180 results about "Invariant feature" patented technology

The invention discloses a metal surface quality detection method and system, and relates to the technical field of metal surface quality detection. The method is used for solving the problems of low microdefect detection precision, weak technological parameter relevance and closed-loop control deficiency of the high-reflection surface. The metal surface is irradiated through multi-angle coherent light field serialization, the phase offset of interference fringes is analyzed to generate three-dimensional shape data, and reflection noise interference is restrained. Defect depth gradient is extracted based on dynamic segmentation of process parameter constraint, deposition temperature and pressure deviation are quantified through deconvolution calculation, and process deviation feature distribution is constructed. Finite element simulation is utilized to generate a process-morphology mapping atlas library, cross-domain invariance features are extracted through depth constraint manifold alignment and comparative learning, and a causal correlation model of defect types and process parameters is established. And dynamically adjusting process parameters according to the weight gradient, and reflowing data to update the manifold rule. And high-precision three-dimensional defect detection, process deviation traceability and adaptive parameter optimization are realized.

The invention discloses a self-adaptive cross-domain remote sensing image semantic segmentation method based on an unsupervised domain, and the method comprises the steps: pre-training an encoder-decoder segmentation skeleton through a source domain image and a semantic tag of the source domain image; the difference between the domain invariant feature and the specific feature of the target domain is measured through the difference loss, and the reconstruction loss is applied to prevent information loss; the domain invariant high-level features of the source domain and the target domain are aligned based on an antagonism method; respectively processing high-layer and shallow-layer invariant features of a target domain through a main decoder and an auxiliary decoder of the decoders, and ensuring that prediction results of different levels of features are consistent by utilizing consistency loss; and fusing the domain-invariant high-level features and the specific features of the target domain to generate target domain fusion features, and optimizing the cross-domain adaptability of the decoder by aligning the prediction result of the target domain fusion features with the prediction result of the domain-invariant features. According to the method, the segmentation performance on the target domain is remarkably improved through feature decoupling and feature representation enhancement in combination with target domain self-learning based on pseudo labels.

The invention relates to an industrial part defect classification method capable of realizing inter-domain category self-adaption, which comprises the following steps of: acquiring a source domain data set with label information and a target domain data set without label information, and preprocessing the source domain data set and the target domain data set; inputting a to-be-detected target domain sample into the trained neural network detection model for defect detection; and removing a domain adaptation structure which is not needed in the domain adaptation detection network, and carrying out defect detection on the target domain scene. According to the industrial part defect classification method and device for realizing inter-domain category self-adaption, the processor and the computer readable storage medium, the domain invariant feature information aiming at the category is decoupled from the middle layer of the feature extraction network and is fused into the original feature so as to enhance the classification capability; an ELA attention mechanism is added to solve the key problems of small defect size and difficult feature extraction of part detection, and finally, class labels are taken as conditions during classification, inter-domain alignment is carried out for classes, and the cross-domain classification capability is further improved.

The invention discloses a remote sensing image omni-directional target detection algorithm based on a multilayer feature interaction pyramid and a lightweight enhanced detection head, and belongs to the field of computer vision. The method comprises the following steps: 1, preprocessing a remote sensing image data set; 2, building a remote sensing image omnibearing target detection model: designing a multi-layer feature interaction pyramid, obtaining an intermediate feature map and a fusion feature map by aggregating multi-layer feature maps, realizing cross-layer feature fusion, preventing information interaction from being limited between adjacent layers, generating rotation-invariant feature representation, and enhancing feature information of a rotating target; a lightweight enhanced detection head is constructed, the parameter quantity of the model is reduced by adopting a shared enhanced convolution strategy, and feature information is extracted through central difference convolution, so that the model can capture detail features of a rotating target; and the complexity of the model is reduced by adopting a Lamp pruning method on the basis that the precision is not lost. And 3, constructing a loss function of the model, and introducing a KLD divergence loss function to solve the problem of periodic angle change in rotating target detection. And 4, iteratively training the model until the model reaches convergence, and obtaining the optimal weight. And 5, testing the test set according to the obtained optimal weight to obtain an evaluation result. According to the method, the parameter quantity and the calculation complexity of the model are reduced while the detection precision of the rotating target is ensured.

