A multi-modal image matching method based on phase consistency structural information distillation

Abstract

The application discloses a multi-modal image matching method based on phase consistency structure information distillation. The method is used on a pair of multi-modal images to be matched. Firstly, stable structure prior information is obtained through phase consistency calculation, and is combined with multi-scale features extracted by a convolutional neural network to form structure-guided feature expression. Then, structure distillation training is performed on the network by using structure consistency loss and cross-modal consistency loss, so that the coarse scale and fine scale features obtained have stronger structure consistency under the cross-modal condition. On this basis, a feature transformation module stacked alternately by self-attention and cross-attention is used to complete coarse scale global matching to generate a candidate matching pair. Then, the candidate matching pair is refined locally by fine scale feature matching, and a final matching set is obtained. The application deeply integrates phase consistency structure prior and a deep learning matching framework, guarantees the generalization of cross-modal matching, and improves matching precision and calculation efficiency.

Description

Technical Field

[0001] This invention relates to the field of computer vision, specifically a multimodal image matching method based on phase consistency structural information distillation. Background Technology

[0002] Multimodal image matching aims to establish corresponding points or pixel-level relationships between images acquired from different imaging mechanisms or sensors. Multimodal images, such as visible light, infrared, and synthetic aperture radar, exhibit significant differences in radiation characteristics and noise distribution due to their different imaging mechanisms. The same object or structure often shows large differences in different modes, leading to problems such as unstable features and increased mismatches during the matching process.

[0003] Existing image matching methods can be broadly categorized into two types: traditional methods based on handcrafted features and deep learning-based methods. Traditional methods typically rely on features such as points and lines in images, constructing local feature descriptors and establishing correspondences based on structural information such as gradients and phase consistency. These methods offer a degree of interpretability and engineering maturity, exhibiting good generalization ability, but are generally slow. Deep learning-based methods learn feature representations and matching relationships through networks, achieving good results in single-modal or weakly differentiated scenarios, but their cross-modal generalization ability is insufficient. Therefore, how to effectively incorporate the structural information from traditional methods into the feature encoding and matching framework of deep learning, providing effective structural priors for deep learning-based matching methods, and improving the generalization ability of cross-modal matching while maintaining computational efficiency, still requires further research. Summary of the Invention

[0004] The technical problem to be solved by this invention is that traditional multimodal image matching methods are slow in extracting phase consistency structural information, while deep learning-based matching methods lack cross-modal matching generalization. This invention proposes a multimodal image matching method based on phase consistency structural information distillation, which aims to combine the cross-modal generalization of phase consistency features with the computational efficiency of deep learning-based matching methods to achieve stable and fast multimodal image matching.

[0005] The technical solution adopted by this invention to solve the above-mentioned technical problems is as follows: a multimodal image matching method based on phase consistency structural information distillation, comprising the following steps:

[0006] S1, Multi-scale feature encoding and phase-consistent structural information distillation:

[0007] S1-1, the multimodal image to be matched and Calculate the phase consistency feature map separately , , as a structural guide diagram;

[0008] S1-2, , Align to the scale set using bilinear interpolation, average pooling, or other equivalent downsampling methods. A coarse-scale phase consistency map was obtained. , and fine-scale phase consistency map , As information for teacher supervision;

[0009] S1-3, multimodal images and Inputting the convolutional neural network extracts multi-scale features, resulting in coarse-scale features from the two images. , and fine-scale features , ;

[0010] S1-4, construct a phase consistency prediction head at each scale, map the features to a phase consistency prediction map, and obtain the coarse-scale phase consistency prediction map. , and fine-scale phase consistency prediction map , ;

[0011] S1-5, Construct the loss function and perform structural distillation training to build a phase-consistent structural distillation loss:

[0012] (6),

[0013] in, , These are the weighting coefficients. For structural consistency loss, For cross-modal consistency loss;

[0014] S1-6, based on the loss function The feature extraction network and prediction head are jointly trained to obtain distilled coarse-scale features that include structural prior information. , and fine-scale features , ;

[0015] S2, Coarse-scale feature matching based on attention mechanism: Position encoding is introduced into the distilled coarse-scale features, and contextual information and cross-image interaction are performed through a feature transformation module to obtain fused coarse-scale features; a similarity matrix is ​​calculated based on the fused coarse-scale features, and a coarse-scale matching set is obtained from it. ;

