2367 results about "Feature aggregation" patented technology

The invention provides a medical decision-oriented multi-modal data dynamic fusion and labeling method and system, and relates to the technical field of medical decision-making, and the method comprises the steps: collecting medical images, electronic medical records and physiological monitoring data of a patient, extracting multi-modal features such as a focus form, a disease type, a symptom, a treatment record and a physiological index, and carrying out the fusion and labeling of the multi-modal features; and constructing a dynamic fusion labeling model. According to the model, multi-modal features are coded, similarity is calculated, a feature association graph is constructed, feature aggregation is carried out, labeling is carried out in combination with a medical knowledge base, and a dynamic labeling vector is generated. And inputting the dynamic labeling vector into a medical decision support model, and based on similar case analysis and deep reinforcement learning, generating a diagnosis and treatment suggestion containing a treatment scheme and early warning information. By dynamically fusing multi-modal data and an online updating mechanism, the accuracy and efficiency of medical decision making can be improved, and personalized diagnosis and treatment are realized.

The invention discloses an image-text matching method based on progressive semantic aggregation and structured cognitive enhancement, and relates to the technical field of image-text matching, comprising the following steps: S1, multi-modal feature extraction; s2, introducing a hierarchical self-enhanced semantic perception pooling mechanism; s3, introducing a structured knowledge auxiliary learning module; s4, introducing a cross-modal semantic alignment mechanism; s5, introducing a multi-scale semantic association matrix; according to the method, through multi-level semantic enhancement and fine-grained level self-adaptive feature aggregation, a significant word field is enhanced, and consistency and unification between local representation and global representation are achieved; extracting structured semantics between entities from an image title corpus, mining high-order relationships in long and short sentences, and expanding concepts missed in text global embedding by utilizing the priori knowledge; a cross-modal relation interaction mechanism is introduced, semantic embedding is enhanced, and more consistent semantic understanding is promoted; a multi-scale semantic association matrix is introduced, and a more comprehensive and accurate similarity measurement matrix is generated.

The invention discloses a transmission line three-dimensional reconstruction method and system based on a monocular channel monitoring camera, and the method comprises the steps: obtaining compensation image data through adaptive illumination compensation processing, obtaining direction feature data based on frequency domain feature analysis, constructing a texture enhancement model through a graph theory, generating enhanced image data, and carrying out the reconstruction of a transmission line. Outputting feature aggregation data through time sequence feature aggregation; generating three-dimensional contour data in combination with multi-wire topology analysis and geometric constraint optimization; depth probability modeling and multi-scale consistency processing are adopted to obtain optimized depth data; and outputting system state data, three-dimensional point cloud data and deformation parameter data through physical model construction and space-time constraint fusion. According to the method, high-precision three-dimensional reconstruction of the power transmission conductor is realized, and the method has relatively high environmental adaptability and calculation efficiency.

The invention, which relates to the technical field of network awareness anomaly detection, discloses a network awareness anomaly detection system based on big data, comprising a data acquisition module, a feature fusion module, a map construction module, a model calculation module, a root cause reasoning module, a threshold decision module and a response control module. The data acquisition module receives network flow data, equipment state data and system log data and outputs a standardized feature set; the feature fusion module is connected with the data acquisition module, dynamically calculates a weight coefficient of each data source based on information entropy, performs feature aggregation of privacy protection through a federated learning framework, and outputs a fusion feature vector; the atlas construction module is connected with the feature fusion module, maintains a network equipment node set and a communication edge set in real time, and updates a space-time association atlas according to a topology change event; and the model calculation module is connected with the atlas construction module, extracts topological features through a space-time diagram convolutional network, and updates a detection model based on an incremental learning mechanism.

