Remote sensing image waveband matching method based on mixed feature points

Abstract

The embodiment of the invention discloses a remote sensing image band matching method based on mixed feature points. The method comprises the following steps: acquiring a multi-band remote sensing image and carrying out noise reduction processing on the multi-band remote sensing image; selecting a reference wave band from the multi-wave band remote sensing image subjected to noise reduction processing to obtain the reference wave band and other wave bands except the reference wave band; calculating mixed feature points of the reference wave band, and calculating mixed feature points of other wave bands; matching the mixed feature points of the reference wave band with the mixed feature points of other wave bands to obtain matched feature point pairs; calculating a transformation model between the reference wave band and the other wave bands according to the matching feature point pairs; and performing coordinate transformation and pixel interpolation on the other wavebands by using the transformation model and the interpolation algorithm to realize remote sensing image waveband matching.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing and digital image processing. More specifically, it relates to a method for band matching of remote sensing images based on hybrid feature points. Background Technology

[0002] Satellite remote sensing images are widely used in military, agriculture, environmental monitoring and other fields. However, due to slight differences in position, attitude or time when remote sensing satellite sensors of different bands acquire images, there are geometric distortions and feature misalignments between images of different bands. This affects the quality of the composite image after band superposition and greatly limits the application of remote sensing images.

[0003] Currently, to ensure spatial and spectral consistency of images across different bands and improve data availability and analytical accuracy, band matching is primarily performed using a single feature point matching method. This method involves selecting a reference band using specific standards or simple mathematical calculations, extracting feature points from both the reference band and the band to be matched using a single feature point extraction model, matching these feature points using a feature point matching method, calculating a transformation model from the matched feature point pairs, applying this transformation model to the image of the band to be matched, performing coordinate transformation and pixel interpolation, and obtaining a multi-band registered image.

[0004] However, this method has certain limitations when dealing with image matching in large-scale and complex scenes. Factors such as spectral shifts, geometric transformations, and atmospheric effects caused by different spectral bands make this method perform poorly in some practical applications. These problems not only affect the accuracy of remote sensing satellite images but also restrict their widespread application in monitoring and analysis. Summary of the Invention

[0005] The purpose of this invention is to provide a remote sensing image band matching method based on hybrid feature points to solve at least one of the problems existing in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of this invention provides a remote sensing image band matching method based on hybrid feature points, the method comprising: Acquire multi-band remote sensing images and perform noise reduction processing on the multi-band remote sensing images; A reference band is selected from the denoised multi-band remote sensing image to obtain the reference band and other bands besides the reference band; Calculate the mixed feature points of the reference band, and calculate the mixed feature points of the other bands; The mixed feature points of the reference band are matched with the mixed feature points of other bands to obtain matched feature point pairs; Calculate the transformation model between the reference band and the other bands based on the matching feature point pairs; The transformation model and interpolation algorithm are used to perform coordinate transformation and pixel interpolation on the other bands to achieve remote sensing image band matching.

[0007] Optionally, selecting a reference band from the denoised multi-band remote sensing image includes: The denoised multi-band remote sensing image is divided into multiple bands; Calculate the wavelet coefficients for each band to obtain multiple wavelet coefficients; Multiple wavelet correlation coefficients are obtained by calculating the wavelet correlation coefficients between the wavelet coefficients of each band and the wavelet coefficients of the remaining bands. Calculate the sum of wavelet correlation coefficients for each band based on multiple wavelet correlation coefficients; The highest wavelet correlation coefficient and its corresponding band are selected as the reference band.

[0008] Optionally, calculating the mixed feature points of the reference band includes: Calculate the first feature point of the reference band and calculate the second feature point of the reference band.

[0009] Optionally, calculating the first feature point of the reference band includes: The reference band is filtered to obtain the scale space; Gradient calculation is performed on the reference band at each scale of the scale space, and potential key points are detected based on the extreme points of the gradient calculation. The location of potential key points is detected based on the potential key points using the difference in Gaussian image pyramid algorithm; Based on the location of the potential key points, the gradient magnitude and direction are calculated for the neighborhood of each potential key point, and the main direction of the potential key points is determined by using the gradient histogram based on the calculation results. A scale- and rotation-invariant local feature descriptor is constructed in the neighborhood of each potential key point to obtain the first feature point of the reference band.

[0010] Optionally, calculating the second feature point of the reference band includes: Set a threshold and detect key points in the reference band, and calculate the response value of each key point; Non-maximum suppression is applied to keypoints, and keypoints corresponding to the maximum response values ​​are retained. A binary descriptor is calculated around the retained key points to obtain the second feature points of the reference band.

