67 results about "Phase congruency" patented technology

The invention belongs to the technical field of image processing, and particularly relates to a screw fastening quality evaluation method and system based on image features, and the method comprises the steps: collecting a screw fastening image, and obtaining a complex response diagram through a Log-Gabor filter; obtaining a phase congruency diagram according to each complex response diagram; acquiring gradient directions of pixel points in the phase consistency graph, and constructing an accumulator graph according to the gradient directions and offset points in a preset radius range; according to salient points in the accumulator graph and local signal-to-noise ratios of the salient points, screw positioning points are determined through two-dimensional quadratic polynomial function fitting; the torque value of the screw is obtained according to the screw positioning point driving torque measuring device, and whether the screw is fastened or not is judged. According to the method, the screw is positioned by using phase consistency in the frequency domain, so that the interference of specular reflection and extreme shadow on positioning is overcome, the accuracy of screw positioning is improved, and screw fastening quality evaluation is assisted.

The invention relates to the technical field of deep foundation pit monitoring, and discloses a deep foundation pit micro-deformation InSAR remote sensing method and system and a storage medium. The method comprises the following steps: synchronously acquiring data through a ground radar and an InSAR satellite, and obtaining a point target set and an original scattering characteristic sequence; obtaining a net deformation phase sequence without influence through registration residual correction, phase deduction and weighted fusion processing, and carrying out phase consistency quality control and unwrapping processing to obtain an accumulated deformation quantity sequence; constructing a deformation mechanism correlation characteristic matrix based on correlation analysis of the cumulative deformation quantity sequence and radar reflection intensity amplitude attenuation characteristics; and fusing the deformation mechanism correlation feature matrix and the real-time differential interference image group, extracting deformation rate and phase gradient change, and further generating a deep foundation pit local micro-deformation risk early warning index set. According to the invention, the tiny deformation of the surrounding area of the foundation pit can be effectively detected, potential risks are warned in advance, and the safety and monitoring efficiency in the construction process are improved.

The invention relates to the technical field of medical data processing, and provides a medical image data processing method based on deep learning, which comprises the following steps of: acquiring data according to an acquisition template, extracting a pulse time sequence by self-adaptive threshold peak detection, calculating an instantaneous phase according to linear interpolation, calculating a statistical magnitude, comparing a quantitative index with a preset threshold value, and calculating a pulse time sequence according to the statistical magnitude. Judging a steady state by combining a peak loss rate and an abrupt change detection rule, and calculating phase consistency between channels for verification; the method comprises the following steps of: splitting acquired data according to a concept entity to form a data relation model, implementing rapid global rigid estimation and applying affine transformation, estimating a pixel-level displacement field by adopting a pyramid dense optical flow network, applying the displacement field to an original pixel, and performing time domain fusion by taking optical flow confidence and a registration residual error as weights; and cutting the short-time image stabilization sequence after registration compensation, and outputting a pixel-level risk thermodynamic diagram, a candidate focus list and each output confidence interval by taking a hybrid network of a convolution front end and a space-time Transform backbone as a prediction model.

The invention relates to the technical field of image recognition, and discloses an image recognition method for cross-domain access authentication, which comprises the following steps: acquiring a to-be-authenticated image uploaded by an image acquisition node; converting the to-be-authenticated image to a frequency domain, and extracting a plurality of response signals in multiple preset directions and scales; calculating a phase consistency measure by using the real part and imaginary part response components, and generating a phase consistency feature map; carrying out statistics on local phase distribution of the phase consistency feature map, determining a phase distribution entropy representing physical attributes of an access medium, and carrying out weight reduction suppression on the phase consistency feature map according to the phase distribution entropy; according to the method, phase interference introduced by an unnatural display medium is recognized and suppressed, in-vivo detection is achieved in identity verification, and the influence of heterogeneous imaging amplitude distortion on recognition precision is eliminated.

The invention provides a circular curve burr detection method and system based on multi-scale phase consistency and a contour model, and a medium, and the method comprises the steps: constructing a multi-scale Log-Gabor filter group, and carrying out the convolution calculation of an input image, and obtaining a response vector; calculating phase consistency measurement of each pixel point, extracting an initial edge point set, fitting an initial circular contour, performing energy minimization evolution on the initial circular contour based on an active contour model, and fitting a real contour of burrs; constructing a two-dimensional feature vector; carrying out fusion discrimination on the two-dimensional feature vectors based on a linear weighted score function, and outputting a burr detection result; according to the method, phase consistency is used as a core image feature, so that the detection algorithm has natural invariance for uneven illumination, brightness change and contrast difference, and under the scene of failure caused by poor image quality in the prior art, the contour edge can still be stably and completely extracted, and the omission ratio is remarkably reduced.