The invention discloses an urban inland inundation risk multi-level prediction method and device based on time-space diagram learning, a storage medium and a computer program product, and relates to the technical field of natural disaster risk prediction, and the method comprises the steps: collecting multi-modal urban hydrological data; performing hierarchical time modeling on the multi-modal urban hydrological data, and extracting a time embedding vector; constructing a heterogeneous graph based on the time embedding vector, and performing spatial feature aggregation calculation on the heterogeneous graph to obtain spatial embedding representation; and performing multi-level prediction according to the spatial embedding representation to obtain a multi-granularity waterlogging risk index. Through multi-modal data acquisition and preprocessing, layered time modeling, heterogeneous graph construction, spatial feature aggregation calculation and multi-level prediction, multi-modal urban hydrological data are effectively fused, and comprehensive and accurate urban inland inundation risk prediction is realized.

The invention provides a dynamic routing selection method and system, electronic equipment and a storage medium, and aims to solve the problem that a routing strategy is difficult to adapt to a dynamically changing network, the method comprises the following steps: a terminal layer collects the state of a terminal and network data, and performs lightweight feature extraction; the edge node layer receives the data of the terminal layer, carries out space-time-semantic feature aggregation, and generates a region-level resource scheduling and routing decision strategy based on fragmented reinforcement learning; the central cloud service layer gathers whole network data, generates a global optimization strategy and issues the global optimization strategy; the edge node layer fuses global optimization and a region-level strategy, and executes dynamic routing selection; and security and privacy protection are provided through the trusted chain layer. According to the invention, adaptive path selection can be realized, the network resource utilization rate is improved, and the network stability is improved.

The invention discloses a traffic flow prediction method based on adaptive dynamic multi-scale space-time hypergraph convolution. The method comprises the following steps: S1, collecting traffic data; s2, forming space-time enhancement data from the traffic data; s3, a multi-scale adaptive causal convolution module extracts multi-scale traffic flow change features; s4, constructing a dynamic hypergraph structure, defining a hypergraph incidence matrix, and generating a self-adaptive hypergraph by adopting an attention enhancement matrix decomposition method; s5, fusing the time features in the S3 and the hypergraph structure in the S4, designing a multi-scale fusion homogeneous convolution module, and realizing fusion of multiple different-scale spatio-temporal features; s6, extracting topological node features through a graph convolutional network and a hypergraph convolutional network by using the multi-scale spatial-temporal features, and performing adaptive weighted fusion through a gating feature fusion unit; s7, performing multi-scale feature integration by a residual feature aggregation module; s8, generating traffic flow prediction results of a plurality of time steps in the future; the method has the advantages of extracting the traffic space-time dependency relationship, improving the prediction precision and generalization ability, and being more suitable for dynamic traffic.

The invention provides a multi-level heterogeneous computing power network task scheduling and resource allocation method and system. In a computing power network, each hierarchical network is connected with an intelligent agent for task scheduling and resource allocation. In the training process, the intelligent agent obtains local observation information of the computing power node in the hierarchical network. And the intelligent agent aggregates and converts local observation information according to the relation feature map to generate a feature matrix and an adjacent matrix. Local feature embedding is calculated by using a graph convolutional network, and feature aggregation is carried out in combination with information of a cooperative agent to obtain a local embedding state; and the intelligent agent executes a joint action according to the local embedding state, wherein the joint action comprises task scheduling and resource allocation. And according to an execution result, the agent obtains a reward value and state information and stores the reward value and the state information into an experience playback buffer area. And sampling data from a buffer area, and updating a value function and strategy parameters by taking optimization of calculation task completion time and resource use efficiency as a target. And after a preset training round, obtaining a trained model. The resource overhead of information interaction can be reduced, and the scheduling efficiency is improved.

The invention provides a remote sensing coastline automatic extraction method and system based on residual space pyramid segmentation and two-dimensional attention, and relates to the technical field of space analysis. The method comprises the steps of high-resolution remote sensing image acquisition and preprocessing, sea-land segmentation network reasoning, probability graph thresholding and edge extraction and vectorization processing. According to the method, the segmentation precision is improved through multi-scale feature aggregation and attention enhancement, coastline vector data with geographic coordinates are generated in combination with edge detection and topological repair, and the method is suitable for spatial analysis and coastline monitoring.