[0016] S3, Attention-based fine-scale feature matching: For each coarse match in the coarse-scale matching set, a local window is cropped on the distilled fine-scale feature map, and feature enhancement is performed through a lightweight feature transformation module to obtain locally enhanced features; based on the locally enhanced features, correlation calculation and expectation solving are performed to obtain the final sub-pixel-level fine-scale matching set. ;

[0017] S4, Joint optimization of phase consistency structural information distillation and matching: Construct coarse-scale matching loss based on steps S2 and S3 respectively. and fine-scale matching loss The phase-consistency structural distillation loss is combined with the coarse-scale matching loss and the fine-scale matching loss to obtain the total loss:

[0018] (14)

[0019] Based on this, a network is trained to obtain a multi-scale feature extraction and matching model for inference.

[0020] Preferably, step S1-1 specifically includes the following steps:

[0021] S1-1-1, Performing an inverse Fourier transform on the Log-Gabor filter yields its even-symmetric filter in the spatial domain. and odd symmetric filters and the image At different scales and direction Convolution operations are performed on each component to obtain the corresponding response components. and :

[0022] (1),

[0023] In the formula, * represents the convolution operation;

[0024] S1-1-2, Calculate the image At different scales and direction amplitude at and phase :

[0025] (2),

[0026] (3);

[0027] S1-1-3, for the same direction Calculate the weighted average phase across all scales. Define the phase deviation function:

[0028] (4),

[0029] Phase consistency of an image is obtained by integrating phase values ​​from multiple directions and scales. The calculation expression is as follows:

[0030] (5),

[0031] In the formula, Point Phase consistency value at the location, For frequency expansion weighting factors, This is the noise compensation value. To prevent constants with a denominator of zero, This represents the positive operator.

[0032] Preferably, steps S1-5 specifically include the following steps:

[0033] S1-5-1, Constructing Structural Consistency Loss :

[0034] (7);

[0035] S1-5-2, Constructing Cross-Modal Consistency Loss :

[0036] (8),

[0037] in, As a scale, it is divided into coarse scale and Fine scale, , For image and In scale Phase consistency prediction plot on the above, , For image and In scale Phase consistency plot on Representing scale The corresponding position mapping function, that is, the image and In scale Points on Mapped to image At the same scale The corresponding point coordinates on, This represents the structural similarity loss.

[0038] Preferably, step S2 specifically includes the following steps:

[0039] S2-1, introduce position encoding into the coarse-scale features obtained in step S-1 to explicitly inject spatial position information;

[0040] S2-1-1, the coarse-scale features obtained in step S-1 , Flattened into a sequence representation based on spatial location , ,in , and These represent the height and width of the feature map, respectively. Number of feature channels, sequence index with two-dimensional coordinates One-to-one correspondence, , ;

[0041] S2-1-2, for each position Constructing a two-dimensional sinusoidal position encoding vector , its first The channels are:

[0042] (9),

[0043] in, , , The above formula forms the position encoding matrix. The positional encoding is added to the coarse-scale feature sequence to obtain and ;

[0044] S2-2, the sequence after position encoding , The input consists of a feature transformation module with multiple layers of self-attention and cross-attention stacked alternately, resulting in features that fuse contextual information and cross-image interaction information. , ,

[0045] (10)

[0046] in Implemented using standard multi-head attention or efficient / linear attention;

[0047] S2-3, for the image At any coarse scale position and images At any coarse scale position Calculate the similarity score matrix :

[0048] (11),

[0049] in, This represents the inner product operation. for Temperature coefficient;

[0050] S2-4, based on similarity score matrix The coarse matching confidence matrix is ​​obtained through the differentiable matching operator. Matches with confidence scores below a threshold are filtered out, and a coarse-scale matching set is obtained by applying a nearest-neighbor consistency constraint. .

[0051] Preferably, step S3 specifically includes the following steps:

[0052] S3-1, perform each coarse match Mapping to fine-scale feature map coordinates and in fine-scale feature maps , The above are respectively , Center-cut size A local window is used to obtain local fine-scale feature blocks. , ;

[0053] S3-2, will , Input is fed into a lightweight feature transformation module consisting of alternating stacks of self-attention and cross-attention. To obtain the locally enhanced features , ;

[0054] S3-3, with The center vector is the query, and... Perform correlation calculations on all locations within the window and The matching probability map is obtained, and the expectation is calculated on the probability map to obtain the graph. pixel position Output the final fine-scale matching set .