[0011] Optionally, calculating the mixed feature points of the other bands includes: Calculate the third feature point of the other bands and calculate the fourth feature point of the other bands.

[0012] Optionally, calculating the third feature point of the other bands includes: Filtering the other bands yields the scale space; Gradient calculations are performed on the other bands at each scale of the scale space, and potential key points are detected based on the extreme points of the gradient calculations. The location of potential key points is detected based on the potential key points using the difference in Gaussian image pyramid algorithm; Based on the location of the potential key points, the gradient magnitude and direction are calculated for the neighborhood of each potential key point, and the main direction of the potential key points is determined by using the gradient histogram based on the calculation results. A scale- and rotation-invariant local feature descriptor is constructed in the neighborhood of each potential key point to obtain the third feature point of the other bands.

[0013] Optionally, calculating the fourth feature point of the other bands includes: Set a threshold and detect key points in the other bands, and calculate the response value for each key point; Non-maximum suppression is applied to keypoints, and keypoints corresponding to the maximum response values ​​are retained. Binary descriptors are calculated around the retained key points to obtain the fourth feature points of the other bands.

[0014] Optionally, matching the mixed feature points of the reference band with the mixed feature points of other bands to obtain matched feature point pairs includes: The first feature point pair is obtained by matching the first feature point of the reference band with the third feature point of the other bands; The second feature point is matched with the fourth feature point of the other bands to obtain the second feature point pair; The first feature point pair and the second feature point pair are merged to obtain a matching feature point pair.

[0015] Optionally, calculating the transformation model between the reference band and the other bands based on the matching feature point pairs includes: The transformation model between the reference band and the other bands is calculated based on the matching feature point pairs using a random sampling consensus algorithm or the least squares method.

[0016] The beneficial effects of this invention are as follows: The technical solution described in this invention achieves high-precision band matching of remote sensing satellite imagery, ensuring geometric consistency and improving matching speed, making it suitable for processing large-scale satellite image data. It utilizes the correlation between band information in multi-band remote sensing imagery to more accurately select the reference band, reducing the randomness of band matching. Simultaneously, it combines two feature operators to compensate for their respective shortcomings, improving the quantity and quality of feature points and achieving high-precision matching between remote sensing image bands, ensuring geometric consistency. Through efficient approximate nearest neighbor search, it improves matching speed, thus making it suitable for processing large-scale satellite image data. Attached Figure Description

[0017] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0018] Figure 1 The flowchart illustrates the remote sensing image band matching method based on hybrid feature points disclosed in an embodiment of the present invention.

[0019] Figure 2 The diagram shows the SIFT operator calculation flowchart of the remote sensing image band matching method based on mixed feature points disclosed in an embodiment of the present invention.

[0020] Figure 3 The diagram shows the ORB operator calculation flowchart of the remote sensing image band matching method based on hybrid feature points disclosed in an embodiment of the present invention. Detailed Implementation

[0021] To more clearly illustrate the present invention, the following description, in conjunction with embodiments and accompanying drawings, further explains the invention. Similar components in the drawings are indicated by the same reference numerals. Those skilled in the art should understand that the specific description below is illustrative rather than restrictive and should not be construed as limiting the scope of protection of the present invention.

[0022] Currently, to ensure spatial and spectral consistency of images across different bands and improve data availability and analytical accuracy, band matching is primarily performed using a single feature point matching method. This method involves selecting a reference band using specific standards or simple mathematical calculations, extracting feature points from both the reference band and the band to be matched using a single feature point extraction model, matching these feature points using a feature point matching method, calculating a transformation model from the matched feature point pairs, applying this transformation model to the image of the band to be matched, performing coordinate transformation and pixel interpolation, and obtaining a multi-band registered image.

[0023] However, this method has certain limitations when dealing with image matching in large-scale and complex scenes. Factors such as spectral shifts, geometric transformations, and atmospheric effects caused by different spectral bands make this method perform poorly in some practical applications. These problems not only affect the accuracy of remote sensing satellite images but also restrict their widespread application in monitoring and analysis.