The invention discloses a credit risk assessment method and system based on a knowledge graph and a storage medium, and the method comprises the steps: carrying out the standardization processing of multi-source credit data through an ontology mapping rule, and obtaining an RDF triple data set; constructing a dynamic knowledge graph containing a guarantee chain, a fund flow direction and a risk factor by adopting a self-organizing algorithm; carrying out modeling through a graph convolution risk propagation operator to obtain a risk state vector; performing time sequence embedding extraction to obtain a five-dimensional comprehensive risk feature vector; and through adaptive attention mechanism aggregation processing, a credit risk assessment result and a confidence interval are obtained. The technical problems of multi-source heterogeneous data fusion, dynamic relation modeling, risk state quantification and multi-dimensional feature aggregation in credit risk assessment based on the knowledge graph are solved.

The invention provides a lightweight underwater image enhancement method based on adaptive feature fusion, and the method comprises the steps: decomposing an underwater image enhancement task into a strong prior enhancement task and a fine-grained enhancement task through employing a lightweight underwater image enhancement network based on adaptive feature fusion, so as to make full use of the multi-scale information in an image; the strong prior enhancement task is used for extracting global and low-layer features of the image; the fine-grained enhancement task is used for improving local details and color recovery of the image so as to improve the visual quality; the lightweight underwater image enhancement network comprises a multi-scale feature aggregation module and a global channel attention module, the multi-scale feature aggregation module is composed of an adaptive feature fusion module and a segmented convolution feed-forward network, and the lightweight underwater image enhancement network is used for combining global and local information of features and adaptively adjusting importance of each channel through information association between channels.

The invention provides a face forgery detection method and system based on reconstruction learning and a hybrid expert mode. The method comprises the following steps: extracting a face frame image from a face video; performing data processing on the face frame image to obtain a two-dimensional face image; swin Transform is used as an encoder, a hybrid expert model is introduced at the same time, and spatial feature and long-distance dependence feature extraction is carried out on the two-dimensional face image; enhancing different types of features in the two-dimensional face image by using a hybrid expert model to obtain a high-level semantic feature map; generating a face reconstruction image by using an attention-guided decoding module; and aggregating the face reconstruction image and the corresponding two-dimensional face image by using a feature aggregation module based on reconstruction guidance, and exploring a possible counterfeit region to obtain a final feature map for face counterfeit detection. According to the method, the detection capability of unknown counterfeit types is better, and the accuracy, robustness and generalization of face counterfeit detection can be improved.

The invention discloses a water surface target detection method based on a visual language large model, which mainly solves the problem of inaccurate water surface target detection in the prior art, and the implementation scheme comprises the following steps: constructing a water surface target detection data set; extracting multi-scale features of the image by using the backbone network; constructing a dynamic feature aggregation network to fuse the multi-scale features of the image; extracting language features of the target category text; constructing a cross-modal cross attention fusion network, and fusing the visual dynamic aggregation features and the language features; utilizing a dynamic target detection head to obtain a prediction target category and a target frame; forming a water surface target detection model based on visual language multi-modal fusion by using the backbone network, the dynamic feature aggregation network, the text encoder, the cross-modal cross attention fusion network and the dynamic target detection head, and training the water surface target detection model; and obtaining a water surface target detection result by using the trained detection model. The method can effectively detect the water surface target, is high in accuracy and robustness, and can be used for intelligent ship driving.