The present application relates to the technical field of image data processing, and more particularly to a numerical control lathe tool wear real-time detection method and system based on machine vision, which comprises the following steps: positioning a region of interest of a tool image; for the pixel points in the region, fusing the gradient amplitude, phase consistency and local structure anisotropy thereof, and calculating an initial boundary probability; constructing an exploration path and performing weighted integration on the initial boundary probability thereon to evaluate the path consistency; fusing the initial boundary probability and the path consistency score to obtain a final boundary probability, and extracting a wear profile therefrom to realize evaluation of the tool wear state. The present application solves the problem of inaccurate wear profile extraction caused by insufficient robustness in a complex machining environment, and improves the accuracy and reliability of detection.

The invention relates to the technical field of image segmentation, and provides a gastrointestinal tract endoscope image segmentation method, which comprises the following steps: carrying out time sequence preprocessing on an obtained intestinal tract video, eliminating fuzzy and shielded invalid frames, constructing a rhythm signal based on a multi-scale time sequence visual feature, and estimating a creeping phase; outputting a phase time sequence and a phase label of the video; selecting a representative phase key frame, performing pixel-level focus segmentation on the key frame, and calculating a pixel-level uncertainty score to obtain a key frame segmentation result with high confidence; carrying out time sequence consistency recovery on the segmentation mask of the key frame based on phase consistency mapping and phase alignment, and recovering an uninterpretable section caused by shielding with the assistance of adaptive segmentation correction of motion perception; and carrying out time sequence aggregation on a cross-frame segmentation result, extracting fragment-level statistical representation and confidence, forming a fragment-level suspicious area report, and outputting structured data for clinical auditing and index retrieval.

The application discloses a silicon carbide grain size detection method based on image processing and relates to the technical field of semiconductor material quality detection. A first feature map is generated by calculating the phase consistency distribution of a silicon carbide metallographic image to suppress scratch interference. A topological anchor sequence is identified by using the eigenvalues of a Hessian matrix, and a balanced first gradient potential field matrix is generated by combining local entropy density to suppress polycrystalline contrast fluctuation. Under the constraint of the topological anchor, an optimization evolution is performed by using a path cost function with curvature penalty and energy saturation characteristics, physical inertia closure of a grain boundary signal fracture is realized, and a closed grid vector is generated. Twin grain boundaries are identified and logically merged based on geometric parameters, and grain size detection data consistent with the physical structure of the material is output.

The invention discloses a broadband random harmonic compression analysis method based on LSTM and multi-feature fusion, and belongs to the technical field of electric digital data processing. Comprising the following steps: collecting a broadband random harmonic signal for time domain analysis, and extracting a period; judging whether each period signal can be compressed or retained or not by using LSTM (Long Short Term Memory); carrying out down-sampling and coding compression on the compressible periodic section signal; decoding and reconstructing the signal through an LSTM decoder; the decoded and reconstructed signals are preprocessed, multi-dimensional features of the signals are extracted, and evaluation indexes are normalized; calculating the compression ratio of the dynamic fusion index and the signal; and dynamically displaying the evaluation index and the dynamic fusion index. According to the method, self-adaptive decision of compression or not is realized, the compression ratio is improved, and meanwhile, the spectrum fidelity and phase consistency are also kept; and parts which do not need to be compressed are reserved as original samples, so that real-time performance and error control are both considered.

This invention belongs to the field of image data processing technology, specifically relating to a connector position offset compensation method and system based on image processing. The method includes: obtaining a connector image and all its image blocks; determining the direction dominance index of each image block, determining the highlight pixels within the image block and the distance variance of their coordinates, determining the highlight aggregation coefficient, filtering direction requirement, and the optimized number of filtering directions for each image block; determining the edge features of the connector image, determining the geometric center coordinates and deflection angle of the connector in the image, determining the position offset and angular offset of the connector in the image, and performing fine-tuning compensation. This invention, through adaptive filtering direction numbering, eliminates computational redundancy in the phase consistency algorithm, eliminates interference from highlights on edge positioning, and ensures high accuracy and real-time performance of connector position offset compensation.