[0055] Preferably, step S4 specifically includes the following steps:

[0056] S4-1, based on the truth transformation relationship between images and the coarse matching confidence matrix obtained from S2-4 Construct coarse-scale matching loss :

[0057] (12);

[0058] S4-2, for each query point By calculating the total variance of the corresponding heatmap To measure its uncertainty, a fine-scale matching loss is constructed. :

[0059] (13)

[0060] in, For query point In the figure The actual corresponding position on the screen. For the final fine-scale matching set;

[0061] S4-3 combines the phase coherence structural distillation loss and matching loss from S1-5 to obtain the total loss:

[0062] (14).

[0063] Preferably, the position encoding in step S2 is sinusoidal position encoding, learnable position encoding, or relative position encoding.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] 1. This invention takes phase-consistent structural information distillation as its core to improve cross-modal structural representation capabilities. It utilizes phase consistency to extract structural information that is insensitive to modal differences and uses it as a structural prior to distill network features at both coarse and fine scales. This results in the learned multi-scale features having stronger structural consistency and discriminability, providing a stable and reliable feature foundation for subsequent matching and effectively improving the generalization ability of cross-modal matching.

[0066] 2. This invention possesses both the robust generalization capability of traditional structural priors and the high expressiveness of end-to-end optimization in deep learning. By introducing phase-consistent structural priors, feature learning maintains robustness and generalization against modal radiation differences, illumination variations, and noise interference under cross-sensor and cross-scene conditions. At the same time, it combines attention-based cross-image feature interaction and joint optimization strategies to enhance global context modeling and feature discrimination capabilities, thereby improving matching accuracy and reducing dependence on manual feature design and complex parameter adjustment.

[0067] 3. This invention introduces phase-consistent structural distillation constraints on both coarse-scale and fine-scale features, enabling cross-modal feature responses to form more consistent representations in stable regions such as structural edges, corners, and linear textures. This results in better peak concentration and class separability in the matching similarity distribution, effectively suppressing non-corresponding responses caused by modal radiation differences, noise, and repetitive textures. Based on this, the matching set output by this invention exhibits better interior point discriminability and consistency when entering geometric model estimation and registration. It can obtain stable and reliable geometric constraints with less reliance on complex heuristic rules and fine-tuning of parameters, improving robustness and reproducibility in engineering deployment. Attached Figure Description

[0068] Figure 1 This is a flowchart of the invention;

[0069] Figure 2 This is a schematic diagram of the phase consistency structure information distillation of the present invention. Detailed Implementation

[0070] The following will combine Figure 1-2 The present invention will be described in detail below. The illustrative embodiments and descriptions herein are used to explain the invention, but are not intended to limit the invention.

[0071] This invention proposes a multimodal image matching method based on phase consistency structural information distillation. For the multimodal image pairs to be matched, this method first obtains stable structural prior information through phase consistency calculation, which, together with multi-scale features extracted by a convolutional neural network, constitutes a structure-guided feature representation. Subsequently, the network is trained using structural consistency loss and cross-modal consistency loss to perform structural distillation, resulting in stronger structural consistency between the coarse-scale and fine-scale features under cross-modal conditions. Based on this, a feature transformation module consisting of alternating stacked self-attention and cross-attention is used to complete coarse-scale global matching to generate candidate matching pairs. Furthermore, local windows are cropped on the fine-scale feature map centered on the coarse matching, and fine-scale sub-pixel-level matching is achieved through lightweight attention enhancement and correlation probability distribution expectation refinement. Finally, a high-confidence matching set is output.

[0072] Multimodal image matching methods based on phase consistency structural information distillation, such as Figure 1 As shown, it includes the following steps:

[0073] S1, Multi-scale feature encoding and phase-consistent structural information distillation;

[0074] S1-1, the multimodal image to be matched and Calculate the phase consistency feature map separately , As a structural guidance graph, it is used to constrain the network to learn cross-modal stable structural representations.

[0075] S1-1-1, Performing an inverse Fourier transform on the Log-Gabor filter yields its even-symmetric filter in the spatial domain. and odd symmetric filters and the image At different scales and direction Convolution operations are performed on each component to obtain the corresponding response components. and :

[0076] (1),

[0077] In the formula, * represents the convolution operation.