The invention provides a multi-modal sentiment analysis method and system based on main modal two-stage guidance, and relates to the technical field of sentiment analysis. Inputting the multi-modal data into a multi-modal sentiment analysis model, and extracting language, visual and acoustic features from the multi-modal data through a feature extraction module; semantically decoupling the multi-modal features into modal invariant features and modal unique features through a feature space distribution alignment module, and realizing feature distribution alignment dominated by language modals through alignment reconstruction constraints; performing self-attention modeling on the modal invariant feature through an attention enhancement module to obtain a first enhanced feature, and adaptively enhancing the visual and acoustic unique features through a cross-modal attention mechanism by taking the language unique feature as a dominant feature to obtain a second enhanced feature; the first enhanced feature and the second enhanced feature are fused through the emotion prediction module, an emotion intensity prediction result is obtained through regression prediction, and the accuracy and robustness of emotion analysis in a complex scene are improved.

The invention provides a large model tuning method and system based on multi-modal information and AI, and relates to the technical field of artificial intelligence, and the method comprises the steps: extracting multi-modal parameter features and interaction features, and dynamically adjusting a contribution coefficient to achieve feature migration, and obtaining a cross-modal fusion vector; aligning the modal features by adopting an adversarial mechanism to obtain domain invariant features; constructing a prompt chain based on the modal confidence, optimizing the prompt chain by using AI reinforcement learning, and iterating to converge to obtain an optimization result; and finally updating the large model parameters. According to the method, effective fusion of multi-modal information is realized, and the tuning effect and generalization ability of the man-machine interaction large model are improved.

The invention discloses a multi-modal emotion recognition method and system based on field generalization and a graph neural network, and relates to the technical field of emotion recognition. According to the technical key points, the method comprises the following steps: acquiring a data set, wherein the data set comprises three types of modal data: a text modal, a voice modal and a visual modal; performing feature extraction on the three types of modal data to obtain text features, voice features and image features; performing field generalization by taking each mode as a field, namely dividing features extracted by each mode into two parts, one part is used for extracting intra-domain invariant features, and the other part is used for extracting inter-domain invariant features; fusing the intra-domain invariant features and the inter-domain invariant features by using a graph neural network to obtain global information and local context information at the same time; inputting the fused features into a classification model for training; inputting the to-be-detected data into the trained classification model for classification, and obtaining an emotion recognition result. According to the method, the emotion recognition performance and robustness are improved.

The invention provides a user consumption behavior multi-dimensional portrait analysis method and system based on a neural network, and relates to the technical field of data analysis, and the method comprises the steps: obtaining user historical consumption behavior data, and constructing a basic feature vector; key time sequence features are determined through a sub-sequence dynamic pruning algorithm and entropy value weighted mapping; extracting sequence features by adopting a bidirectional long-short-term memory network; constructing a multi-task adversarial feature extraction network to obtain scene invariant features; performing feature fusion to obtain multi-dimensional combined features; constructing a feature index tree to calculate user similarity; and hierarchical clustering is carried out to obtain a consumption behavior portrait. According to the invention, high-precision user portraits are realized, and scene adaptability and calculation efficiency are improved.

The invention discloses an efficient method for unsupervised domain adaptive target detection, and relates to the technical field of image processing. According to the invention, a field query module is designed, and the characteristics output by an encoder are subjected to adversarial alignment, so that a detector can extract more domain invariant characteristics, in addition, a category prototype alignment module is provided, the category prototype characteristics of each category can be extracted from a decoder, category-known cross-domain characteristic alignment is realized, and the accuracy of the category-known cross-domain characteristic alignment is improved. In each training iteration, a local category prototype gradually generates a global category prototype, a detector is guided to realize global feature alignment through comparative learning and confrontation loss, and a large number of experimental results show that the proposed model realizes excellent detection performance on a multi-domain adaptive reference data set.

The invention discloses a cross-subject electroencephalogram emotion recognition method and system based on dynamic domain invariant representation decoupling and recombination, and belongs to the technical field of artificial intelligence. The invention provides a non-personalized decoupling and recombination framework for cross-subject electroencephalogram emotion recognition, and aims to separate emotion-related individual invariant features from cross-subject individual invariant features through complex electroencephalogram signal characterization obtained through dynamic decoupling, so that individual differences are eliminated while emotion classification performance is guaranteed. Specifically, the method comprises the following steps: carrying out original EEG data analysis and preprocessing by using MATLAB and Python MNE libraries; frequency spectrum and space features of EEG signals are extracted through a multi-channel frequency spectrum space self-attention mechanism module, and capture of emotional features is enhanced in combination with a self-attention mechanism and a cross-attention mechanism; a joint distribution alignment method based on a category prototype is adopted, and decoupled feature distribution is optimized, so that invariant features in subjects and invariant features among subjects have higher distinction degree in emotion classification; the optimized decoupling features are recombined through a linear network, the two features are coordinated to perform more sufficient emotion representation extraction, and emotion classification is performed through a multi-layer perceptron. According to the method, excellent cross-subject emotion recognition performance is obtained on multiple data sets, and the emotion recognition rate of the electroencephalogram signals in a cross-subject scene can be effectively improved.

