88results about How to "Improve inference speed" patented technology

The invention discloses a thermodynamic system analysis method and system based on a physical information neural network. The method comprises the following steps: constructing a thermodynamic system mathematical model; converting a solving problem of the mathematical model of the thermodynamic system into an optimization problem taking a minimum control equation set residual error as a target function; constructing a physical information neural network; constructing a composite loss function based on the control equation set residual sum of squares, and performing end-to-end training on the physical information neural network to obtain a trained physical information neural network; taking the trained physical information neural network as an approximate solver, and for any given thermodynamic system input working condition, predicting a group of approximate solutions as initial predicted values through one-time forward propagation; and inputting the initial predicted value into a traditional numerical solver for accurate solving to obtain a to-be-solved state variable. According to the method, a feasible technical path is provided for applications with strict requirements on the analysis and calculation speed of the thermodynamic system, such as real-time simulation, online optimization and digital twinning.

The invention provides a lightweight human shape detection method and device based on improved RTDETR, a medium and a product, and the method comprises the steps: obtaining a data set for human shape detection, and dividing the data set into a training set, a verification set and a test set; an improved RTDETR model is constructed, an ES Block is used in the improved RTDETR model to replace a Basic Block in the RTDETR feature extraction network before improvement, and the ES Block comprises an OfficientVIT Block and a SimAM attention module which are connected in sequence; the NWD Loss is used as a loss function of the improved RTDETR model; training the improved RTDETR model by using a training set, verifying the improved RTDETR model by using a verification set and testing the improved RTDETR model by using a test set to obtain a trained improved RTDETR model; and inputting the to-be-detected image into the trained improved RTDETR model for reasoning to obtain a human shape detection result of the to-be-detected image. By means of the technical scheme, the human shape detection precision and small target human shape detection in a security and protection monitoring complex scene can be improved.

The present application provides a kind of topography probability distribution intelligent identification method based on lightweight convolutional neural network, constructs special terrain identification network by fusing the advantages of MobileNetV3 and ShuffleNetV2, improves model accuracy by combining channel attention mechanism and knowledge distillation technology, realizes the real-time inference of embedded platform using structured pruning and INT8 quantization.At the same time, this method not only outputs the probability distribution of five types of terrain and the confidence evaluation based on information entropy, but also realizes the unsupervised ROI automatic positioning through the improved Grad-CAM++ algorithm, without additional labeling of boundary box data.The present application can realize high-precision, high-robustness terrain identification with very low hardware cost and computing overhead, and provide reliable environment perception basis for exoskeleton proactive adaptive control.

The invention provides a method for optimizing an artificial intelligence model based on a tensor structure, and the method comprises the steps: replacing a linear mapping weight matrix in an initial artificial intelligence model with the tensor structure, and obtaining a target artificial intelligence model, the tensor structure comprising a target tensor and / or a target tensor network, the target artificial intelligence model is used for processing a target task, and the target task comprises at least one of natural language processing, logic and mathematical reasoning, code programming and multi-modal content generation. Structured compression of the artificial intelligence model is achieved through the target tensor and / or the target tensor network technology, the calculation complexity is effectively reduced while the model parameter quantity, hard disk occupation and video memory occupation are greatly reduced, and the reasoning and training speed of the model can be remarkably increased.

The invention discloses a sparse time sequence Bayesian network construction method and system based on power distribution network topological constraints, and the method specifically comprises the steps: constructing a power distribution network topological graph, and calculating an adjacent matrix between nodes and a k-hop neighborhood matrix Nk; based on the power distribution network topology distance information and the matrix Nk, generating a hard constraint rule, and constructing a topology dependence mask matrix M; calculating an electrical influence range of the fault point on surrounding nodes to obtain an electrical influence matrix E; if the electrical influence coefficient of one node on the other node is smaller than a set value, deleting the corresponding dependent edge; introducing a data source reliability matrix R, and performing hard deletion or soft weakening on edges with reliability lower than a threshold value; combining the matrixes M, E and R to synthesize a sparse structure matrix S; and taking the matrix S as a space skeleton, adding a time dimension autoregression edge, and constructing a complete sparse time sequence Bayesian network. According to the invention, by guiding the rarefaction of the network structure, the number of network edges and the number of parameters are effectively reduced, and the trainability and reasoning efficiency of the model are improved.

