Multi-modal remote sensing image matching method

Abstract

The invention discloses a multi-modal remote sensing image matching method, particularly relates to the technical field of remote sensing image processing, and is used for solving the problem of multi-modal image matching. The method mainly comprises the following steps of: 1, improving a phase consistency model, and constructing a phase-moment weighted joint direction feature in combination with a maximum moment and a minimum moment to replace the feature expression of the traditional image gradient; 2, implementing a point product fusion strategy on the phase-amplitude characteristics extracted by the phase consistency model and the maximum moment, and constructing phase-moment weighted joint amplitude characteristics; and 3, on the basis of the steps 1 and 2, identifying the direction of the feature points and screening local peak values to determine the main direction. Three values adjacent to a peak value are selected, and the peak value position is interpolated through parabola fitting so as to improve the matching precision; and step 4, constructing a logarithm polar coordinate descriptor based on regularization non-uniform partition to generate a feature description vector. Through the mode, high-precision and high-efficiency matching of the multi-mode remote sensing image can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a multimodal remote sensing image matching method. Background Art

[0002] In the remote sensing and photogrammetry disciplines, image matching, as a core technology for visual information processing, provides an important theoretical foundation and technical support for achieving key tasks such as image fusion, ground feature change monitoring, spatial positioning, and navigation. Given the complexity and diversity of remote sensing application scenarios, single-modality data suffers from limitations such as insufficient information representation and incomplete semantic understanding. There is an urgent need to collaboratively process multimodal data and establish a data fusion mechanism with complementary advantages to improve the accuracy and reliability of remote sensing information extraction. However, traditional image matching methods face challenges in multimodal data processing, such as large feature differences and low matching accuracy. There is an urgent need to explore new matching strategies and algorithms to adapt to the characteristics and fusion requirements of multi-source heterogeneous data.

[0003] With the rapid development of sensor technology, imaging devices such as visible light cameras, infrared cameras, synthetic aperture radars, and lasers have emerged, providing a rich and diverse data source for Earth observation. Effectively integrating multi-sensor, multi-resolution, and multi-temporal data and conducting in-depth analysis has become a research hotspot, and multimodal remote sensing image matching is one of the core issues that needs to be addressed.

[0004] Remote sensing images exhibit severe nonlinear radiometric differences and complex geometric discrepancies. These discrepancies make image matching a challenging task. In recent years, many remote sensing image matching methods have been proposed, which can be broadly categorized into two types: region-based methods and feature-based methods.

[0005] Region-based methods, such as normalized cross-correlation and mutual information, achieve matching through template similarity measurement. Although they can circumvent the complexity of feature extraction and obtain sub-pixel accuracy, their sensitivity to geometric deformation and computational complexity seriously restrict their practical applications.

[0006] Feature-based methods typically involve three main steps: feature detection, feature description, and feature matching. These methods typically begin by extracting salient image features such as points, lines, regions, and edges. Subsequently, image correspondences are established by calculating the similarity of these feature descriptors. Early classic algorithms, such as SURF, KAZE, and PSO-SIFT, suffer from low matching accuracy and insufficient robustness when processing multimodal remote sensing images (MRSI) with intensity variations and nonlinear radiometric distortion (NRD). To overcome these bottlenecks, researchers have turned to frequency domain image feature mining and proposed the Log-Gabor filter-based LGHD algorithm, which utilizes phase features to optimize the matching process. The Histogram of Absolute Phase Consistency Gradients (HAPCG) algorithm further enhances the expressive power of phase features, significantly improving the matching accuracy and reliability of MRSI images. However, its applicability is limited by rotational and scale variations, as well as computational complexity. Therefore, the design and development of a robust multimodal remote sensing image matching method is highly desirable. Summary of the Invention

[0007] In view of the problems existing in the background technology, the present invention proposes a multimodal weak texture remote sensing image matching method comprising the following steps:

[0008] Step 1: Improve the phase consistency model, introduce a weighted bandwidth function to generate the weight coefficient of the phase directional feature, and combine the maximum moment and minimum moment to construct a phase-moment weighted joint directional feature model that is resistant to radiation differences, replacing the traditional feature expression based on image gradient direction;

[0009] Step 2: By applying a dot product fusion strategy to the phase-amplitude features extracted from the phase congruency model and the maximum moment, a phase-moment weighted joint amplitude feature model was successfully constructed. This model excels in dealing with complex conditions such as illumination fluctuations, noise contamination, and geometric distortion, and serves as an ideal alternative to traditional gradient amplitude features, providing a more robust feature foundation for subsequent image processing and analysis.