The application provides a surface feature-based anti-counterfeiting information generation method, device, equipment and medium, which comprises the following steps: firstly, a microstructure intensity map is obtained by extracting surface micro-relief features through a phase consistency model; a first height map is obtained by reconstructing an initial surface height field by using a MiDaS depth estimation network; a local displacement field caused by embedding a preset mark is quantified by using a RAFT optical flow model; an initial stress disturbance distribution is constructed by fusing the height field and the displacement field as physical input of a coupling prediction model, and then the surface morphology of a concrete sample after solidification is predicted; finally, the microstructure intensity map, the local displacement field and the first height map are taken as initial surface features at the sampling time, and the initial surface features and target surface features are encoded as anti-counterfeiting information, so that the reliability of sample identification is effectively improved.

The present application relates to the technical field of medical data processing, and provides a medical image data processing method based on deep learning, which comprises the following steps: collecting data according to a collection template, extracting pulse timing by adaptive threshold peak detection, calculating instantaneous phase by linear interpolation, comparing quantitative indicators with preset threshold values, judging steady state by combining peak loss rate and mutation detection rules, and verifying by calculating inter-channel phase consistency; splitting the collected data according to concept entities to form a data relationship model, implementing fast global rigidity estimation, applying affine transformation, estimating pixel-level displacement field by using a pyramid dense optical flow network, applying the displacement field to original pixels, and performing time domain fusion by taking optical flow confidence and registration residual as weights; cropping the short-time steady image sequence after registration compensation, taking a mixed network of a convolution front end and a space-time Transformer backbone as a prediction model, and outputting a pixel-level risk heat map, a candidate lesion list and confidence intervals of each output.

The invention discloses a self-adaptive periodic signal noise reduction method and system for an event camera and electronic equipment, and the method comprises the steps: S1, a spatial domain self-adaptive filtering step: carrying out spatial domain filtering on an original event stream output by the event camera, and generating a spatially filtered event stream; s2, a time domain density cutting step: performing event density analysis on the event stream after spatial filtering on a time axis, identifying and retaining a time interval higher than a set event density, and generating an event stream after time cutting; and S3, a pixel-level period phase filtering step: performing pixel-level period estimation and phase consistency filtering on the event stream after time cutting to generate a de-noised event stream. According to the method, the problems that the signal-to-noise ratio is too low due to high-density background noise interference when an event camera detects periodic flicker signals and an existing fixed parameter filtering method is poor in adaptability are solved, and the high-fidelity extraction capability of weak periodic signals in a complex noise environment is improved.

The invention relates to the technical field of data analysis, in particular to a fault signal phase synchronization method and device based on a data fusion algorithm. The method comprises the following steps: preprocessing a collected bearing operation data signal; the preprocessed data signals are fused, and reliable phase offset estimation of each operation period and position is calculated; the fused data signals are processed through waveform matching and a phase correction algorithm, phase deviation caused by driving speed and load changes is eliminated, and data signals after phase synchronization are obtained; and based on the data signals after phase synchronization, bearing fault features are extracted, and a phase consistency coefficient is calculated. According to the technical scheme, the identifiability of fault signal characteristics is improved through multi-sensor data fusion and an advanced phase synchronization algorithm, and an important technical means is provided for fault diagnosis of rotating machinery.

The invention discloses a fatigue test crack detection method based on thermal imaging. In the periodic fatigue loading process, the surface temperature change of a sample is detected in real time through infrared thermal imaging, and an amplitude diagram and a phase diagram of weak temperature fluctuation are obtained through calculation; segmenting the sample through local phase consistency, prefabricating a gap and a background, and only performing subsequent analysis on a sample area so as to eliminate background interference; after the amplitude diagram is binarized, the boundary displacement influence is eliminated through a simple boundary displacement review strategy and by means of the boundary displacement direction and stress concentration point position relation, complex feature point matching and displacement estimation between images do not need to be conducted through the algorithm, the calculated amount is small, and reliability is high; the crack length is calculated according to the detected movement track of the crack tip feature point, the method does not depend on static crack features, the basic assumption that sample changes inevitably come from damage generation and development is utilized, and crack length evaluation is not affected even if wrong feature points are extracted through single-frame heat map analysis.