This invention discloses a heterogeneous acceleration system and method for analog modulation recognition of large-scale multi-channel signals, comprising: a CPU host processing end and a GPU device computing end; the CPU host processing end is used for asynchronous pipelined processing of data reading, task scheduling, and result distribution, realizing parallel execution of data reading, GPU computing, and result distribution on the time axis; the GPU device computing end completes the entire process of signal preprocessing and deep learning model inference within the GPU memory after the original signal is transmitted to the video memory via the PCIe bus, and does not exchange data with the CPU host processing end until the recognition result is generated. This invention eliminates data transport and I / O blocking through a three-stage asynchronous pipeline architecture and a closed-loop processing link in the entire video memory, and combines parallel preprocessing operators, inference acceleration engines, multi-stream scheduling mechanisms, and dynamic video memory pools to achieve high throughput and low latency recognition of large-scale multi-channel analog modulation signals.

The application relates to the field of image segmentation technology in computer vision, in particular to a zero-shot image segmentation method and system based on a three-dimensional model of an object. The method comprises the following steps: for any untrained object, a three-dimensional model of the object is rendered into a two-dimensional RGB reference image under a plurality of preset discrete viewing angles by using a graphics rendering engine, the reference image is input into a visual basic model DINOv3, a global feature vector representing semantic information of the object is extracted, and a reference feature library is constructed; a target image to be segmented is input into a visual basic model SAM2, and a plurality of candidate object masks are generated; for each candidate mask, an object image corresponding to the candidate mask is cropped from the target image, and the object image is input into the DINOv3 to extract a semantic feature vector of the candidate object; by calculating the cosine similarity between the candidate object feature vector and each template feature vector in the reference feature library, the top similar degrees with the highest values are selected, and an arithmetic mean value of the similar degrees is calculated, and the mean value is taken as the classification confidence of the candidate object mask; the class label of each candidate object mask is determined according to the confidence, and the class label of the target object and a corresponding pixel-level segmentation mask are output.

This invention provides an integrated method for pruning and controlling large-scale models of unmanned aerial vehicles (UAVs), belonging to the field of model lightweighting technology. The proposed method significantly reduces model size and computational load, greatly improves inference speed and control response performance, and enables the deployment of complex neural network controllers on resource-constrained airborne platforms. Simultaneously, the lightweight model maintains or approaches the control accuracy and stability of the original uncompressed model. This method has wide applicability and strong scalability, covering various scenarios from basic flight control to complex mission control, providing a feasible, efficient, and cost-effective solution for applying large models to airborne controllers of UAVs, and possesses high engineering practical value.

This invention provides a method, system, and related devices for adaptive state estimation of distribution network topology, belonging to the field of power system monitoring and control. It includes: obtaining measurement vectors and topology states based on acquired real-time operating data of the distribution network; using the measurement vectors and topology states as input to a physical information autoencoder, outputting estimated values ​​for the real and imaginary parts of the voltage of the distribution network; converting the obtained voltage real and imaginary part estimates into voltage amplitude and phase angle estimates to complete the adaptive state estimation of the distribution network topology. The physical information autoencoder sequentially includes an input projection layer, a first residual network module, a topology-aware latent module, a second residual network module, and an output projection layer. This invention integrates the advantages of physical constraints and data-driven approaches, possessing high estimation accuracy, excellent computational efficiency, strong physical consistency, and adaptive topology changes, effectively adapting to operating scenarios with frequent changes in distribution network topology.

The present application relates to the cross field of target detection and emotion recognition, and particularly relates to a facial expression recognition method for autistic children based on deep learning. The method comprises the following steps: embedding a three-dimensional attention (Tri-Attention) after a C3k2 module of a YOLOv11 backbone network to model the interaction relationship among height, width and channel dimensions in parallel, and enhancing the response to key expression regions such as eyes and corners of the mouth; introducing the three-dimensional attention (Tri-Attention) again at the multi-scale feature output end of the neck network to realize effective positioning of small-scale faces in a complex background and suppression of background noise; training and verifying the YOLOv11-TA model by using a public data set and a self-built autistic children data set respectively, and outputting the detection frame and expression category of each face; and comparing various indexes of the original YOLOv11 and the YOLOv11-TA on multiple data sets, and actually verifying that the method significantly improves the accuracy and robustness of the expression recognition of autistic children.