The invention relates to the technical field of artificial intelligence, and discloses an unmanned aerial vehicle flight control-oriented intelligent obstacle recognition method, which comprises the following steps of constructing an adaptive wind disturbance fuzzy kernel based on unmanned aerial vehicle attitude data, and synthesizing a degraded image for training; a specific obstacle recognition model integrating a frequency domain and space domain joint feature decomposition module, a turbulence invariant feature enhancement module guided by physical prior information and a double-branch anti-fuzzy feature extraction network is adopted, and a fuzzy robustness contrast loss function is combined to carry out progressive training so as to learn feature representation insensitive to fuzziness; after the trained model is deployed, the wind disturbance fuzzy intensity of a real-time input image is estimated on line, the internal workflow of the model is dynamically adjusted according to the wind disturbance fuzzy intensity, the attention weight is adaptively adjusted, and the output of a trunk or an anti-fuzzy branch is selected. According to the invention, through coupling of the physical model and deep learning, the recognition robustness, accuracy and stability of the model in a dynamic wind disturbance environment are improved.

The invention provides a cross-domain fault diagnosis method based on a self-learning convolution domain adversarial network, and the method comprises the following steps: constructing a self-adaptive feature extractor based on self-learning convolution, and achieving the dynamic perception of signal features through a convolution parameter self-learning mechanism; constructing a self-learning convolution domain adversarial network comprising a self-learning convolution-based self-adaptive feature extractor, a domain classifier with a gradient inversion layer and a fault classifier; and inputting the source domain data set and the target domain data set into the self-learning convolutional domain adversarial network by adopting a cross-domain joint training strategy, synchronously optimizing cross-domain feature distribution confusion loss and source domain classification loss, and realizing feature alignment of the source domain and the target domain. According to the method, the advantage that self-learning convolution can dynamically adjust convolution kernel parameters according to different domain data is fully utilized, so that the model better adapts to cross-domain data distribution with difference in the feature extraction process, invariant features of a source domain and a target domain can be effectively extracted and aligned, and the cross-domain fault diagnosis performance of the model is remarkably improved.

The invention relates to the technical field of multi-modal data processing, and discloses a multi-modal sentiment analysis method and system based on feature decoupling and variational optimization, and the system comprises a feature decoupling module, a modal variational alignment module, an adversarial representation module and a multi-path interactive fusion network. According to the system, modal features are separated into modal specific features and modal invariant features through a feature decoupling technology, uncertainty modeling and re-parameterization processing are carried out on exclusive features by adopting a single-modal uncertainty perception fusion technology, and robustness of feature learning is enhanced by adopting an adversarial training mechanism for shared features. Model optimization is carried out by combining a cross-modal attention mechanism modeling inter-modal interaction relationship and a multi-level loss function, so that efficient feature extraction, robust feature fusion and accurate emotional intensity prediction of multi-modal data are realized.

The invention discloses a multi-modal image matching method and system for three-dimensional phase orientation invariant feature transformation, and belongs to the field of image processing. The method comprises the following steps of: firstly, constructing a multi-direction difference graph through a difference filter, then calculating three-dimensional phase tensor characteristics by utilizing a three-dimensional phase consistency model, and further calculating a phase difference graph; thirdly, calculating the maximum direction value of the phase difference diagram to obtain a significant difference phase feature; and finally, obtaining feature points in the differential phase torque space. Then, optimal feature indexing is carried out under multi-scale and multi-direction three-dimensional phase features, the positions of feature points are positioned, a histogram of oriented gradients of local areas of the feature points is calculated and counted, and then a three-dimensional phase orientation feature descriptor is constructed; finally, Euclidean distance is adopted as a matching measure, homonymy points are obtained by describing the ratio of the nearest neighbor distance to the next nearest neighbor distance through three-dimensional phase orientation features, mismatching points are removed through a random sampling consistency algorithm, and multi-modal image matching is completed.

