87results about How to "Enhance feature expression" patented technology

This invention provides a method and apparatus for correcting sea surface temperature (SST) predictions based on spatiotemporal axial attention. The method includes inputting target SST data into a target SST prediction correction model, comprising a convolutional input layer, an encoder, and a decoder. The convolutional input layer performs feature extraction, temporal encoding, and positional encoding on the target SST data to obtain a target feature vector. The encoder performs attention calculations on the target feature vector in three dimensions to obtain a first target output vector. The decoder outputs the target SST prediction result based on the first target output vector and historical prediction values ​​output by the decoder. Temporal and positional encoding are performed during the model input stage to enhance the representation of temporal and positional information contained in the original data. By performing attention calculations on the target feature vector in the spatiotemporal, longitude, and latitude dimensions respectively by the encoder, effective fusion of features in different dimensions is achieved, further improving feature representation capabilities and increasing the accuracy of SST prediction.

The invention discloses a PHSM-YOLO-based multi-target wearing identification method. The PHSM-YOLO-based multi-target wearing identification method comprises the following steps: constructing a PHSM-YOLO target detection model used for identifying whether a constructor correctly wears a safety appliance; collecting a target image containing constructors in the construction environment; and inputting the target image into a PHSM-YOLO target detection model, and outputting a wearing identification result of the safety wearing tool of each construction worker in the target image by the PHSM-YOLO target detection model. According to the multi-target wearing identification method, the wearing of the safety wearing tool can be identified on the target image containing the construction personnel through the constructed PHSM-YOLO target detection model, so that the personal safety of the construction personnel is guaranteed.

This invention relates to the field of smart agriculture information technology, and discloses an agricultural pest and disease prediction system and method based on multimodal data fusion and edge AI, comprising: a data acquisition module, an edge AI analysis module, and a communication interaction module; the data acquisition module is used to collect agricultural monitoring data containing agricultural environmental information and crop status information; the edge AI analysis module is used to call a pest and disease prediction model that integrates multimodal agricultural feature extraction to perform multimodal joint feature extraction on the agricultural monitoring data containing agricultural environmental information and crop status information, and infer pest and disease prediction results based on the multimodal joint features, and generate pest and disease early warning information based on the pest and disease prediction results; the communication interaction module is used to upload the agricultural monitoring data collected by the data acquisition module and the pest and disease prediction results and pest and disease early warning information output by the edge AI analysis module to a cloud platform. This invention enables agricultural pest and disease prediction on resource-constrained edge devices.

The invention discloses an ocean sub-mesoscale process instance segmentation method based on frame supervision, and relates to the field of ocean remote sensing information processing. The invention aims to solve the problem of low segmentation precision of the existing frame supervision instance segmentation method. The method comprises the following steps: forming a first training set by using a chlorophyll concentration remote sensing image and a frame label of an ocean sub-mesoscale process; training a teacher model by using the first training set, and outputting an initial prediction mask set by the trained teacher model; performing mask correction on the initial prediction mask set to obtain a pseudo mask; forming a second training set by using the chlorophyll concentration remote sensing image, the pseudo mask and the frame label of the ocean sub-mesoscale process, and training and optimizing the student model by using the second training set to obtain a trained student model; and inputting a chlorophyll concentration remote sensing image to be tested into the trained student model to obtain an ocean sub-mesoscale process instance segmentation result. The method is used for obtaining the ocean sub-mesoscale process region.

The application belongs to the technical field of target detection, and discloses a UAV image target detection method based on an improved RT_DETR model, which introduces an FAPPA module in a backbone network, effectively alleviating the information loss problem of small target features in a deep network. The method also replaces the feedforward neural network of the AIFI module in the encoder with an EDFFN module, which selectively enhances high-frequency features such as edges and textures by using a patch_wise FFT strategy and learnable frequency domain weights. In addition, the method also uses a Zoom_cat module to realize adaptive alignment of multi-scale features, and uses the convolution_attention dual-path architecture of the CAFMFusion module to fuse local details and global semantics, thereby fully utilizing complementary information at different levels. The method can significantly improve the UAV image target detection capability and also identify small target features in a complex background.