The invention discloses a reconfigurable crop image processing method and system based on software and hardware cooperation, and belongs to the technical field of embedded systems and image processing. The problems that in the prior art, an FPGA neural network accelerator model is solidified, the hardware resource utilization rate is low, and the detection precision and the reasoning speed are difficult to consider at the same time are solved. According to the method, the complexity score of an input image is calculated through an edge detection operator, a color histogram and a gray level co-occurrence matrix, a rapid detection mode, a fine recognition mode or a cooperative reasoning mode is selected according to the complexity score and delay constraint, and dynamic partition configuration is carried out on a processing unit array in the reconfigurable accelerator; in a cooperative reasoning mode, executing the lightweight target detection model through the first processing unit group to quickly detect and output a candidate box, extracting a region-of-interest feature map through the ROI cutting unit, routing the region-of-interest feature map to the second processing unit group, executing the high-precision target detection model to perform fine recognition, and fusing double-model output to generate a final detection result; the two models realize hardware resource sharing through a three-level weight storage architecture. The precision and speed balance capability of agricultural image detection are effectively improved, the resource utilization rate is improved, and the method can be applied to intelligent agricultural scenes such as crop disease recognition and fruit grading.

The invention discloses a coarse-grained power utilization data non-intrusive load monitoring method and system based on a YOLO deep neural network. According to the method, an encoding mode for converting a one-dimensional power consumption time sequence into a two-dimensional load image is constructed, so that coarse-grained power consumption data can be efficiently utilized by a mature convolutional neural network and a target detection framework; a set of automatic labeling mechanism based on power change characteristics is designed, and bounding boxes and category labels required by training are generated on the premise of not depending on manual labeling; a non-intrusive load monitoring problem is uniformly modeled as a target detection task, equipment category identification and operation time interval regression are realized through YOLOv5 or an improved structure thereof, and the overall identification performance is improved; and a complete system design scheme is given. According to the invention, the data acquisition and communication cost is obviously reduced, and the system has the advantages of simple model structure, high training efficiency, easy transplantation, engineering landing and the like, and can be widely applied to scenes of energy efficiency monitoring, demand response, power consumption behavior analysis and the like of families, buildings and parks.

A real-time target detection network training method based on prior information guidance, the target detection network comprises a feature extraction module, an importance scoring module, a feature aggregation module and a target detection module, the real-time target detection network training method comprises: acquiring a to-be-detected image of a target object; inputting the to-be-detected image into a pre-trained teacher supervision module to obtain a teacher attention map; inputting the to-be-detected image into the feature extraction module to obtain a first feature map; inputting the first feature map into the importance scoring module to obtain an importance score map; inputting the importance score map and the first feature map into the feature aggregation module to obtain a first feature sequence; inputting the first feature sequence into the target detection module to obtain a target detection result; and training the target detection network according to a total loss function. The above real-time target detection network training method enables the trained target detection network to accurately identify a foreground target during inference, thereby improving target detection precision.

This invention provides a rapid landslide area segmentation method based on high-resolution satellite remote sensing data and deep learning algorithms, belonging to the field of ground scene technology. The invention downsamples the original satellite remote sensing image to obtain a first image, extracts reflectance data from the red and near-infrared channels, calculates the reflectance difference and sum of reflectance values, and normalizes them to obtain a vegetation index to construct a vegetation layer. The first image and the vegetation layer are superimposed and stitched together, and feature enhancement is performed to obtain a vegetation feature map. This map is input into a convolutional neural network to output a landslide probability map. A two-dimensional coordinate set is constructed and cropped to obtain high-resolution test patches. The spatial gradient magnitude of the local vegetation index is calculated and weighted by a preset enhancement coefficient to generate edge-enhanced patches. The edge-enhanced patches are input into a semantic segmentation network to output a local landslide map, which is then backfilled into the initial image to obtain a landslide area segmentation map. This invention constructs a cascaded segmentation system to achieve efficient and accurate segmentation of landslide areas.