[0078] S1-1-2, Calculate the image At different scales and direction amplitude at and phase :

[0079] (2),

[0080] (3);

[0081] S1-1-3, for the same direction Calculate the weighted average phase across all scales. Define the phase deviation function:

[0082] (4),

[0083] Phase consistency of an image is obtained by integrating phase values ​​from multiple directions and scales. The calculation expression is as follows:

[0084] (5),

[0085] In the formula, Point Phase consistency value at the location, For frequency expansion weighting factors, This is the noise compensation value. To prevent constants with a denominator of zero, This represents the positive operator.

[0086] S1-2, , Align to the scale set using bilinear interpolation, average pooling, or other equivalent downsampling methods. A coarse-scale phase consistency map was obtained. , and fine-scale phase consistency map , This serves as information for teacher supervision.

[0087] S1-3, multimodal images and Inputting the convolutional neural network extracts multi-scale features, resulting in coarse-scale features from the two images. , and fine-scale features , .

[0088] S1-4, construct a phase consistency prediction head at each scale, map the features to a phase consistency prediction map, and obtain the coarse-scale phase consistency prediction map. , and fine-scale phase consistency prediction map , .

[0089] S1-5, Construct the loss function and perform structural distillation training to build a phase-consistent structural distillation loss:

[0090] (6),

[0091] in, , These are the weighting coefficients.

[0092] S1-5-1, Constructing Structural Consistency Loss :

[0093] (7);

[0094] S1-5-2, Constructing Cross-Modal Consistency Loss :

[0095] (8),

[0096] in, As a scale, it is divided into coarse scale and Fine scale, , For image and In scale Phase consistency prediction plot on the above, , For image and In scale Phase consistency plot on Representing scale The corresponding position mapping function, that is, the image and In scale Points on Mapped to image At the same scale The corresponding point coordinates on, This represents the structural similarity loss.

[0097] S1-6, based on the loss function The feature extraction network and prediction head are jointly trained to obtain distilled coarse-scale features that include structural prior information. , and fine-scale features , The aforementioned phase consistency prediction head and loss function are only used in the training phase. In the inference phase, the prediction head and phase consistency calculation module are removed, and the distilled multi-scale features are directly used in the subsequent coarse-scale matching and fine-scale matching processes, thereby reducing inference overhead.

[0098] S2, coarse-scale feature matching based on attention mechanism;

[0099] S2-1, introduce position encoding into the coarse-scale features obtained in step S-1 to explicitly inject spatial position information;

[0100] S2-1-1, the coarse-scale features obtained in step S-1 , Flattened into a sequence representation based on spatial location , ,in , and These represent the height and width of the feature map, respectively. Number of feature channels, sequence index with two-dimensional coordinates One-to-one correspondence, , .

[0101] S2-1-2, to enhance the spatial discriminability of features, for each location... Constructing a two-dimensional sinusoidal position encoding vector , its first The channels are:

[0102] (9),

[0103] in, , , The above equation forms the position encoding matrix. The positional encoding is added to the coarse-scale feature sequence to obtain and (Alternatively, learnable positional coding or relative positional coding can be used to replace the two-dimensional sinusoidal positional coding described above.)

[0104] S2-2, the sequence after position encoding , The input consists of a feature transformation module with multiple layers of self-attention and cross-attention stacked alternately, resulting in features that fuse contextual information and cross-image interaction information. , .

[0105] (10)

[0106] in Standard multi-head attention or efficient / linear attention can be used to reduce computational complexity.

[0107] S2-3, for the image At any coarse scale position and images At any coarse scale position Calculate the similarity score matrix :

[0108] (11),

[0109] in, This represents the inner product operation. for Temperature coefficient.

[0110] S2-4, based on similarity score matrix The coarse matching confidence matrix is ​​obtained through the differentiable matching operator. Matches with confidence scores below a threshold are filtered out, and a coarse-scale matching set is obtained by applying a nearest-neighbor consistency constraint. .

[0111] S3, fine-scale feature matching based on attention mechanism;

[0112] S3-1, for each pair of coarse matches Map it to fine-scale feature map coordinates and in fine-scale feature maps , The above are respectively , Center-cut size A local window is used to obtain local fine-scale feature blocks. , .

[0113] S3-2, will , Input is fed into a lightweight feature transformation module consisting of alternating stacks of self-attention and cross-attention. To obtain the locally enhanced features , .