The invention relates to the technical field of three-dimensional sensing, in particular to a point cloud self-supervised learning method based on multi-feature sensing and auxiliary reconstruction. The method comprises the following steps: S1, grouping point clouds of a random mask strategy; s2, local mark multi-feature fusion embedding is carried out; s3, pre-training the backbone network of the auto-encoder; s4, performing GaussPoint enhancement in a fine tuning stage; s5, auxiliary branch reconstruction in a fine tuning stage; and S6, carrying out multi-task coverage classification segmentation evaluation. According to the method, a Point-MAR network model is constructed on the basis of a mask self-encoding normal form by combining a multi-feature perception fusion embedder, an auxiliary reconstruction branch and a GaussPoint data enhancement method, feature expression, geometric perception and generalization capabilities are remarkably improved, and the method shows good competitiveness compared with numerous mainstream models in downstream tasks.

The invention discloses a power equipment fault diagnosis method and system based on multi-source data fusion. The method comprises the following steps: extracting linear features of data sources from preprocessed multi-source data through an improved principal component analysis (PCA) algorithm; fusing the linear features of the multiple data sources through a weighted fusion strategy, and constructing a fault diagnosis model in which a bidirectional long-short-term memory network BiLSTM and a convolutional neural network CNN are mixed based on an attention mechanism; and the pre-trained fault diagnosis model is used to carry out power equipment fault diagnosis, and a fault type identification result and a fault degree evaluation value are output. According to the invention, through multi-source data acquisition, hierarchical fusion and a hybrid diagnosis model, accurate and real-time diagnosis of power equipment faults and quantitative evaluation of fault degrees are realized; according to the method, the BiLSTM and CNN hybrid model based on the attention mechanism is adopted, the spatial information and the time sequence information of the features are captured at the same time, key features are highlighted, and the accuracy of fault diagnosis is improved.

The invention discloses a wake detection model construction method based on frequency domain multi-scale perception, which relates to the technical field of ocean observation, is used for capturing weak wake features to carry out sea surface internal wave wake detection, and comprises the steps of constructing a wake image data set, constructing a wake detection model by taking an RT-DETR model as a baseline model, and constructing a multi-scale texture information selection network. A backbone network of the baseline model is replaced, a frequency domain multi-scale attention perception module is constructed, an AIFI module of the baseline model is replaced, a golden cudgel convolution down-sampling module is adopted to replace a standard convolution down-sampling module of the baseline model, training and optimization are performed on the wake detection model based on the wake image data set, and a target detection result is obtained. According to the method, the wake detection model AUVRT-DETR is constructed, the detection precision is improved while the advantage of light weight is kept, the optimal balance is achieved between the detection precision and the calculation efficiency, and the model parameter quantity and the calculation overhead are remarkably reduced while the highest detection precision is achieved.

A method and system for super-resolution reconstruction of infrared images based on convolutional neural networks, relating to the field of electronic digital data processing, is disclosed. The method includes: inputting a low-resolution infrared image into a preset convolutional neural network to obtain multi-layer feature maps of the low-resolution infrared image; scaling the feature maps of different layers in the multi-layer feature maps according to a preset ratio and then stitching and fusing them to obtain a fused feature map; generating feature vectors from the fused feature map using global average pooling, and transforming the feature vectors using a fully connected layer to obtain attention weights; weighting the attention weights with the fused feature map to obtain an enhanced feature map; and performing upsampling reconstruction processing on the enhanced feature map using an upsampling structure, introducing a residual connection structure during the upsampling reconstruction process to generate a high-resolution infrared image. Implementing this method generates high-resolution infrared images with more detail.

The invention discloses an electric welding machine intelligent operation and maintenance management method and system based on machine learning, and belongs to the technical field of operation and maintenance management, and the method comprises the steps of electric welding machine operation and maintenance data integration, electric welding machine fault detection, electric welding machine residual life prediction and electric welding machine intelligent operation and maintenance management. According to the scheme, the original importance score is calculated based on the Gini impurity reduction amount, the weighted importance score of the features is obtained in combination with the fault correlation coefficient, the key feature set is constructed based on the accumulated score, the comprehensive confidence of the decision tree is obtained, the class weight compensation factor is introduced for correction, and the accuracy and reliability of fault detection are improved; based on the calibrated degradation complexity, a dynamic window increment is calculated, a multi-scale window set is constructed, a multi-scale sample is generated, a full connection graph is constructed, and weighted frequency domain features are obtained in combination with the degradation stage, so that the prediction precision of the remaining service life is improved, and a reliable basis is provided for operation and maintenance management of the electric welding machine.