[0010] Step 3: Within the circular neighborhood of the feature point's characteristic scale space, the feature point's direction is first identified based on the phase-moment weighted joint direction and amplitude features. Then, 36 histogram components are generated to statistically analyze the gradient direction distribution, and the local peak (no less than 80% of the global maximum peak) is screened to determine the main direction. Finally, the three adjacent values ​​of the peak are selected and the peak position is interpolated by parabola fitting to improve the matching accuracy.

[0011] Step 4: Combine the multi-level quantization mechanism of distance and angle to perform a non-uniform partitioning strategy on the regularized log-polar coordinates.

[0012] Furthermore, in step 1, the phase consistency model is improved, and a weighted bandwidth function is introduced to generate the weight coefficient of the phase directional feature. The specific calculation formula is as follows:

[0013]

[0014] Where wc′ and wc″ represent the maximum and minimum weight coefficients, exp is an exponential operator, Cutoff is the fractional measure of frequency expansion, and width o (x, y) is the fraction of the image frequency, g controls the sharpness of the PC model transition, and ξ is the minimum value that excludes zero. The phase energy component in each direction and scale is weighted and then added to the odd and even symmetric functions of the 2D-Log-Gabor function respectively. The calculation formula is as follows:

[0015]

[0016] Where WO represents the weighted phase direction feature, OO o (x, y) represents the odd symmetric convolution result after normalization of the oth layer direction, EO o (x, y) represents the even symmetric convolution result after normalization of the direction of the oth layer, represents the rotation angle, and φ is the minimum value excluding zero (φ=0.0001).

[0017] In order to effectively overcome the negative impact of extreme phase mutations, the weighted maximum moment, minimum moment and weighted phase direction features are fused. The specific formula is as follows:

[0018]

[0019] In the formula, WO m represents the phase-torque weighted joint directional feature, Indicates the dot multiplication operation, M max and M min Represent the maximum torque and minimum torque respectively, and the calculation formula is as follows:

[0020]

[0021] The three distance components are:

[0022]

[0023] Where θ o Indicates the angle in the o direction.

[0024] Finally, in order to eliminate the angle reversal caused by multiple directional convolution values, the azimuth angle is converted to between 0° and 360° and W is used to represent the final phase-torque weighted joint directional feature. The specific formula is as follows:

[0025]

[0026] Where W represents the final directional feature, WP represents the radian value of the directional feature, and ω represents a non-negative constant term (ω = 360). denotes a convolution operation, and Φ is the minimum value excluding zero (Φ=0.0001).

[0027] In step 2, the phase-amplitude feature is fused with the maximum moment. The specific formula is as follows:

[0028]

[0029] Where, F Fused is the amplitude feature after fusion, W o (x) is a one-dimensional weight function, A no (x, y) is the phase component at scale n and direction o. The constant ε (ε>0) can prevent the denominator from becoming zero. T o is the noise compensation amount, φ no-avg is the average phase angle.

[0030] In step 3, within the circular neighborhood of the feature point feature scale space, the feature point direction is first identified based on the phase-moment weighted joint direction and amplitude features; then, 36 histogram components are generated to statistically calculate the gradient direction distribution, and the local peak value (not less than 80% of the global maximum peak value) is screened to determine the main direction; finally, three adjacent values ​​are selected for the peak value, and the peak position is interpolated by parabola fitting to improve the matching accuracy.

[0031] In step 4, the logarithmic polar coordinate local area is divided into concentric rings of different radii, and the gradient information is dynamically allocated according to the distance and angle. The specific partitioning method is: the inner area is divided into 15 sector areas, the middle inner area is divided into 13 sector areas, the middle outer area is divided into 11 sector areas, and the outermost area is divided into 9 sector areas. The pixels in each sub-area have 8 directional dimensions. Therefore, the 48 sub-areas of the logarithmic polar coordinates and the 8 directional dimensions of each feature point are multiplied to generate a new 384-dimensional logarithmic polar coordinate descriptor vector. The formula is as follows:

[0032] UNRGLOH=[D1,D2,…,D N ] T ⑧

[0033] Where UNRGLOH represents the set of all feature point descriptors, Di T Describes the descriptor of the feature point, T represents the matrix transpose, and N represents the number of feature points. Its 384-dimensional descriptor component is D i T =[V1, V2, ..., V 384 ].