[0114] S3-3, with The center vector is the query, and... Perform correlation calculations on all locations within the window and The matching probability map is obtained, and the expectation is calculated on the probability map to obtain the graph. pixel position Output the final fine-scale matching set .

[0115] S4, joint optimization of phase-consistent structural information distillation and matching;

[0116] S4-1, based on the truth transformation relationship between images and the coarse matching confidence matrix obtained from S2-4 Construct coarse-scale matching loss As a preferred option, cross-entropy loss can be used:

[0117] (12);

[0118] S4-2, for each query point By calculating the total variance of the corresponding heatmap To measure its uncertainty, a fine-scale matching loss is constructed. As a preferred option, L2 loss can be used:

[0119] (13)

[0120] in, For query point In the figure The actual corresponding position on the screen. This is the final fine-scale matching set.

[0121] S4-3 combines the phase coherence structural distillation loss from S1-5 with the matching loss from S4 to obtain the total loss:

[0122] (14);

[0123] Based on this, a multi-scale feature extraction and matching model for inference is obtained by training the network.

[0124] This invention achieves high-precision and robust matching of multimodal images (such as visible light-infrared, visible light-SAR, and visible light-depth). On one hand, by introducing phase-consistent structural prior distillation during the feature extraction stage, this invention uses stable structural information extracted by the Log-Gabor filter bank as a guiding constraint, significantly enhancing the network's adaptability to cross-modal radiation differences and illumination variations, thus obtaining reliable cross-modal structural consistency representations. On the other hand, this invention employs a coarse-to-fine two-stage matching framework based on an attention mechanism: in the coarse-scale stage, self-attention and cross-attention are used to achieve global context aggregation and cross-image information interaction, obtaining stable global candidate matches; in the fine-scale stage, local window refinement and correlation matching are performed centered on the coarse match, outputting sub-pixel-level positions, thereby further improving matching accuracy while ensuring robustness. This invention introduces interpretable structural priors and consistency constraints into the overall framework, preserving the robustness and generalization of traditional methods under cross-modal conditions; simultaneously, by utilizing attention-driven end-to-end feature learning and a coarse-to-fine matching strategy, it leverages the advantages of deep learning methods in feature representation capabilities, matching accuracy, and joint optimization.

[0125] The technical solutions provided by the embodiments of the present invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the embodiments of the present invention. The descriptions of the embodiments above are only for helping to understand the principles of the embodiments of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the embodiments of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A multi-modal image matching method based on phase consistency structure information distillation, characterized in that, The method comprises the following steps: S1, multi-scale feature coding and phase consistency structure information distillation: S1-1, for the matched multi-modal images and respectively compute phase consistency feature maps , as structure-guided maps; S1-2, to , Align to the scale set by bilinear interpolation, average pooling or other equivalent downsampling methods , get coarse scale phase consistency map , and fine scale phase consistency map , , as teacher supervision information; S1-3, extracting multi-modal images and inputting a convolutional neural network to extract multi-scale features to obtain coarse scale features of the two images 、 and fine scale features 、 ; S1-4, constructing the phase consistency prediction head on each scale, mapping the features into the phase consistency prediction map, obtaining the coarse scale phase consistency prediction map as , and the fine scale phase consistency prediction map , ; S1-5, constructing a loss function and performing structure distillation training, constructing a phase consistency structure distillation loss: （6）， wherein, , is a weight coefficient, is a structural consistency loss, is a cross-modal consistency loss; S1-6, according to the loss function Joint training of the feature extraction network and the prediction head to obtain distilled coarse-scale features containing structural prior information 、 and fine-scale features 、 ; S2, coarse-scale feature matching based on attention mechanism: introducing position encoding to the distilled coarse-scale feature, and performing context information and cross-image interaction through a feature transformation module to obtain a fused coarse-scale feature; calculating a similarity matrix based on the fused coarse-scale feature, and obtaining a coarse-scale matching set from the similarity matrix ; S3, fine-scale feature matching based on an attention mechanism: for each pair of coarse matches in the coarse-scale matching set, a local window is cropped on the distilled fine-scale feature map, and feature enhancement is performed through a lightweight feature transformation module to obtain locally enhanced features; correlation calculation and expectation solving are performed based on the locally enhanced features to obtain a final fine-scale matching set at a sub-pixel level ; S4, phase consistency structure information distillation and joint optimization of matching: based on steps S2 and S3, respectively constructing coarse scale matching loss and fine scale matching loss , jointing the phase consistency structure distillation loss with the coarse scale matching loss, fine scale matching loss to obtain total loss: （14）， And train the network accordingly to obtain a multi-scale feature extraction and matching model for inference.