This application belongs to the fields of artificial intelligence and finance, and relates to a semantic-based text classification method. The method includes inputting a training sample set and a knowledge graph into a knowledge-enhanced language model to obtain knowledge-enhanced text semantic feature vectors; inputting the text semantic feature vectors into a capsule network model to output classification prediction results; calculating the loss value between the predicted classification result and the classification label; adjusting the model parameters based on the loss value to output a model to be validated; validating the model to be validated using a test sample set to obtain a text semantic classification model; and inputting the text to be classified into the text semantic classification model for classification. This application also provides a semantic-based text classification device, computer equipment, and storage medium. Furthermore, this application relates to blockchain technology, allowing the classified text dataset to be stored in the blockchain. This application can effectively identify multi-labeled text, improving the efficiency and accuracy of text classification.

The invention discloses a target tracking algorithm-based sowing seed hole-falling counting method of a sowing machine. The method comprises the following steps of A, acquiring a sowing seed hole-falling video in a sowing process of the sowing machine in real time; b, constructing a sowing grain hole-falling data set based on the hole-falling video; c, constructing a sowing seed detection model for detecting and positioning sowing seeds; d, training to obtain a trained sowing grain detection model; and E, performing real-time target detection on the seeding grain hole-falling video to be subjected to hole-falling counting from a first frame by using the trained seeding grain detection model, and calculating the inter-frame grain distance of the same grain among different frame numbers by combining a target tracking algorithm from a second frame, when the inter-frame grain distance meets a set threshold condition, continuous tracking is carried out, and hole falling event judgment and counting are completed. According to the invention, the sowing grain hole falling condition in the sowing process can be accurately detected in real time, the non-sowing position and the sowing quantity are fed back, and the sowing precision and efficiency are improved.

The invention discloses a complex weather small target detection system based on multi-scale image enhancement and dynamic domain adaptation, and the system comprises an LMIENet image enhancement module, an SC2f deformable convolution feature extraction module, and a DDA dynamic domain adaptation module, and the detection method of the detection system comprises the following steps: a, collecting a plurality of types of weather scene images, and constructing an original image data set; b, data cleaning and refined labeling; c, performing image enhancement preprocessing; d, dividing a structured data set; e, training and calibrating an LMIENet image enhancement module; f, constructing a detection network integrated with deformable convolution; g, introducing joint training adapted to a dynamic domain; h, pipeline construction and end-to-end optimization; and i, system performance verification. According to the method, multi-scale feature fusion, complex weather preprocessing and dynamic domain self-adaption are integrated, and the detection precision and robustness of the small target under the complex weather condition can be effectively improved.

The application discloses a kind of multi-robot navigation reinforcement learning methods based on buffer Voronoi unit, it is related to multi-agent navigation technical field.Based on the current position information of robot, the corresponding Voronoi unit is constructed, and each boundary of Voronoi unit is translated inwardly by the distance of the physical radius of robot, to generate buffer Voronoi unit;The distance between the current navigation point and each adjacent robot navigation point is calculated;If there is distance less than the physical radius of robot, then the current navigation point is offset angle in counterclockwise direction along the boundary of Voronoi unit, to generate new navigation point;New navigation point is projected on the boundary of buffer Voronoi unit, to obtain the current target navigation point;According to the self-observation information and external observation information of robot, determine the observation space, generate the action of robot based on the observation space and the trained strategy network, and determine the linear velocity and angular velocity of robot in the next iteration round based on action.The method improves the success rate of robot obstacle avoidance.

The invention provides a method and a system for fraud detection by using a contrast graph neural network based on intra-class substructure differentiation, and solves the technical problem of low detection precision caused by heterogeneity and class imbalance of existing similar subclusters. The method comprises the following steps: acquiring original data, and carrying out manual annotation to obtain annotated data; modeling the marked data into a graph structure, inputting the graph structure into a comparison graph neural network model based on substructure differentiation for training, and obtaining a trained comparison graph neural network model after reaching a preset training round; and reasoning the original data modeled as a graph structure by using the trained comparison graph neural network model to obtain a classification result. The method can be widely applied to the technical field of fraud detection.