The invention provides a decoding and coding method, device and equipment, and the decoding method comprises the steps: decoding a first code stream of a current image block, and obtaining a coefficient hyper-parameter feature of the current image block; for each stage sub-block of the current image block, determining a probability distribution parameter of the stage sub-block based on the coefficient hyper-parameter feature, and decoding a second code stream of the current image block based on the probability distribution parameter to obtain a residual feature of the stage sub-block; based on a mean value prediction network, determining mean value features of the sub-blocks in the stage; determining a reconstruction feature of the stage sub-block based on the mean feature of the stage sub-block and the residual feature of the stage sub-block; performing feature aggregation on the reconstruction features of the sub-blocks in each stage to obtain the reconstruction features of the current image block; a reconstructed image block of the current image block is determined based on the reconstruction feature of the current image block. Through the scheme of the invention, the coding performance and the decoding performance are improved.

The invention provides a high-resolution remote sensing image semantic segmentation method based on multi-scale depth supervision, and the method comprises the following steps: 1, obtaining an ISPRS Potsdam data set and a Vaihingen data set; 2, building a multi-scale deep supervision network MDSNet, wherein the multi-scale deep supervision network MDSNet comprises a spatial branch network, a semantic branch network and a multi-scale feature aggregation part; each of the spatial branch, the semantic branch network and the multi-scale feature aggregation part is of a four-layer structure; step 3, inputting the training set into a multi-scale deep supervised network MDSNet for training, calculating a loss function and performing back propagation, updating network parameters, and obtaining a trained parameter model; and step 4, inputting the test set into the trained parameter model, and outputting a segmentation image of the target image. According to the method provided by the invention, the feature extraction capability is improved, and meanwhile, the semantic segmentation accuracy, the training efficiency and the stability are improved.

The invention provides a time sequence prediction method and system for a KAN enabling graph attention parallel loop network, and relates to the technical field of time sequence prediction, and the method comprises the steps: obtaining time sequence input data; inputting the time sequence input data into a feature extraction layer based on fast Fourier transform and discrete wavelet transform, and preprocessing the time sequence input data by the feature extraction layer to obtain spliced feature extraction data; inputting the spliced feature extraction data into a feature aggregation layer of graph attention based on KAN, and preprocessing the spliced feature extraction data by the feature aggregation layer to obtain feature aggregation data; and inputting the feature aggregation data into a time sequence prediction model of the parallel gating circulation unit network based on KAN and attention, and outputting time sequence prediction data by the time sequence prediction model. According to the method, the calculation complexity of time sequence prediction is reduced, and the time sequence prediction efficiency and accuracy are improved.

The invention discloses a multi-unmanned aerial vehicle collaborative adversarial learning method based on reinforcement learning, and the method comprises the steps: designing a communication mechanism based on multiple agents, combining a feature aggregation extraction module of a long-short-term memory network, and accurately and efficiently extracting battlefield situation features; a hierarchical reinforcement learning method is adopted, a bottom maneuvering control module and an upper combat decision module are introduced, bottom control and upper decision are achieved respectively, training is performed based on an efficient environment parallel module, and combat strategy training efficiency in a complex combat scene is improved; in addition, the generalization ability of the combat strategy is improved through a multi-agent game training module based on population evolution diversification. The method can effectively extract battlefield situation features, carries out efficient cooperative combat training, adapts to a complex and changeable battlefield environment, and improves the generalization ability of unmanned aerial vehicle autonomous cooperative combat.

The invention relates to the field of power data management, and discloses a power big data adaptive management method and system fused with spatial-temporal feature mapping, and the method comprises the steps: obtaining a real-time operation data flow from a multi-source power terminal, and constructing an original power data set with a time sequence label and a device identifier; the method comprises the following steps: dividing an original power data set into parallel processing units based on a storage-while-computing architecture, performing dynamic index updating by adopting an event-driven index mapping rule, and constructing a multi-dimensional data index cache system with real-time responsiveness; identifying a key abnormal trajectory through a time-varying feature nesting mechanism, and performing hierarchical measurement and entropy disturbance analysis on a data fluctuation degree in the key abnormal trajectory by using a streaming feature aggregation network; identifying potential security risk nodes in combination with the structure matching degree between the historical abnormal event evolution graph and the key abnormal trajectory; and generating a multi-level response instruction chain based on the risk assessment result. The method has the advantage of improving the operation safety of the power grid.