This invention discloses an intelligence generation method and system based on feature fingerprint storage and spatiotemporal geometric correction, belonging to the field of intelligent remote sensing image processing technology. The method involves: accessing heterogeneous data streams generated from multi-source remote sensing images in a synchronized time sequence; inputting a shared backbone network and a modality adaptation layer to generate a current feature stream in a unified semantic space; asynchronously retrieving historical baseline features from an in-memory feature fingerprint database based on geographic coordinates; fusing satellite imaging parameters with the current feature stream and inputting it into a spatial transformation network, resampling historical baseline features using a regression transformation matrix to achieve spatiotemporal geometric correction in the feature domain; weighted fusing of the current feature streams from each modality to generate a fused feature stream; and parallel interpretation and semantic encapsulation of the fused feature stream to generate a natural language intelligence report. This invention reduces memory usage and data throughput, shortens the intelligence generation cycle, reduces the false alarm rate, and achieves synergistic optimization of timeliness, accuracy, and automation.

The invention provides a multi-dimensional self-adaptive large model inference engine method and system, and relates to the technical field of artificial intelligence, and the method comprises the steps: collecting hardware identification information of calculation hardware, and calling a hardware fingerprint database for matching to generate a hardware type result and a hardware characteristic parameter; constructing a hardware abstraction layer, loading a hardware abstraction driving plug-in according to a hardware type result, and virtualizing computing hardware into a unified logic computing power unit through a hardware abstraction interface; when a large model reasoning task is triggered, dynamically routing to a back-end executor according to a hardware type result by querying a back-end registry; on the basis of hardware characteristic parameters and a hardware abstract interface, a hardware compiling tool chain is called through a rear-end executor, graph optimization and operator fusion are executed on a large model calculation graph, and therefore self-adaptive deployment and efficient reasoning of a large model on heterogeneous hardware can be achieved; the problems that in the prior art, heterogeneous hardware is insufficient in compatibility, configuration is tedious and rigid, and compilation optimization barriers and performance losses are caused during model cross-platform migration are solved.

The present disclosure provides a target detection model construction method, a target detection method, a device and a computing device to solve the problem of poor small object detection performance of single-stage target detectors in existing solutions. The target detection model construction method comprises: constructing a feature extraction network, the feature extraction network being configured to perform feature extraction on an input image to obtain a plurality of feature maps, the plurality of feature maps comprising a first feature map and a second feature map; and constructing a target detection network, the target detection network comprising a plurality of network layers corresponding to the plurality of feature maps, the plurality of network layers comprising a first network layer and a second network layer; the first network layer being configured to perform a query operation on the first feature map and transmit a query result obtained to the second network layer, the query result comprising a query point of a specific target in the first feature map; and the second network layer being configured to determine a mapping region of the query point in the second feature map and perform a detection operation in the mapping region to obtain a detection result.

The invention discloses a battery maintenance method based on a dynamic sparse neural network and a large language model, and belongs to the technical field of battery management. The method comprises the following steps: acquiring sensor time sequence data and an operation and maintenance text maintenance log of a battery pack, and respectively carrying out standardized preprocessing on the sensor time sequence data and the operation and maintenance text maintenance log; extracting time domain and frequency domain features of the preprocessed sensor time sequence data to obtain sensor features; carrying out semantic coding on the preprocessed text maintenance log by adopting a large language model subjected to battery domain knowledge fine adjustment to obtain text semantic features; performing feature fusion on the sensor features and the text semantic features based on a cross attention mechanism to generate fusion features; and inputting the fusion features into a dynamic sparse neural network which adaptively adjusts the sparseness of each layer of network according to the real-time health state of the battery, and outputting a battery fault probability and a fault root cause weight.

The application provides a scale-bias double-parameter self-calibration single-frame monocular absolute depth estimation method, mainly improving the precision of single-frame monocular absolute depth estimation and reducing the application cost; the application sets two key parameters of "scale (s)" and "bias (b)" which can be autonomously learned in the network, and uses the geometric relationship between the continuous frames of the video as a "scale" for self-calibration, constantly tries to synthesize the next frame with the depth of the previous frame and the motion of itself, and automatically adjusts the two parameters in the reverse direction by comparing the difference between the synthesized picture and the real picture, so as to form a self-consistent optimization closed loop, and finally drives the depth value output by the network to automatically converge to the real physical scale in "meters", so that the relative depth which is difficult to use directly can be reliably mapped to absolute depth, and the depth estimation precision of the application is significantly improved, and the power consumption cost is greatly reduced.