The invention discloses a document image classification method and system based on a multi-modal contrast learning framework and a readable storage medium, and relates to the field of computer vision and natural language processing. Comprising the following steps: constructing a training data set, extracting positive and negative sample pairs, sending the positive and negative sample pairs into a multi-modal large model, setting cue words for model training, and optimizing the model by supervising and comparing loss; respectively selecting examples from a plurality of categories as templates, respectively sending the templates into a model, and storing a plurality of output feature embedding vectors and corresponding categories into a template library; and sending a document image of which the category is to be confirmed into the model, outputting feature embedding vectors, reading the template, and taking the category corresponding to the feature embedding vector with the maximum similarity as a classification result. According to the method, algorithms such as multi-modal learning and comparative learning are combined, multi-source information such as images and texts is fused, discriminative features in the multi-source information are extracted, accurate classification of document images is achieved, and the method has the advantages of being universal, efficient and high in precision.

This invention relates to the field of power equipment safety monitoring technology, and addresses the problems of existing technologies such as reliance on physical tags, insufficient accuracy in key point positioning, unstable depth estimation, and difficulty in end-side deployment. The invention constructs a database of key points and depth estimation in substations, designs a lightweight key point detector, and achieves accurate positioning of key points in the risky areas of workers and energized parts of equipment. A stereo perception module outputs dense depth information, mapping two-dimensional key points to three-dimensional space, calculating the minimum safe distance between human key points and energized parts, and combining threshold values ​​to achieve graded early warning and visualization. This invention eliminates the need for additional physical tags, achieving significant improvements in detection accuracy, real-time performance, lightweight deployment, and engineering reliability through key point detection and stereo vision collaborative perception. It effectively solves the problems of insufficient accuracy and difficult deployment in safety distance monitoring in complex substation scenarios.

The application discloses a kind of spatio-temporal perception remote sensing image change detection methods based on SAM2.The application introduces spatio-temporal perception adapter as the weight sharing twin encoder, difference modeling and attention enhancement are carried out to double time phase features, the consistency of unchanged area features is improved, and the difference of change area features is expanded, so as to effectively alleviate the interference of radiation difference.Simultaneously, a dynamic adaptive scanning change decoder is designed, and the scanning position and scanning path are adaptively adjusted according to the local change characteristics, so as to enhance the modeling capability of irregular change area under complex background.Finally, the proposed method cooperates to improve the double time phase feature representation and change information extraction capability from the encoding end and the decoding end, and effectively improves the accuracy of change detection under the condition of serious radiation difference interference.

The invention relates to a non-ideal array DOA estimation method based on multi-scale weight distribution and transfer learning and electronic equipment, and the method comprises the steps: constructing a fine tuning data set under a non-ideal array condition, the fine tuning data set comprising array signal data simulating various array physical defects; constructing an improved denoising and classification network, and training the improved denoising and classification network by adopting a transfer learning strategy; inputting a test signal in a non-ideal array environment into the trained improved denoising and classification network, and performing DOA estimation; the electronic equipment is realized based on the method. According to the method, the calculation overhead is reduced, the model is more robust to the interference of array errors and noise, the feature expression ability of signals is improved, the robustness and precision of the model in a complex environment are improved, transfer learning enables the model to obtain better performance in a low signal-to-noise ratio environment through a parameter sharing mode, and the method is suitable for large-scale popularization and application. And new non-ideal array data can be quickly adapted, and the retraining time and computing resources are reduced.

The application provides a kind of overhead power line surveying and mapping image key area identification processing method and system, wherein the method comprises: obtaining the original image data collected by camera and the geographic position data of camera shooting position;Match the geographic position data with the preset geographic position database, obtain the natural material prior probability data set corresponding to the camera shooting position;Extract polarization feature and image content feature from the original image data, use double-branch attention network, establish material optical property table based on polarization feature in the first branch processing path, fuse natural material prior probability data set and image content feature in the second branch processing path;Based on material optical property table and fusion result, identify key area from original image data;Conduct confidence assessment on key area, generate key area confidence map.The application improves the identification accuracy of natural material key area in outdoor complex scene.

This invention discloses an infrared weak target recognition method based on skip connections enhanced by Unet, belonging to the fields of computer vision and artificial intelligence. The method includes: extracting features from infrared images using the Unet encoder; adjusting the channel and spatial dimensions of the feature map using an adaptive convolutional embedding mechanism; inputting the feature map into a two-dimensional Mamba backbone network; fusing the original features with the processed features through residual connections; and performing batch normalization and activation function optimization on the fused feature map; reconnecting the feature map to the Unet decoder using the adaptive convolutional embedding mechanism; employing DiceLoss and Adam optimizers for collaborative training; and extracting the target region based on a locking mechanism with a pixel value of 255 in the mask. This invention enhances the feature representation capability of Unet skip connections by introducing an MCA module, enabling better mining of multi-level feature information. Through multi-level feature fusion, it improves the accuracy and robustness of infrared weak target detection.