The invention provides a multi-source domain invariant acoustic feature extraction method and system for an equipment operation state, and belongs to the technical field of equipment maintenance, and the method comprises the steps: constructing a multi-source domain invariant acoustic feature extraction network based on a DANN model, and the network comprises a feature extractor, a classifier, a domain discriminator and a multi-domain acoustic feature class boundary constraint module; the feature extractor extracts high-dimensional features of sound signals, the classifier carries out fault mode recognition, and the domain discriminator realizes domain prediction. The multi-domain constraint module generates an embedding space, and calculates the maximum mean value difference, the local maximum mean value difference and the Euclidean distance among different source domain features. The network constructs a loss function by taking minimization of inter-domain difference and classification loss and maximization of inter-domain distance and domain discrimination loss as targets, and carries out adversarial training through multi-source tagged acoustic data. According to the method, domain invariant features are extracted by using adversarial learning and hidden space constraint alignment of multi-source domain acoustic features, so that the influence of feature offset under a cross-working-condition condition is effectively reduced, and the fault mode recognition accuracy is improved.

The invention discloses a series arc fault detection method based on DSCAN and feature fusion. The method comprises the following steps: collecting source domain data and target domain data; preprocessing data of a source domain and a target domain to obtain multi-dimensional features such as a grayscale image after GASF conversion, an IMF component after VMD adaptive decomposition, a frequency domain component obtained by DT-CWT and the like; constructing a four-channel feature fusion module, inputting the original current signals and the converted multi-dimensional features into corresponding channel networks, extracting signal features, and splicing the signal features; and constructing a DSCAN network to learn domain invariant features of a source domain and a target domain, and detecting a series arc fault through a model. Through multi-dimensional feature fusion and cross-domain transfer learning, arc fault features in a complex load scene are effectively extracted, the robustness of the model to different load environments is enhanced, the problem that a traditional method is insufficient in detection precision under mixed loads is solved, efficient and accurate detection of series arc faults is achieved, and the fault detection efficiency is improved. And a reliable technical scheme is provided for electrical safety monitoring.

A method for training a neural network to extract domain invariant features suitable for image registration comprises collecting a first set of multi-modal images comprising a first image of at least one first modality and a corresponding second image of at least one second modality (3). A first feature is extracted from at least one first image while a second feature is extracted from at least one second image (5, 7) using a feature extraction subnet of a neural network. Domain-invariant loss and homography loss are estimated for the images (9, 11), and the neural network is trained to minimize a multi-objective loss function comprising domain-invariant embedding loss and homography loss (13).

Provided is a cross-period brain fingerprint identification method with paradigm adaptive decoupling and a system thereof. The method includes: extracting a feature representation from original electroencephalogram data by a feature extractor; effectively separating identity-related features and paradigm task-related features from highly coupled electroencephalogram information; and further learning domain invariant features with identity identification ability through domain adversarial training. According to the method, three decouplers are introduced to perform feature decoupling on features extracted by a feature extraction module. At the same time, three classifiers are introduced to pass through a domain label, an identity label and a paradigm task label, and the decouplers are guided to decouple paradigm task features and identity features effectively through adversarial training.

The invention provides a generalized SAR target detection method for a unified framework of a cross-source scene, and the method comprises the following steps: 1, defining the scattering features of an SAR image, matching the scattering features of the SAR image through dynamic statistics, and expanding a domain feature space; and 2, designing a domain invariant feature and domain specific feature decoupling module for the feature space, and removing redundant domain discrimination information to realize final feature optimization. Through the optimized feature distribution contour, domain offset can be eliminated, and downstream detection tasks can be completed at a detection head of the baseline network. Experimental results show that the proposed model framework runs on a Faster RCNN baseline network, and a target detection result is obtained through a downstream detection head of the Faster RCNN baseline network.

The invention discloses a domain adaptive DINO model target detection method based on feature fusion and alignment, and the method comprises the steps: sequentially inputting features outputted by a backbone network into a first feature fusion module and a first domain discriminator, and obtaining the image-level adversarial training loss of the backbone network; inputting the features output by the encoder into a second field discriminator to obtain encoder pixel-level adversarial training loss; sequentially inputting the features output by the decoder into a second feature fusion module and a third domain discriminator to obtain decoder target-level adversarial training loss; according to the backbone network image level adversarial training loss, the encoder pixel level adversarial training loss, the decoder target level adversarial training loss and the source domain detection loss, the DINO model learns cross-domain invariant features, and cross-domain target detection of target domain image data is completed. According to the method, the data distribution offset can be reduced, and the transferable feature representation is learned, so that the model also has good detection performance in the target domain.