[0034] Compared with the prior art, the present invention has the following benefits:

[0035] This paper improves on the features in the phase congruency model, replacing traditional image gradient features with new ones. This effectively addresses the difficulty in accurately matching remote sensing images due to nonlinear radiometric distortion. Furthermore, the proposed gradient features are used to calculate the principal direction of feature points, and parabolic interpolation is used to further accurately determine this principal direction, enhancing rotational invariance and improving matching accuracy.

[0036] The principle of the invention is intuitive and easy to understand, has good operability, and can be effectively applied to matching scenarios of various remote sensing images.

[0037] The multimodal remote sensing image matching method proposed in the present invention can effectively achieve correct matching of remote sensing images with various types of differences such as rotation differences and scale differences.

[0038] The multimodal remote sensing image matching method proposed in this invention supports flexible adjustment of parameters such as the number of feature points and scale factor, which can be selected according to the actual matching task requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0040] Figure 1 Flowchart of the multimodal remote sensing image matching method of the present invention;

[0041] Figure 2 Comparison results of weighted phase feature and phase-moment weighted joint directional feature, where (a) is the weighted phase directional feature map and (b) is the phase-moment weighted joint directional feature map;

[0042] Figure 3 Comparison results of phase amplitude and phase-moment weighted joint amplitude features, where (a) is the optical remote sensing image, (b) is the phase amplitude feature, and (c) is the phase-moment weighted joint amplitude feature;

[0043] Figure 4 This is a flow chart for calculating the main direction of feature points of the present invention;

[0044] Figure 5 is the non-uniform regularized log-polar coordinate;

[0045] Figure 6 : 6 sets of multimodal remote sensing image pairs in an embodiment of the present invention, wherein (a) is a depth-optical image pair, with the depth image on the top and the optical image on the bottom; (b) is an infrared-optical image pair, with the infrared image on the top and the optical image on the bottom; (c) is a map-optical image pair, with the map on the top and the optical image on the bottom; (d) is a SAR-optical image pair, with the SAR image on the top and the optical image on the bottom; (e) is a night-day image pair, with the optical remote sensing image at night on the top and the optical remote sensing image at daytime on the bottom;

[0046] Figure 7 The index results of several matching algorithms, among which (a) is the result of the number of correct matches of the six methods, (b) is the result of the correct matching rate of the six methods, and (c) is the result of the root mean square error of the six methods;

[0047] Figure 8 Results of six matching algorithms, including (a) LNIFT algorithm, (b) CoFSM algorithm, (c) RIFT algorithm, (d) HAPCG algorithm, (e) HOWP algorithm, and (f) JW-PMDH algorithm.

[0048] Figure 9 These are the results of matching five sets of multimodal remote sensing image pairs using the method proposed in this invention, where (a) is the depth-optical image matching result, (b) is the infrared-optical image matching result, (c) is the map-optical image matching result, (d) is the SAR-optical image matching result, and (e) is the night-day image matching result. DETAILED DESCRIPTION

[0049] The preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more precise definition of the protection scope of the present invention.

[0050] Example 1

[0051] like Figure 1 As shown, an embodiment of the present invention provides a multimodal remote sensing image matching method, comprising the following steps:

[0052] Step 1: Improve the phase consistency model, introduce a weighted bandwidth function to generate the weight coefficient of the phase directional feature, and combine the maximum moment and minimum moment to construct a phase-moment weighted joint directional feature model that is resistant to radiation differences. This replaces the traditional feature expression based on image gradient direction. The model includes the following sub-steps:

[0053] The phase-oriented features generated by the weighted bandwidth function can effectively address the problem of image shape component changes caused by phase direction reversal, phase extreme value mutation, etc. Such problems easily destroy the image structural features and increase the difficulty of descriptor recognition. Its calculation formula is as follows:

[0054]

[0055] Where wc′ and wc″ represent the maximum and minimum weight coefficients, exp is an exponential operator, Cutoff is the fractional measure of frequency expansion, and width o (x, y) is the fraction of the image frequency, g controls the sharpness of the PC model transition, and ξ is the minimum value that excludes zero.

