, and splice as the strategy network input with the body state of robot, the present application effectively improves the motion stability, environmental adaptability and simulation to reality migration reliability of quadruped robot in complex terrain.