This invention belongs to the field of remote sensing data processing technology and discloses a method, system, terminal, and storage medium for multimodal remote sensing data representation and fusion. The method includes: acquiring and preprocessing a multimodal image dataset to obtain an initial image dataset; inputting the initial image dataset into corresponding autoencoders to obtain corresponding time-invariant features, and pre-training the corresponding autoencoders using feature reconstruction; classifying the obtained features using the geographic location information of the initial image data to generate positive and negative sample pairs, and adjusting the trained autoencoders using a contrastive loss function constrained by cosine similarity; based on the adjusted autoencoders, extracting features with spatiotemporal invariant information from the corresponding modal images, and performing feature fusion using a spatiotemporal invariant information integration method that considers modal importance; and outputting the fused image features. This invention improves the fusion accuracy of multimodal remote sensing image features.

The invention provides a multi-source-domain distributed migration target identification method and system based on evidence fusion, and the method comprises the steps: inputting image data through an initialization module, and setting the maximum number of iterations and an importance coefficient; the feature learning module performs domain invariant feature learning and migration classification on the image data according to the setting to obtain a fusion soft classification result of the target domain image; a weight calculation module estimates the relative weight of the source domain based on the accuracy and the relative weight of the source domain based on the distribution distance, and calculates the final weight of the source domain; and the evidence fusion module discounts the fusion soft classification result according to the final weight of the source domain, weighted evidence fusion is carried out, and a category decision is carried out to obtain the category of the target image to be identified. According to the method, the problem that the classification performance is limited due to insufficient multi-domain information mining in multi-source migration classification is solved, an accurate recognition result is still obtained under the condition that the inter-domain difference is large, the classification accuracy of target domain samples is improved, and the negative influence of information source quality on fusion is effectively reduced.

The invention provides a student learning condition and writing habit evaluation method based on multivariate data analysis, and belongs to the technical field of education, and the method comprises the steps: 1, collecting motion data and pressure sensing data of writing equipment, and semantic vector data of text content; 2, track invariant features of handwriting are extracted based on the motion data, dynamic pen pressure distribution features are determined based on the pressure sensing data and the track invariant features, and feature alignment is conducted on the pressure sensing data and semantic vector data; step 3, extracting continuous homology features from the track invariant features, constructing a persistent interval graph, and further generating a state transition matrix of writing behaviors in combination with a feature alignment result; and 4, determining a persistent feature vector based on the trace invariant feature of the handwriting, and determining a writing habit evaluation result based on the state transition matrix and the persistent feature vector. A more accurate learning evaluation tool is provided for an educator, and personalized learning of students is promoted.

The invention discloses a remote sensing change interpretation deep network combining distributed sampling and feature decoupling, and belongs to the technical field of remote sensing image processing. In order to solve the problem of inaccurate classification of remote sensing change pixels caused by fuzzy semantic boundaries easily caused by mixed feature extraction, the mixed features are decoupled into change and invariant features through joint distribution sampling so as to complete remote sensing change interpretation. In the training stage, the posterior distribution of the decoupled features is calibrated and learned through labels and is used for training a change prior generator; feature decoupling is realized by combining posterior distribution and a feature separator, and features are further gathered through prototype learning; a super-expectation push-pull loss regular term is provided, and the inter-class distance is increased by improving the prediction expectation push-pull positive and negative sample features to the farther end. In the test stage, remote sensing image change detection is completed through modules such as a feature separator and a change detection head without posterior distribution support. Experiments prove that the method has remarkable effects on qualitative and quantitative indexes.

The invention discloses a training and online diagnosis method and system for a ship fault diagnosis model. The method comprises the following steps: generating a dynamic weight based on multi-source domain and target domain data; using the weight to guide a feature extractor and a domain discriminator to carry out adversarial training so as to learn domain invariant features; fusing the weighted supervision loss and domain adversarial loss, and combining a target domain regularization item optimization model; screening a target domain high-confidence sample through a dynamic threshold value to generate a pseudo label and performing iterative training; and finally, performing real-time fault diagnosis by utilizing the trained model, and adaptively updating model parameters according to online data distribution change. According to the method, the problem of negative migration in multi-source domain fusion is effectively solved, and the generalization ability, the diagnosis precision and the long-term operation stability of the model in the target domain are improved.