[0056] like Figure 2 As shown in (a), the phase energy components in each direction and scale are weighted, and then the calculation results are added to the odd symmetry function and even symmetry function of 2D-Log-Gabor respectively. The calculation formula is as follows:

[0057]

[0058] Where WO represents the weighted phase direction feature, OO o (x, y) represents the odd symmetric convolution result after normalization of the oth layer direction, EO o (x, y) represents the even symmetric convolution result after normalization of the direction of the oth layer, represents the rotation angle, and Φ is the minimum value excluding zero (Φ=0.0001).

[0059] like Figure 2 As shown in (b), in order to effectively overcome the negative impact of phase extreme value mutation, the weighted maximum moment, minimum moment and weighted phase direction feature are fused. The specific formula is as follows:

[0060]

[0061] In the formula, WO m represents the phase-torque weighted joint directional feature, Indicates the dot multiplication operation, M max and M min Represent the maximum torque and minimum torque respectively, and the calculation formula is as follows:

[0062]

[0063] The three distance components are:

[0064]

[0065] Where θ o Indicates the angle in the o direction.

[0066] Finally, in order to eliminate the angle reversal caused by multiple directional convolution values, the azimuth angle is converted to between 0° and 360° and W is used to represent the final phase-torque weighted joint directional feature. The specific formula is as follows:

[0067]

[0068] Where W represents the final directional feature, WP represents the radian value of the directional feature, and ω represents a non-negative constant term (ω = 360). denotes a convolution operation, and Φ is the minimum value excluding zero (Φ=0.0001).

[0069] Step 2, such as Figure 3 As shown in (b), the phase-amplitude feature is fused with the maximum moment. The specific formula is as follows:

[0070]

[0071] Where, F Fused is the amplitude feature after fusion, W o (x) is a one-dimensional weight function, A no (x, y) is the phase component at scale n and direction o. The constant ε (ε>0) can prevent the denominator from becoming zero. T o is the noise compensation amount, φ no-avg is the average phase angle.

[0072] Step 3, such as Figure 4 As shown in the figure, within the circular neighborhood of the feature point feature scale space, the feature point direction is first identified based on the phase-moment weighted joint direction and amplitude features; then, 36 histogram components are generated to statistically calculate the gradient direction distribution, and the local peak value (not less than 80% of the global maximum peak value) is screened to determine the main direction; finally, the three adjacent values ​​​​are selected for the peak value, and the peak position is interpolated by parabola fitting to improve the matching accuracy.

[0073] Step 4, such as Figure 5As shown in the figure, the local area of ​​the logarithmic polar coordinates is divided into concentric rings of different radii, and the gradient information is dynamically allocated according to the distance and angle. The specific partitioning method is: the inner area is divided into 15 sector areas, the middle inner area is divided into 13 sector areas, the middle outer area is divided into 11 sector areas, and the outermost area is divided into 9 sector areas. The pixels of each sub-area have 8 directional dimensions. Therefore, the 48 sub-areas of the logarithmic polar coordinates and the 8 directional dimensions of each feature point are multiplied to generate a new 384-dimensional logarithmic polar coordinate descriptor vector. The formula is as follows:

[0074] UNRGLOH=[D1,D2,…,D N ] T ⑧

[0075] Where UNRGLOH represents the set of all feature point descriptors, D i T Represents the descriptor of the feature point, T represents the matrix transpose, N represents the number of feature points, and its 384-dimensional descriptor component is D i T =[V1, V2, ..., V 384 ], the dimension of each feature point can be expressed as 4×d×n, where n represents the number of log-polar coordinate sub-regions and d represents the direction dimension of each feature point.

[0076] Step 5: Euclidean distance similarity measurement and two-way matching strategy are used to ensure that the matching points are one-to-one corresponding and a fast sample consistency algorithm is used to remove outliers generated after two-way matching (to eliminate false matches).

[0077] Five sets of multimodal remote sensing image pairs are used to compare the effects of different matching methods. Figure 6 (a)- Figure 6 (e) As shown, (a) is a depth-optical image pair, with the depth image on top and the optical image on the bottom; (b) is an infrared-optical image pair, with the infrared image on top and the optical image on the bottom; (c) is a map-optical image pair, with the map on top and the optical image on the bottom; (d) is a SAR-optical image pair, with the SAR image on top and the optical image on the bottom; (e) is a night-day image pair, with the optical remote sensing image at night on top and the optical remote sensing image at daytime on the bottom.