This paper discloses a method, system, storage medium, and electronic device for predicting flood extent based on SAR imagery assisted by geographic environment data. The method includes the following steps: data preprocessing; extracting features using a dual-branch feature encoder to obtain backscattering features of SAR data at different scales and physical driving features of auxiliary data; using an adaptive cross-gated fusion module to fuse the backscattering features and physical driving features of the encoder part layer by layer to obtain multimodal fusion features; adaptively fusing the fusion features of the encoder part with the upsampling features of the decoder part; upsampling and restoring the image size to obtain the prediction result. This invention enables rapid and accurate prediction of flood extent. This method can provide a reference for real-time flood disaster early warning, emergency command and dispatch, and flood control and disaster reduction decision-making. It also provides a new idea and research direction for the future development of more timely and accurate flood extent prediction technologies.

This invention discloses a 3D target detection method based on sparse dynamic attention and star-shaped interaction. The method obtains basic voxel features from the original LiDAR point cloud through voxelization and sparse convolution, then introduces a sparse dynamic parallel attention module. This module achieves efficient enhancement of global context and channel dimensions through dynamic attention branches and parallel channel interaction branches. A sparse star-shaped interaction module is then used to construct a star-shaped neighborhood interaction structure with a central voxel, completing local geometric modeling and nonlinear feature interaction only on non-empty voxels. Finally, keypoint sampling, RoI pooling, and a detection head output the 3D detection box, category, and confidence score. This invention, through the synergistic complementarity of SDPA and SSB, significantly improves the detection accuracy of long-distance, small-scale, and occluded targets while maintaining linear growth in computational complexity and meeting real-time requirements. It achieves balanced performance optimization across multiple categories, including vehicles, pedestrians, and cyclists, and is suitable for 3D perception scenarios with high precision and real-time requirements, such as autonomous driving.

This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for detecting and locating multiple sound events based on a neural network model. The method includes: innovatively designing time-frequency multi-scale residual convolutional blocks, which, together with a Conformer module and a cross-stitch unit module, form a network model to extract features at multiple scales, enhance long sequence modeling, and promote task-based collaborative optimization, thereby improving performance and accuracy; in terms of data processing, pre-emphasis and frame-by-frame windowing improve feature quality, audio channel swapping and spectrum enhancement increase data diversity and reduce overfitting, and SALSA-Lite features are used to enhance feature representation; in terms of training strategy, a multivariate loss function is used to accelerate convergence while considering task requirements, and hyperparameters are flexibly adjusted using a validation set. This makes the method highly efficient in training, has excellent practical performance, strong generalization ability on unknown data, and can accurately cope with complex and ever-changing real-world scenarios, effectively overcoming the shortcomings of traditional methods.

This application provides a method, apparatus, device, and storage medium for matching and repairing lip movements and speech content in a video, belonging to the field of image processing. It acquires original video data and supplementary text information, performs speech synthesis, generates speech feature parameters, and then generates a lip movement animation sequence synchronized with the speech content. A semantic segmentation model is used for pixel-level recognition of the mouth region, and a physical constraint model is combined to control the tooth region. If the tooth region exceeds the lip contour, dynamic regression adjustment is performed to ensure the naturalness and plausibility of the generated lip movement. Furthermore, the accuracy of tooth region generation is improved through dense cross-layer connections, a feature pyramid structure, specialized skip connections for the tooth region, and a heatmap gating mechanism for key tooth points. A spatial-channel dual attention algorithm is introduced into the decoder to enhance the feature representation of key mouth regions. A tooth region focus loss function is used to optimize the generation results and improve model robustness. A differential physics engine simulates a point mass spring system and adversarial physical constraints to achieve physical plausibility control of key lip points. Finally, the generated lip movement animation is fused with the original video to output the repaired video, significantly improving the consistency and realism of the speech and lip movement in the video.