The invention discloses a domain offset mitigation method and system based on a power equipment image, and relates to the technical field of cross-domain detection, and the method comprises the steps: constructing a generalization model based on insulator defect detection; based on an IRM risk extrapolation method, regularization items are added to extract invariant features, so that the aligned sample loss of each domain is the same; by measuring the Zensen-Shannon divergence of each domain feature of the selected layer in the generalization model, the distribution difference in the feature space is minimized, so that the similarity of the feature space is the same, and the domain offset between the source domain and the target domain is reduced. According to the method, the generalization of real data is enhanced by reducing the difference between the synthetic domain and the real domain to the greatest extent, so that the problem of domain migration is solved.

The present invention discloses a visual odometry method and system based on image depth prediction and monocular geometry. The method includes inputting two consecutive image frames, detecting and describing local features of the images using the Scale-Invariant Feature Transform (SIFT) algorithm, and then using the Fast Library for Approximate Nearest Neighbors (FLANN) algorithm to match corresponding feature point pairs between the two frames; solving for the essential matrix using epipolar geometry constraints to obtain the relative pose transformation of the camera; constructing and training a monocular depth prediction model to predict dense depth information for each input image frame; if the number of valid depth information pairs formed by two consecutive image frames is greater than a given threshold, using triangulation to estimate the scale factor to obtain the corrected relative pose transformation, otherwise using a combination of the Perspective-n-Point (PnP) projection algorithm, the Random Sample Consensus (RANSAC) algorithm, and local non-linear optimization to solve for the absolute pose transformation. The present invention effectively integrates the advantages of deep learning and traditional geometric methods, can adapt to dynamic environments, and improves the robustness and accuracy of monocular visual odometry.

This invention provides a method, apparatus, device, and storage medium for fault diagnosis of rotating components of a crane, relating to the field of mechanical fault diagnosis technology. The method includes: processing source and target domain data of the crane's rotating components in a unified format to output a normalized vibration signal; aligning feature distributions through a shared encoder and adversarial training to output domain-invariant features; constructing conditional embedding vectors based on the domain-invariant features and generating target domain fault data using a diffusion model; combining the generated fault data with real data to form a diagnostic dataset, extracting multimodal features, fusing them, and outputting the fault diagnosis result through a classifier. This invention effectively solves the problem of scarce crane fault data through a diffusion model and, combined with multimodal feature fusion technology, improves the accuracy, robustness, and cross-domain generalization ability of fault diagnosis.

The application discloses a hyperspectral wetland image classification method based on a graph capsule neural network, which comprises the following steps: S1, learning feature transformation is performed on an adversarial domain self-adaptive framework, so that source domain samples and target domain samples of a hyperspectral wetland image are matched in features; S2, a graph capsule neural domain self-adaptive network structure is constructed, domain-invariant features and domain-related features are extracted, and transferable features are discovered and shared across domains; and S3, a coupling structure two-classifier is designed, the two-classifier is trained by using the source domain samples, classification differences of the target domain samples are maximized, and precise classification of the hyperspectral wetland image is realized by identifying a classification boundary. Meanwhile, the application discovers transferable knowledge and realizes cross-domain sharing, enhances effective discrimination of a class boundary, and finally realizes precise classification of the hyperspectral wetland image under conditions of unknown regions, complex scenes, and lack, deficiency and imbalance of data types.

The application discloses a feature extraction method and system based on difficult sample mining and multi-granularity division, which is used for a cross-view geographical image retrieval task. The method first preprocesses cross-view street view images and satellite images, and uses a generative model to generate cross-view images, reducing the visual difference between different view images. Then a two-stage difficult sample mining model is constructed, including a sampling strategy based on geographical location and visual similarity, mining difficult negative samples in different ranges, and enhancing the inter-class discrimination ability. Then a multi-granularity feature division module is introduced, the image features are extracted through a ResNet50 backbone network, and the features are divided and fused according to different granularities to obtain rich and robust view-invariant feature representation. Finally, the satellite image to be retrieved is input into the trained feature extraction model, the features are extracted, and similarity matching is performed with the street view image library to obtain the cross-view retrieval result.