[0078] The multimodal remote sensing image matching method (JW-PMDH) proposed in this paper is used to compare the matching results of five groups of multimodal weak texture remote sensing image pairs with five existing matching methods (RIFT, LNIFT, CoFSM, HOWP and HAPCG). The performance of each matching method is quantitatively evaluated using three indicators: correct match rate (SR), number of correct matches (NCM), correct match rate (RCM) and root mean square error (RMSE). The comparison results are shown in Table 1. Figure 7 and Figure 8 shown. Table 1 Comparison of SR, NCM, RCM and RMSE

[0079] The number of correct matches (NCM) indicates the number of correctly matched feature points in the reference image and the image to be matched. The formula for determining the correct matching points is as follows:

[0080] ||H·(x r ,y r )-x s ,y s )||2≤T1 ⑨

[0081] In the formula, (x r ,y r ) and (x s ,y s ) represent the positions of the feature points in the reference image and the registered image respectively, H is the uniformity matrix between the two images, and T1 is the threshold, which generally defaults to 2.

[0082] The correct match rate (RCM) reflects the ratio of correct match points to all match points given by the algorithm (the sum of correct match points and incorrect match points). This indicator can reflect the matching success rate obtained by the algorithm. The calculation formula is as follows:

[0083]

[0084] Where N C Indicates the number of correct matching points, N F Indicates the number of incorrect matching points.

[0085] The root mean square error (RMSE) is calculated using the conversion model of the correct matching estimate and can reflect the accuracy of the matching points. The smaller the RMSE value, the higher the accuracy of the matching model. The calculation formula is as follows:

[0086]

[0087] Where N represents the number of selected points with the same name, (x i ,y i) represents the coordinates of the same-name points in the two images, (x″ i , y″ i ) represents the coordinates of the same-name point in the i-th image to be matched after the matching correspondence transformation.

[0088] Table 1 shows that the RIFT algorithm achieves a correct match rate (RCM) of 5.10%, an average number of correct matches (NCM) of 71.72, and an average root mean square error (RMSE) of 5.96 pixels in image matching. Although the algorithm does not construct a scale space, it exhibits a certain degree of resistance to small scale factor variations. In comparison, the LNIFT algorithm has an average NCM of 53.98, an average RCM of 4.35%, and an average RMSE of 6.37 pixels, indicating limitations in handling scale and rotation differences. In infrared and lidar image matching scenarios, LNIFT's normalization process fails to completely eliminate radiometric differences and is susceptible to noise. Furthermore, its reliance on the HOG descriptor makes it less capable of distinguishing complex textures or low-contrast areas, resulting in unsatisfactory correct match numbers and matching rates in multimodal remote sensing imagery (MRSI) matching. The CoFSM algorithm utilizes a co-occurrence filter (CoF) to construct a co-occurrence scale space, effectively alleviating the radiometric distortion caused by differences in imaging mechanisms in multimodal images and increasing the number of correct matches (NCM) to 85.50. However, due to its reliance on gradient and edge feature extraction, the feature point detection and matching process is limited in areas with simple textures or sparse features, such as cloud cover and low contrast, as seen in the experimental dataset. The resulting RCM is 9.19% and the RMSE is 7.04 pixels. Both the HAPCG and HOWP algorithms are based on the principle of phase consistency, employing phase feature regularization and phase weighting strategies, respectively, to optimize features and construct a multi-scale space, significantly improving matching stability. The proposed JE-PMDH algorithm achieves efficient matching in all tested image pairs, achieving an average NCM of 162.66, making it the only method among the compared algorithms to exceed 150. Furthermore, its RCM is 19.18% and RMSE is 2.02 pixels, significantly outperforming other algorithms in both correct match rate and error control, fully demonstrating its superior performance in image matching tasks.

[0089] Example 2

[0090] Based on the same inventive concept, the present invention also provides a multimodal remote sensing image matching system, including a processor and a memory, the memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute a multimodal remote sensing image matching method as described above.

[0091] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention's description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A multimodal remote sensing image matching method, characterized in that: The steps include: Step 1: Improve the phase consistency model, introduce a weighted bandwidth function to generate the weight coefficient of the phase directional feature, and combine the maximum moment and minimum moment to construct a phase-moment weighted joint directional feature model that is resistant to radiation differences, replacing the traditional feature expression based on image gradient direction; Step 2: By implementing a dot product fusion strategy on the phase-amplitude features extracted from the phase consistency model and the maximum moment, a phase-moment weighted joint amplitude feature model is successfully constructed; Step 3: Within the circular neighborhood of the feature point's characteristic scale space, the feature point's direction is first identified based on the phase-moment weighted joint direction and amplitude features. Then, a histogram component statistical gradient direction distribution is generated, and the local peak is screened to determine the main direction. Finally, three adjacent values ​​are selected for the peak, and the peak position is interpolated through parabola fitting to improve matching accuracy. Step 4: Combine the multi-level quantization mechanism of distance and angle to perform a non-uniform partitioning strategy on the regularized log-polar coordinates.