2. The multi-modal image matching method based on phase consistency structural information distillation according to claim 1, characterized in that, The step S1-1 specifically comprises the following steps: S1-1-1, inverse Fourier transform of Log-Gabor to get its even-symmetric filter in spatial domain and odd-symmetric filter , and convolve the image at different scales and orientations to get the corresponding response components and : （1）， In the formula, * represents convolution operation; S1-1-2, compute image at different scales and directions with amplitudes and phases : （2）， （3）； S1-1-3, to the same direction Phase-computing weighted mean phase of all scales , define a phase deviation function: （4）， By integrating the phase values of the image in multiple directions and multiple scales, the phase consistency of the image is obtained, and the calculation expression is as follows: （5）， wherein phase coherence value at a point phase coherence value at a point is a frequency spreading weight factor, is a noise compensation value, is a constant to prevent the denominator from being zero, denotes the positive operator.

3. The multi-modal image matching method based on phase consistency structural information distillation according to claim 1, characterized in that, The step S1-5 specifically comprises the following steps: S1-5-1, construct structure consistency loss : （7）； S1-5-2, constructing cross-modal consistency loss : （8）， wherein, is a scale, divided into coarse scales and fine scales, , is an image and a phase consistency prediction map at scale , , is an image and a phase consistency map at scale , denotes a corresponding position mapping function at scale , i.e. mapping a point and at scale to a corresponding point coordinate of an image at the same scale , denotes a structural similarity loss.

4. The multi-modal image matching method based on phase consistency structural information distillation according to claim 1, characterized in that, The step S2 specifically comprises the following steps: S2-1, introducing position coding into the coarse scale feature obtained in step S-1 to explicitly inject spatial position information; S2-1-1, flattening the coarse-scale features obtained in step S-1 into a sequence list according to spatial positions 、 , 、 , , , , , , , , ; S2-1-2, for each position constructing a two-dimensional sinusoidal position encoding vector its first channel is （9）， wherein , , , the position encoding matrix is formed from the above formula , the position encoding is added to the coarse scale feature sequence to obtain and ; S2-2, the position-encoding sequence , The input is stacked into a feature transformation module alternately stacked by multi-layer self-attention and cross-attention to obtain the feature fused with context information and cross-image interaction information , , （10）， wherein implemented with standard multi-head attention or efficient / linear attention; S2-3, to the image any coarse scale position in the image and the image any coarse scale position in the image , computing a similarity score matrix : （11）， wherein denotes an inner product operation, is temperature coefficient of S2-4, based on the similarity score matrix , obtaining a coarse matching confidence matrix by a differentiable matching operator , filtering out matches with confidence below a threshold and obtaining a coarse scale matching set by mutual nearest neighbor consistency constraints .

5. The multi-modal image matching method based on phase consistency structural information distillation according to claim 1, characterized in that, The step S3 specifically comprises the following steps: S3-1, coarse matching mapping to fine-scale feature map coordinates , and respectively cutting a local window of size centered at , on the fine-scale feature map , to obtain local fine-scale feature blocks , ; S3-2, to , input to a lightweight feature transformation module stacked by self-attention and cross-attention alternately , to obtain locally enhanced features , ; S3-3, to the center vector as query, with all positions in the window do correlation computation and get the matching probability map, do expectation computation on the probability map, get the map pixel position on the map output the final fine scale matching set .

6. The multi-modal image matching method based on phase consistency structural information distillation according to claim 1, characterized in that, The step S4 specifically comprises the following steps: S4-1, the true value transformation relationship between images and the coarse matching confidence matrix obtained by S2-4 , construct a coarse scale matching loss : （12）； S4-2, for each query point , measure its uncertainty by computing the total variance of the corresponding heat map , construct a fine-scale matching loss : （13）， wherein, is the query point the true corresponding position on the image , and is the final fine-scale matching set; S4-3, combining the phase consistency structure distillation loss in S1-5 with the matching loss to obtain a total loss: （14）。 7. The multi-modal image matching method based on phase consistency structural information distillation according to claim 1, characterized in that: The position coding in the step S2 is sine position coding, learnable position coding or relative position coding.

Images