The invention provides an anti-fake information generation method and device based on surface features, equipment and a medium. The method comprises the steps that firstly, surface micro-fluctuation features are extracted through a phase consistency model to obtain a microstructure strength diagram; reconstructing the initial surface height field by using a MiDaS depth estimation network to obtain a first height map, and quantifying a local displacement field caused by embedding of a preset identifier through an RAFT optical flow model; the height field and the displacement field are fused to construct initial stress disturbance distribution to serve as physical input of a coupling prediction model, and then the surface appearance of the concrete sample after solidification is predicted; and finally, the microstructure strength diagram, the local displacement field and the first height diagram are used as initial surface features during sampling, and the initial surface features and the target surface features are coded into anti-counterfeiting information, so that the reliability of sample identification is effectively improved.

This invention provides a method and system for detecting minute defects in proton exchange membranes (PEMs) in fuel cells. The method includes acquiring an original image of the PEM to be detected, preprocessing the original image to obtain a preprocessed grayscale image, the preprocessing including at least grayscale conversion and resolution normalization, extracting full-image phase consistency features from the preprocessed grayscale image to generate a phase consistency feature map, wherein the phase consistency feature map characterizes the structural abrupt changes of pixels in the grayscale image, and then performing multi-scale residual enhancement processing based on the preprocessed grayscale image and the phase consistency feature map to generate a defect-enhanced image. Finally, adaptive threshold segmentation and contour analysis are performed on the defect-enhanced image to locate the defect region in the original image. This method can locate the defect region in the original image of the PEM, thus overcoming the problems in the prior art.

The application discloses a method for automatic defect detection and classification of sapphire substrates based on visual detection, relates to the technical field of visual detection and image processing, and comprises the following steps: performing adaptive threshold segmentation and phase consistency detection on a multi-modal image dataset, fusing the adaptive threshold segmentation and the phase consistency detection, and obtaining a joint defect probability graph; extracting a defect candidate region from the joint defect probability graph and generating a unified feature representation; inputting the unified feature representation into a double-branch neural network, fusing and classifying the unified feature representation through a cross-attention mechanism, outputting a defect classification result, reconstructing the defect classification result by using a NeRF algorithm, outputting a three-dimensional morphology of the defect, and generating a three-dimensional morphology atlas; generating a defect heat map based on the three-dimensional morphology atlas, performing space-time correlation analysis in combination with equipment process parameters, and outputting a sapphire substrate defect process analysis report. The application improves image contrast and defect characterization capability, thereby significantly enhancing the comprehensiveness and sensitivity of detection.

The invention provides a tooth image caries recognition method based on frequency domain enhancement, and relates to the technical field of image processing, the method adopts a dynamically adjusted annular band-pass filter to adaptively enhance an intermediate frequency component of a tooth image caries region in a frequency domain, so that the feature discrimination can be effectively improved, and the recognition accuracy is improved. Based on improved phase consistency edge optimization and gamma correction, the quality of a tooth image can be remarkably improved, multi-modal features are extracted through an improved local binary pattern and a gray-level co-occurrence matrix, and a dual-channel deep convolutional network is constructed to realize fusion of global and local features, so that the image quality is improved. The caries probability heat map of the tooth image can be accurately generated, and accurate positioning of the caries area can be realized by adopting adaptive threshold segmentation and morphological processing.

The embodiment of the invention relates to an image cross-modal saliency detection method and device, equipment and a medium. The method comprises the steps that a color image and a depth image corresponding to the color image are acquired; determining a first phase diagram corresponding to the color image, and determining a second phase diagram corresponding to the depth image; performing saliency phase updating processing based on preset distribution according to the color image, the depth image, the first phase diagram and the second phase diagram to obtain a saliency phase diagram; and according to the color image and the depth image, carrying out conversion from a frequency domain to a space domain on the saliency phase diagram to obtain a saliency map. According to the embodiment of the invention, on the basis that the discovered color image and the depth image have phase consistency of frequency spectrums with the saliency map, that is, the phase-related data obeys the preset distribution, the two modes of color and depth are taken as input, the saliency map is taken as output, and three different image modes are involved; and cross-modal saliency target detection with interpretability and relatively high accuracy is realized.