The invention discloses a power grid data acquisition and analysis system based on big data, particularly relates to the technical field of power network monitoring, and is used for solving the problem of key feature loss caused by cross-hierarchy data semantic association missing of an existing hierarchical aggregation mechanism. Performing standardized filling and time label layered alignment processing on the original data through the equipment acquisition module; the feature extraction module screens an abnormal waveform fragment and a steady-state parameter offset based on the equipment type and the real-time fluctuation amplitude; the semantic association module generates a space-time association weight in combination with the regional physical topological relation and the reactive circulation path sensitivity; the feature clustering module reconstructs a region-level feature aggregation packet through kernel density estimation and spatial weighted fusion; the cross-layer analysis module dynamically corrects the aggregation weight based on the current flow direction and the transient energy distribution; and the decision generation module generates an equipment positioning instruction and load scheduling strategy combination through alarm template matching, and finally realizes accurate positioning of fault equipment and generation of a scheduling strategy.

The invention relates to the technical field of audio testing, and discloses an intelligent audio testing system frequency response analysis method, device and equipment, and the method comprises the steps: carrying out the audio parameter initialization of an intelligent audio testing system, and obtaining a testing parameter matrix and an excitation signal; performing dual-channel acquisition and time-frequency feature extraction to obtain a time-frequency feature data set; executing double-branch processing of time domain noise reduction and frequency domain enhancement to obtain a purified feature data set and a feature weighting matrix; performing segmented recursive calculation and feature aggregation to obtain a frequency response feature vector; decomposing the frequency response feature vector according to a test frequency band, and performing environment compensation and nonlinear calibration to obtain a standard frequency response data set; and inputting the standard frequency response data set into the deep neural network for feature recognition and mode classification to obtain a frequency response curve and a phase characteristic curve, thereby improving the test precision, the anti-interference capability, the automation degree and the adaptability of the intelligent audio test system.

The invention provides a whole vehicle point cloud windage coefficient prediction method and system based on multi-scale learning and convolution, and relates to the technical field of windage coefficient prediction, and the method specifically comprises the steps: obtaining point cloud data of a whole vehicle model, and carrying out the preprocessing of the point cloud data; sampling the preprocessed point cloud data through a farthest point sampling method so as to reserve geometric key points; constructing a convolutional neural network model, inputting the point cloud data into the network, sequentially carrying out two times of multi-scale convolution operations, extracting local details and global structure features through convolution kernels of different scales, mapping the features from low dimensions to 512 dimensions and expanding the features to 1024 dimensions, and completing feature aggregation; the feature expression capability is further enhanced through two-layer convolution; utilizing a maximum pooling layer to aggregate global features, and performing dimension reduction on the features through a multi-layer perceptron; introducing a physical guidance attention mechanism to physically constrain the spatial weight of the features; and establishing a mapping relation between the point cloud features and the wind resistance coefficient through a full connection layer, and outputting a wind resistance coefficient prediction result. According to the invention, the accuracy and efficiency of wind resistance coefficient prediction can be improved.

The invention discloses a flowmeter-oriented embedded multi-parameter fusion calibration control method and system, relates to the technical field of industrial measurement and embedded control, and is used for solving the problem of unstable calibration caused by misalignment of multi-link flow signals, compensation lag and fusion oscillation. The method relates to synchronous acquisition, pretreatment and calibration of differential pressure type, ultrasonic wave and thermal type mass measurement links and environment parameters such as temperature, pressure and density. All-channel unified time scales are established through ternary binding, filtering and denoising, zero snapshot and alignment rules, a stability scoring mechanism is constructed in combination with offset behavior recognition, event classification and feature aggregation, and dynamic switching of a compensation path and a fusion strategy is driven. In a double-path fusion structure, table look-up compensation, model estimation, multi-algorithm parallelism and fusion track recording are adopted, version traceability and calibration result closed-loop control are achieved, and the stability and rechecking performance of mass flow output under complex disturbance are effectively improved.