The invention discloses a progressive unmanned aerial vehicle detection method and system based on super-resolution guidance. According to the method, aiming at the problems that the low-resolution unmanned aerial vehicle image detection precision is low, and a traditional super-resolution + detection series framework has insufficient feature collaboration, false texture interference, calculation redundancy and the like, a dual-resolution feature collaboration framework is provided, and features are extracted in parallel through fixed scale and random scale branches; the fixed scale feature enhancement is guided by using the random scale feature; target area focusing is realized by adopting a candidate box mapping mechanism, only feature-level local super-division enhancement is carried out on a target area, and invalid calculation of full-image super-division is avoided; through cross-scale mapping loss and cosine similarity loss constraint, space consistency and semantic alignment of the double-branch target area are ensured. According to the method, deep coupling of super-resolution reconstruction and unmanned aerial vehicle detection is realized, the target identification degree of the low-resolution unmanned aerial vehicle is improved, false texture interference is effectively suppressed, and the detection precision is improved.

This invention relates to a multimodal orbital classification and recognition method, terminal, and medium based on edge AI. The method, applied to a mobile terminal, includes the following steps: acquiring a two-dimensional color image of a face to be recognized; performing region segmentation on the two-dimensional color image of the face to obtain a color image of the orbital region; acquiring three-dimensional face data corresponding to the two-dimensional color image of the face; obtaining a face depth image based on the three-dimensional face data; performing region segmentation on the face depth image to obtain a depth image of the orbital region; and using the color image of the orbital region and the depth image of the orbital region as input to a lightweight multimodal recognition model to obtain the classification result. Compared with existing technologies, this invention has advantages such as high reliability and small scale.

The application belongs to the technical field of image recognition segmentation, and discloses a medical image cell segmentation and tracking method, which comprises the following steps: step 1: data processing; feature extraction: the backbone part in the model is used to extract features from the preprocessed image, and the CSPDarknet structure is adopted in YOLOv8; step 3: FPN-PAN multi-scale feature fusion; step 4: Head prediction according to multi-scale features. The application realizes real-time tracking of the motion trajectory of cells by combining with a tracking algorithm such as deepsort. The method is mainly based on the YOLOv8 framework, and the Simam attention mechanism and the multi-scale proto method are adopted to optimize the model, so that the detection effect of YOLOv8 is further improved. The application can automatically complete the analysis and detection of medical images, is high in convenience and easy to use.

The embodiment of the application provides a kind of optical remote sensing image slice level change detection method and device, the method comprises: obtaining the image slice pair of two time phase optical remote sensing image;Feature extraction is carried out to image slice pair using twin encoding, and two time phase multilevel feature map is obtained, and twin encoding is obtained by processing based on sensitivity network pruning method;Two time phase global feature vector is obtained by using multilevel feature compression module to compress processing two time phase multilevel feature map;According to two time phase global feature vector, difference feature vector is obtained, and difference feature vector is input into decision network to carry out change detection, and the change detection result of image slice pair is obtained.The application is based on multilevel feature compression and network pruning technology to realize optical remote sensing image slice level change detection, while guaranteeing detection precision, maximum degree compression model complexity and improve reasoning speed.

The application provides a lightweight millimeter wave radar two-dimensional feature map classification method and system for FPGA hardware deployment, and solves the technical problems of poor real-time performance and high power consumption of existing millimeter wave radar two-dimensional feature map classification methods in edge computing scenarios. It includes obtaining a feature map and preprocessing it to obtain a preprocessed feature map, constructing a dataset, and dividing the dataset into a training set and a validation set; building a classification network model based on LeNet, training the classification network model using the training set, and obtaining a trained classification network model; verifying whether the performance of the trained classification network model meets the performance indicators using the validation set, and if so, obtaining a validated classification network model, deploying it to an FPGA, classifying the feature map based on the classification network model, and obtaining a classification result; otherwise, continue training to update the parameters of the classification network model until the performance after verification meets the performance indicators. The application can be widely applied in the field of image classification technology.