This application provides a fault identification method based on UHF phase characteristic changes, belonging to the field of fault identification technology. The method includes: acquiring the original electromagnetic signal of the target power equipment, performing preprocessing and phase demodulation to generate initial phase time series data; extracting the phase consistency coefficient, phase mutation index, and phase drift gradient within a preset time window to construct an initial three-dimensional phase feature vector; dynamically correcting the initial three-dimensional phase feature vector based on preset performance parameters of the phase-locked loop and historical data processing errors to generate an optimized three-dimensional phase feature vector; inputting the vector into a pre-trained fault classification model to calculate the matching probability between the output and each typical fault mode; performing multi-dimensional decision analysis based on the matching probability and the correlation weight matrix to identify and output the fault type identification result. This method solves the technical problems of insufficient identification accuracy and low precision in existing fault identification methods based on UHF phase characteristics.

The present application relates to the technical field of single crystal silicon wafer defect detection, more specifically, to a product single crystal defect locking system and method based on the principle of black cross light disappearance of silicon wafer, which is used to solve the problem that the prior art usually lacks effective verification of the mechanical orthogonal relationship between the fracture characteristics and the local stress direction, and only relies on image intensity or simple threshold to extract the fracture endpoint, which is easy to be interfered by noise, imaging non-uniformity or crystal microstructure, resulting in false fracture misjudgment; the present application fuses polarization response and mechanical priori through a black cross fracture stress verification module, positions the fracture endpoint by modulating the consistency of depth and phase, combines weighted robust fitting to inverse the local stress direction, and strictly verifies the fracture-stress vertical relationship by statistical test, finally takes the verified points as the boundary to drive anisotropic diffusion to generate a physically reasonable, noise-robust and full-field guided feature field, which provides reliable priori for high-precision stress reconstruction.

The application discloses a kind of ice thickness measurement method and device based on visual perception of ice storage coil, and the method comprises: setting annular or fan ring-shaped calibration disc on ice storage coil.The calibration disc is provided with multiple different color stripe regions arranged concentrically and along radial direction.Judge whether each stripe region is completely covered by ice layer according to the brightness change of calibration disc covered by ice layer in image, and each stripe region is divided into completely covered interval, partially covered interval and uncovered interval.For partially covered interval, the ice layer covering ratio of partially covered interval is obtained, and thus the ice layer thickness is obtained.The application provides the accuracy of ice thickness detection by stage ice thickness analysis, specifically, the ice layer completely covered by ice layer is determined by using the segmentation and brightness of different color stripe regions, and then for the partially covered stripe region, the comprehensive edge strength of integrated gradient edge strength and phase consistency edge strength is used for coverage ratio calculation.

The application provides an IC equipment alignment mark image enhancement method and an image enhancement device, wherein the method comprises: acquiring a process degradation mark image to be enhanced; performing real-time quality evaluation on the process degradation mark image to be enhanced by using a preset image quality evaluation algorithm; if the evaluation result is lower than a preset quality threshold, inputting the process degradation mark image to a trained cycle generative adversarial network to output an enhanced alignment mark image; the cycle generative adversarial network is trained in the following manner: training the network by using a non-paired training data set, and in the training process, optimizing the network by using a loss function comprising a structural similarity loss and an edge phase consistency loss, and training and reasoning by using a dynamic slicing strategy. By the application, the signal-to-noise ratio, edge sharpness and geometric position fidelity of the alignment mark image are improved without paired data, thereby improving the alignment accuracy.

This invention discloses a CBCT image restoration method based on a contrastive learning-guided diffusion model. The steps are as follows: Obtaining raw CBCT projection data and respiratory state signals for phase binning reconstruction to obtain phase binning reconstructed images and non-phase complete reconstruction prior images; extracting structured contrastive learning features from the phase binning reconstructed images, and combining phase text information and phase embedding information to obtain phase conditional semantic embeddings; performing semantically guided intermediate reconstruction based on the phase binning reconstructed images and phase conditional semantic embeddings to obtain intermediate reconstruction results; applying diffusion forward noise to the intermediate reconstruction results to obtain a noisy image; performing conditional diffusion reconstruction by combining the noisy image, phase conditional semantic embeddings, and non-phase complete reconstruction prior images; and iteratively optimizing based on projection domain data consistency constraints to obtain the CBCT restoration result. This invention eliminates the need for explicit motion field estimation, effectively suppresses fringe artifacts, and improves image structural details and phase consistency.