The embodiment of the invention discloses a method and system for identifying transaction risks, and the method comprises the following steps: obtaining multi-source heterogeneous data, and initializing a dynamic transaction graph according to a multi-source heterogeneous data flow; the multi-source heterogeneous data comprises invoice data, fund flow data, an enterprise stock equity relationship and geographic data; receiving a multi-source event stream through a message queue, and updating the dynamic transaction graph in real time based on the multi-source event stream; for enterprise nodes affected by updating in the dynamic transaction graph, feature aggregation is executed, topological risk features are extracted, and risk scores of the enterprise nodes are output; verifying the consistency of the invoice hash and the fund flow hash through a relay chain protocol, and obtaining a cross-chain verification result; and according to the risk score and the cross-chain verification result, identifying a transaction risk. According to the embodiment of the invention, systematic risk identification is realized through data integration, graph construction, risk mining and cross-chain verification, and the accuracy of invoice risk identification can be improved.

The invention provides an RGB-D salient target detection method and system based on cross-modal correlation learning, and the method comprises the steps: constructing a feature extraction network of an RGB mode and a depth mode based on a double-flow encoder architecture; performing feature fusion by using an adaptive cross-modal fusion module to obtain cross-modal fusion features; performing attention enhancement on the refined large-scale features and the small-scale up-sampling features by utilizing element-by-element multiplication to obtain output fusion features of a multi-scale feature aggregation module; performing object boundary prediction on output fusion features of the multi-scale feature aggregation module by using an up-sampling layer and a convolution layer to obtain features of an edge prediction map; performing point-by-point addition operation on the saliency features to obtain a final saliency prediction map; inputting the RGB-D data into a network model for training so as to generate an RGB-D salient target detection result; according to the method, the relevance and complementarity between the modals are mined, the characteristic information of different modals is fully integrated, and the accuracy and robustness of salient target detection are improved.

The invention relates to the technical field of network communication, and discloses an intelligent control method and system of a switch. The method comprises the following steps: receiving a heterogeneous data stream in a target network domain; performing data processing by using a preset flow feature aggregation model, and generating a flow dynamic pressure field distribution diagram and a protocol priority mapping matrix; constructing an adaptive scheduling threshold model, and generating a multi-target flow control instruction set; a cascade effect is simulated based on a network congestion chained conduction model, and collaborative routing parameters are optimized; and iteratively optimizing through a mixed integer programming framework, and outputting a network scheduling action sequence to a switch control engine. The system correspondingly comprises a data receiving module, a data processing module, an instruction generation module, a simulation optimization module and an instruction output module. According to the method, heterogeneous data can be effectively processed, congestion can be accurately predicted, flow control can be optimized, the overall performance of the network can be improved, and communication requirements in a complex network environment can be met.

The invention discloses an industrial defect detection method based on self-supervised fine tuning, and solves the problems of scarcity of industrial scene defect samples and weak model generalization ability. The method comprises the steps that a data set is divided and preprocessed, and the data robustness is improved through size scaling, random luminosity transformation, geometric enhancement and the like; extracting a foreground mask by using a saliency model, synthesizing a Perlin Noise and DTD texture fused pseudo-abnormal image, carrying out self-supervised fine tuning on the ImageNet pre-trained WideResNet-50, and enhancing the industrial data feature extraction capability; a model containing a visual trunk, feature aggregation mapping, noise feature adaptation and a discriminator is established, local neighborhood features are fused through Unfold operation, Gaussian noise is superposed to generate pseudo-abnormal features, and an abnormal score is output by the discriminator after multi-scale fusion. And the training adopts binary cross entropy and focus loss optimization parameters. The innovation points of the method are that self-supervised fine tuning adapts to industrial data distribution, feature aggregation improves fine-grained detection, and multi-scale fusion considers different defects.