2. A multimodal remote sensing image matching method according to claim 1, characterized in that: Step 1 specifically includes: Step 1.1, calculate the weighted function model, the weighted function model is: Where wc′ and wc″ represent the maximum and minimum weight coefficients, exp is an exponential operator, Cutoff is the fractional measure of frequency expansion, and width o (x, y) is the fraction of the image frequency, g controls the sharpness of the PC model transition, and ξ is the minimum value that excludes zero; In step 1.2, weight the phase energy components in each direction and scale, and then add the calculated results to the odd-symmetry function and even-symmetry function of the 2D-Log-Gabor, respectively. The calculation formula is as follows: Where WO represents the weighted phase direction feature, OO o (x, y) represents the odd symmetric convolution result after normalization of the oth layer direction, EO o (x, y) represents the even symmetric convolution result after normalization of the direction of the oth layer, represents the rotation angle, Φ is the minimum value excluding zero; Step 1.3: Add the maximum moment and minimum moment to step 1.

2. The specific formula is: In the formula, WO m represents the phase-torque weighted joint directional feature, ° represents the dot product operation, M max and M min Represent the maximum torque and minimum torque respectively, and the calculation formula is as follows: The three distance components are: Where θ o Indicates the angle in the o direction; Finally, in order to eliminate the angle reversal caused by multiple directional convolution values, the azimuth angle is converted to between 0° and 360° and W is used to represent the final phase-torque weighted joint directional feature. The specific formula is as follows: Where W represents the final directional feature, WP represents the radian value of the directional feature, and ω represents a non-negative constant term (ω = 360). represents the convolution operation, and Φ is the minimum value excluding zero.

3. The multimodal remote sensing image matching method according to claim 1, wherein: Step 2: The formula for fusing the maximum moment with the phase-amplitude feature is as follows: DF no (x, y)=cos(φ no -f no-avg )-|cos(φ no -f no-avg )|. Where, F Fused is the amplitude feature after fusion, W o (x) is a one-dimensional weight function, A no (x, y) is the phase component at scale n and direction o. The constant ε (ε>0) can prevent the denominator from becoming zero. T o is the noise compensation amount, φ no-avg is the average phase angle.

4. The multimodal remote sensing image matching method according to claim 1, wherein: Step 3 specifically includes: Step 3.1, within the circular neighborhood determined by the feature point feature scale space, the direction of each feature point is identified based on the phase-moment weighted joint direction feature and the phase-moment weighted joint amplitude feature; In step 3.2, several histogram components covering the full angle range are generated. Each component corresponds to a certain angle interval. This is used to calculate the distribution of gradient directions in the neighborhood of the feature point. The gradient directions of each pixel in the neighborhood of the feature point are then mapped to the corresponding histogram components to complete the direction information statistics. Finally, the local peaks in the histogram are screened, and the angle range corresponding to the peak that meets the requirements is determined as the main direction of the feature point. In step 3.3, for each peak, the three histogram values ​​closest to it are selected and the peak position is interpolated by parabola fitting to improve rotation invariance and matching accuracy.

5. The multimodal remote sensing image matching method according to claim 1, wherein: In step 4, the local area of ​​the logarithmic polar coordinates is divided into concentric rings of different radii and the gradient information is dynamically allocated according to the distance and angle. The specific partitioning is to divide the inner area into a certain number of sector areas, the middle inner area into another number of sector areas, the middle outer area into a different number of sector areas, and the outermost area into a specific number of sector areas. The pixels of each sub-area have an 8-dimensional gradient direction histogram. The number of logarithmic polar coordinate sub-area grids is multiplied by the dimension of the feature direction to generate a new logarithmic polar coordinate descriptor vector. The formula is as follows: UNRGLOH=[D1,D2,…,D N ] T ⑧ Where UNRGLON represents the sum of all feature point descriptors, D i T Represents the descriptor of the feature point, T represents the matrix transpose, and N represents the number of feature points.

Images