The invention belongs to the technical field of crossing of optical communication and machine vision, and particularly relates to a high-speed real-time visual navigation and image matching and positioning method for micron-sized features of an optical communication module, which comprises the following steps of: constructing a system architecture integrating multi-scale feature guidance, dynamic disturbance modeling and closed-loop feedback correction; hardware synchronous image acquisition, multi-scale coarse and fine matching, disturbance state estimation driven parameter adaptive adjustment, geometric consistency verification and thermal and mechanical coupling compensation are carried out to realize end-to-end deterministic time sequence control. Through a multi-scale feature guide mechanism, an invalid region is quickly eliminated in a coarse positioning stage, and the calculation overhead caused by global high-dimensional matching is avoided; the environmental interference is inhibited at an image generation source by utilizing dynamic disturbance modeling and physical layer parameter adjustment instead of depending on back-end algorithm compensation; weighted fusion of local phase consistency and gradient direction information is realized, so that the matching strategy can adapt to feature reliability changes under different noise levels.

The application discloses an infrared image diffusion super-resolution reconstruction method combining frequency domain enhancement and infrared perception, and belongs to the technical field of infrared image processing and deep learning. The method performs frequency domain enhancement on the infrared image through Fourier transform, and improves the recovery ability of high-frequency details and texture information. In combination with the characteristics of the infrared image, a physically guided diffusion scheduling strategy is designed to ensure the rationality of the generated image in thermal physical consistency. In order to further optimize the image quality, a spectrum consistency and phase consistency loss function is used to strengthen the feature consistency of the infrared image from the frequency domain, and the restoration effect of edges and textures is effectively enhanced. In addition, a self-supervised learning technology is introduced to automatically extract feature information from unlabeled infrared image data, thereby improving the super-resolution reconstruction performance. Through multi-level feature fusion and diffusion model optimization, the super-resolution reconstruction of the infrared image is realized, and the deficiencies of the traditional super-resolution technology, such as the loss of infrared image details and the physical consistency, are solved.

This invention provides a method and system for displaying veins based on image processing, belonging to the field of image processing. The method includes acquiring a near-infrared vein image to be processed, calculating the structure tensor, selecting the minimum eigenvalue map as a guide image, and calculating the pixel value variance. The original image is decomposed into an initial illumination component and an initial reflection component through guided filtering. The order of a two-dimensional least-squares surface fitting is determined based on the pixel value variance, and the initial illumination component is processed using the fitting to generate a background illumination model. A first vessel response map of the initial reflection component is obtained by calculating the Hessian matrix at multiple scales, and a second vessel response map is obtained by calculating local phase consistency. The two are fused to obtain an enhanced reflection component. A correction matrix is ​​generated by combining the background illumination model and gradient magnitude to correct the enhanced reflection component. The corrected reflection component is multiplied by the background illumination model to reconstruct the vein display image. The display image obtained by this invention is more in line with the needs of practical operation.

The present application provides a kind of dumpling micro crack identification method, system, the identification method includes collecting the image of the region to be measured, fusion pixel main curvature tensor and multi-direction phase consistency feature, constructs multi-scale feature pyramid;Rely on the multi-scale micro crack membership degree feature map generated by convolution network containing channel attention, weighted fusion obtains main membership degree atlas, calculates threshold screening pixel average membership degree as main membership value, screening pixel constructs binary connected domain and skeleton, extracts topological node, constructs local topological atlas according to the difference of node quantity;Solve Laplace matrix eigenvalue to obtain algebraic connectivity, combined with the eigenvalue proportion of covariance matrix to calculate linear structure degree, fusion three core indexes input classification decision model, realize accurate identification of dumpling surface micro crack.

The invention discloses a multi-modal image matching method based on phase consistency structure information distillation. According to the method, on a to-be-matched multi-modal image pair, firstly, stable structure prior information is obtained through phase consistency calculation, and the stable structure prior information and multi-scale features extracted by a convolutional neural network jointly form structure-guided feature expression; performing structural distillation training on the network by using structural consistency loss and cross-modal consistency loss, so that the obtained coarse-scale and fine-scale features have higher structural consistency under a cross-modal condition; on the basis, coarse-scale global matching is completed by adopting a self-attention and cross-attention alternately stacked feature transformation module to generate candidate matching pairs, and then local refinement is performed on the candidate matching pairs through fine-scale feature matching to obtain a final matching set. According to the method, the phase consistency structure prior and the deep learning matching framework are deeply fused, and the matching precision and the calculation efficiency are improved while the generalization of cross-modal matching is ensured.