The embodiment of the invention provides a transmitting device, a receiving device, a system and a method of a terahertz wireless communication system, and the method comprises the steps: generating an initial signal modulation symbol based on an initial random bit sequence, and generating a transmitting end digital intermediate frequency signal based on the initial signal modulation symbol; converting the transmitting end digital intermediate frequency signal into a transmitting end analog intermediate frequency signal; generating a terahertz signal according to the local oscillator signal and the transmitting end analog intermediate frequency signal; generating a receiving end analog intermediate frequency signal according to the terahertz signal and the local oscillator signal, and converting the receiving end analog intermediate frequency signal into a receiving end digital intermediate frequency signal; and processing the digital intermediate frequency signal of the receiving end to obtain a baseband signal, processing the digital baseband signal to obtain a carrier phase synchronization digital signal, and processing and demapping the carrier phase synchronization digital signal based on the sliding window bidirectional LSTM neural network to obtain a target random bit sequence. Nonlinear distortion caused by a terahertz device in terahertz communication is relieved, and EVM is reduced.

This invention, belonging to the field of computational processing technology, proposes a method to accelerate sparse 3D convolution based on thread bundle alignment and memory access rearrangement. To address the issues of discontinuous memory access and low memory bandwidth utilization caused by the sparsity of convolution kernels in GPU architectures, this invention establishes a thread bundle aligned convolution kernel format method based on the direct convolution method. This method uses the width of the thread bundle as the basic unit for grouping and compression to obtain the thread bundle aligned convolution kernel format. Considering the single-instruction multithreading (SIMT) characteristics of graphics processors and the properties of convolution operations, this method is used to merge sparse convolution kernel data and perform vectorized memory access. A memory access conflict resolution rearrangement method is also established: Input data is stored in shared memory using a DHWC layout, and then a greedy algorithm is used to rearrange the weight storage order within each basic unit of the WAF, thereby eliminating shared memory conflicts for input data access by manipulating the address of each memory access.

The invention discloses a network planning method and device based on a hybrid reinforcement learning strategy, and the method comprises the steps: firstly, providing a basic network structure and an adjustment strategy through building a data link network topological structure model and a strategy support library; and then, an optimization strategy model is constructed based on multi-relation constraints, and efficient network planning decision optimization is performed through a hybrid reinforcement learning strategy so as to cope with a complex network environment. And then, generating an initial network planning structure through small sample data input, and finally, performing security review through the trusted network planning decision model to obtain a target network planning structure conforming to the standard. According to the method, a mixed strategy reinforcement learning method is adopted, high-dimensional target optimization and agile and efficient decision making under the condition of dynamic environment change are achieved, an intelligent planning technology based on a small data set and empirical knowledge is introduced, and the credibility and safety of network planning decision making are enhanced.

The invention relates to the technical field of artificial intelligence, and particularly provides a forklift tray tracking method, system and device based on a lightweight model and a medium, and the method comprises the steps: constructing a rotating target data set of a standard tray, and carrying out the targeted data enhancement and preprocessing; secondly, performing three-point improvement on a YOLOv12 algorithm: introducing a brand new StarNet backbone network, constructing a high-dimensional implicit feature space through star operation, and improving feature expression capability while reducing parameter quantity; a dynamic hybrid convolution module is designed in the neck network, multi-scale features are adaptively extracted by using multi-branch deep convolution and a dynamic weight fusion mechanism, and the flexibility of the model is enhanced; and a lightweight rotation detection head is provided, and the rotation angle of the tray is efficiently predicted while the parameter quantity is reduced and the small-batch training stability is improved through application group normalization and a shared convolution structure. And finally, combining the improved detection model with a ByteTrack tracking algorithm of the optimized adaptive rotating frame to form a complete identification tracking system.

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.

The application discloses an image splicing method based on unsupervised learning and an adversarial generative network, and comprises the following steps: (1) feeding two images to be spliced as a reference image and a target image into an alignment model, and obtaining a grid vertex offset through calculation; (2) projecting and transforming the target image according to the grid vertex offset to obtain an aligned target image; and (3) inputting the aligned target image and the reference image into a splicing model to splice and obtain a spliced image. The method can accurately realize the splicing of the images.