The invention provides an ocean wind field prediction method based on a neural network, and belongs to the technical field of ocean wind field prediction.The method comprises the steps that sparse ocean observation data are collected, a spatial covariance matrix is established, the spatial covariance matrix is converted into a graph structure, and then multi-hop neighborhood feature aggregation is conducted through a graph convolutional network; a tensor decomposition algorithm is combined for modeling high-order feature interaction to generate a gridding wind field, a bidirectional long-short-term memory network encoder is used for extracting space-time invariant features, a multi-layer perceptron predictor is used for directly mapping a future multi-step wind field, and a course learning strategy and a Shenchang differential equation boundary layer are matched for correction. The technical problem that sparse ocean observation data are difficult to accurately reconstruct into a high-resolution gridding wind field is solved.

The invention discloses a PHP taint type vulnerability detection method based on a heterogeneous graph neural network, and belongs to the field of software security. The method comprises the following steps: performing annotation removal, variable naming standardization and character string standardization processing on a PHP source code through a code preprocessing module to generate a standardized code; based on a vulnerability sub-attribute graph extraction module, reversely tracking vulnerability sinks to a taint source, extracting a simplified vulnerability sub-attribute graph, and removing redundant nodes and edges; fusing BERT semantic features and node type features through a graph node embedding module to generate an initial embedding vector, and constructing a heterogeneous graph comprising an abstract syntax tree edge, a program flow graph edge and a control dependence graph edge; a heterogeneous graph neural network vulnerability detection module is adopted to perform independent feature aggregation on multiple types of edges, dynamic weighted fusion is performed in combination with an attention mechanism, and key nodes are screened through Top-k graph pooling; and finally, inputting the graph-level features into a classifier to realize vulnerability detection.

The invention provides a surface roughness measuring method and device based on a multispectral laser speckle image, and belongs to the field of surface quality visual inspection. Comprising the following steps: respectively acquiring speckle images formed by irradiating the surface of a workpiece with lasers of N wavelengths at different angles, and constructing a multispectral laser speckle data set; the speckle images of N wavelengths at a single angle are spliced according to channels and subjected to feature fusion, and then a multispectral laser speckle fusion image is generated; the fused images of the multiple angles form an image sequence, and single images in the sequence are subjected to multi-scale feature extraction and fusion to generate single-angle features; the multi-angle features generate global illumination information through a multi-angle feature aggregation module, and finally a roughness value is calculated through a multi-layer perceptron. Multi-spectral laser speckle coding surface features are adopted, multi-angle information is fused through the multi-angle feature aggregation neural network, and the relation between the light source wavelength, the incident angle and the surface three-dimensional shape and roughness is comprehensively considered, so that high-precision measurement is achieved.

The invention discloses a multi-scale attention aggregation-based target point cloud segmentation method for a battery replacement robot, and the method comprises the steps: 1, collecting an RGB image, a depth image and three-dimensional point cloud data of a target fastener through a binocular structured light depth camera, and carrying out the fusion to generate a FastSeg3D data set; 2, a two-stage preprocessing method is provided, noise points are removed through radius filtering, and background point clusters far away from a target are removed through DBSCAN density clustering; 3, the network encoder uses a local feature aggregation module to extract geometric features, and the calculation complexity is reduced in combination with a random sampling strategy; 4, embedding a multi-scale attention aggregation module into the jump connection of the encoder and the decoder, fusing the features through a channel and a space attention unit, and achieving the self-adaptive weight weighting of the features of each layer of the encoder; and 5, recovering the resolution of the original point cloud by adopting nearest neighbor interpolation up-sampling, outputting a segmentation semantic tag, and obtaining a high-quality point cloud target. According to the method, the operation time of the battery replacement robot is shortened, and the balance problem of large-scale target point cloud segmentation speed and precision is solved.