This invention belongs to the field of power system operation and control technology, and relates to a voltage optimization control method for three-phase distribution networks based on vertical and horizontal flexible resource coordination and large language model-enhanced federated reinforcement learning. The method establishes a three-phase distribution network power flow model and defines a voltage optimization problem, setting up a state space, action space, and initial reward function for multi-agent reinforcement learning. Each agent in each region uses a large language model and data distillation algorithm to obtain a low-dimensional state vector, constructs a phase correlation matrix using phase features, and generates weighted features. Each agent in each region independently performs local reinforcement learning training, solves constraints based on the three-phase distribution network power flow model, generates a semantic summary which is uploaded to the computing center, generates global policy guidance information, and distributes it to each agent. The trained agents are deployed to the distribution network control system, generating action commands based on real-time collected system states to perform coordinated voltage regulation of vertical and horizontal flexible resources.

The application is suitable for the technical field of computer application, and provides a two-dimensional code image feature extraction method and device based on a CNN-MALT hybrid architecture, which comprises the following steps: inputting a target two-dimensional code image into a CNN preprocessing module of a preset lightweight image feature extraction network to generate first structure perception features; inputting the first structure perception features into a MALT encoder of the preset lightweight image feature extraction network, performing feature extraction on the first structure perception features based on a proxy attention mechanism to generate deep features; and inputting the deep features into a feature fusion module of the preset lightweight image feature extraction network to generate target features. Thus, the structure features of the two-dimensional code image are perceived by the CNN preprocessing module, the deep features are generated by the MALT encoder based on the proxy attention mechanism, and the deep features are fused by the feature fusion module, so as to ensure that the calculation complexity is reduced while the two-dimensional code image feature extraction accuracy is improved.

The application discloses a hydrocarbon detection method based on an adjacency matrix constraint and bias correction of a mixed residual network, and comprises the following steps: extracting prestack, poststack seismic and spectral attributes and performing quality control; constructing a loopless directed graph geological network based on reservoir structure; taking an adjacency matrix generated by the network as a constraint to construct a mixed residual network model, the model fusing seismic spectrum, attribute vectors, geological node codes and depth data through a multi-source input layer, using an adjacency matrix geological constraint module to perform spatial topological optimization on a preliminary prediction result, and integrating a clustering bias correction module to eliminate background value differences; and using a training strategy based on geological unit iteration to train the model for hydrocarbon detection. The application effectively fuses multi-modal seismic information, and embeds geological rules as strong prior knowledge into a deep learning model, thereby significantly improving the accuracy and spatial rationality of hydrocarbon detection, and having stronger generalization ability and engineering application value in a few well areas.

The application discloses an active landslide identification method, and relates to the field of disaster identification, which comprises the following steps: acquiring image data of a region to be identified, and constructing an active landslide data set; constructing a context-aware adaptive fusion model; the context-aware adaptive fusion model comprises an encoder and a decoder; the encoder comprises a plurality of feature processing layers and a convolution layer connected in sequence; the feature processing layer comprises a convolution layer and a wavelet transform down-sampling WDB module connected in sequence; the decoder comprises a plurality of multi-branch scale adaptive aggregation MSA modules, and a plurality of feature extraction refinement layers and a convolution layer connected in sequence; the feature extraction refinement layer comprises a convolution layer and a convolution up-sampling CUB module connected in sequence; based on the active landslide data set, the context-aware adaptive fusion model is trained to obtain an active landslide identification model; the active landslide data set to be detected is input into the active landslide identification model to obtain an active landslide identification result.

The invention belongs to the technical field of image classification and recognition processing, and relates to a combat target recognition method based on multi-scale space-frequency domain feature enhancement, which comprises the following steps of: establishing a multi-scale space-frequency domain feature fusion network MSSFF-Net consisting of a dual-path spatial domain feature extractor DP-SFE, a multi-scale frequency domain feature enhancement module MS-FDFEM and a space-frequency domain feature fusion module SFFFM; the method is used for space-frequency domain feature extraction, fine-grained expression enhancement and effective fusion. According to the method, space detail information in video frames and inter-frame frequency domain discriminative features are effectively combined through lightweight design, the robustness and discriminative power of feature representation are enhanced under the condition of limited labeled samples, and scene constraints of complex backgrounds, posture dynamic changes and noise interference in video information are adapted; according to the method, the defect of insufficient feature abstraction capability on a shallow-layer backbone network can be remarkably overcome, and meanwhile, certain performance improvement can be brought to a deep-layer backbone network.