[0024] In view of the above, one embodiment of the present invention provides a remote sensing image band matching method based on mixed feature points. The method includes: acquiring a multi-band remote sensing image and performing noise reduction processing on the multi-band remote sensing image; selecting a reference band from the noise-reduced multi-band remote sensing image to obtain the reference band and other bands besides the reference band; calculating mixed feature points of the reference band and calculating mixed feature points of the other bands; matching the mixed feature points of the reference band with the mixed feature points of the other bands to obtain matching feature point pairs; calculating a transformation model between the reference band and the other bands based on the matching feature point pairs; and performing coordinate transformation and pixel interpolation on the other bands using the transformation model and an interpolation algorithm to achieve remote sensing image band matching.

[0025] In a specific example, the method includes: performing noise reduction processing on the acquired multi-band remote sensing images that need to be matched; selecting a certain band from the denoised multi-band images as a reference band based on the wavelet correlation coefficient; calculating the mixed feature points of the reference band and other bands other than the reference band; performing mixed feature point matching on the reference band and other bands other than the reference band; calculating the transformation model between the other bands other than the reference band and the reference band for the matched feature point pairs; and applying the transformation model to the other bands other than the reference band to perform coordinate transformation and pixel interpolation.

[0026] In a specific example, the method includes: selecting a reference band from the denoised multi-band remote sensing image based on wavelet correlation coefficients; calculating mixed feature points for the reference band and other band images to be matched; matching the feature point pairs; generating a transformation model between bands based on the set of matched feature point pairs; and performing coordinate transformation and pixel interpolation on the band images to be matched.

[0027] This embodiment achieves high-precision band matching of remote sensing satellite imagery, ensuring geometric consistency and improving matching speed, making it suitable for processing large-scale satellite image data. It utilizes the correlation between band information in multi-band remote sensing imagery to more accurately select the reference band and reduce the randomness of band matching. Simultaneously, it combines two feature operators to compensate for their respective shortcomings, improving the quantity and quality of feature points and achieving high-precision matching between remote sensing image bands, ensuring geometric consistency. Through efficient approximate nearest neighbor search, it improves matching speed, thus making it suitable for processing large-scale satellite image data.

[0028] In a specific example, the method includes: step S110, denoising the multi-band image; Furthermore, for the acquired remote sensing multi-band images, a suitable filter is selected for band-by-band denoising. Here, an adaptive median filter is used. The specific method is as follows: the adaptive median filter consists of two processes, A and B. Process A first determines whether Z is satisfied. min <Z med <Z max If the condition is met, proceed to procedure B; otherwise, increase the window size W. xy If the enlarged window size satisfies W xy ≤ W max If the result is positive, repeat process A; otherwise, replace the original grayscale value with the median to filter out noise. Process B determines whether Z is satisfied. min <Z xy <Z max If the condition is met, the grayscale value of the current pixel is retained; otherwise, the median value is used to replace the original grayscale value to filter out noise.

[0029] Furthermore, W xy W represents the size of the rectangular neighborhood of the image. max W xy Maximum allowed size, Z min W xy The minimum gray value in Z med W xy The median grayscale value in Z. max W xy The maximum grayscale value in Z xy This represents the gray value at image coordinates (x, y).

[0030] In this embodiment, the adaptive median filter can handle noise with a higher probability, smoothing out non-impulse noise while attempting to preserve image details.

[0031] In a specific example, noise reduction processing is performed on the acquired multi-band remote sensing images that need to be matched to improve the image signal-to-noise ratio and reduce the band registration error. In one possible implementation, selecting a reference band from the denoised multi-band remote sensing image includes: dividing the denoised multi-band remote sensing image into multiple bands; calculating wavelet coefficients for each band to obtain multiple wavelet coefficients; calculating wavelet correlation coefficients between the wavelet coefficients of each band and the wavelet coefficients of the remaining bands to obtain multiple wavelet correlation coefficients; calculating the sum of wavelet correlation coefficients for each band based on the multiple wavelet correlation coefficients; and selecting the band corresponding to the largest sum of wavelet correlation coefficients as the reference band.

[0032] In a specific example, such as Figure 1 As shown, the method includes: step S120, selecting a reference band based on the maximum wavelet correlation coefficient; Furthermore, the wavelet correlation coefficients between each pair of bands are calculated, and the band with the largest sum of wavelet correlation coefficients with other bands is selected as the reference band.

[0033] Furthermore, a two-dimensional wavelet basis function is used to perform discrete wavelet transform on the two band images, decomposing the images into wavelet coefficients at different scales and directions. For each scale and direction, the correlation coefficient between the wavelet coefficients of the two band images is calculated. This can be done using common correlation coefficient calculation methods; here, the Pearson correlation coefficient is used. The correlation coefficients at different scales and directions are then weighted and averaged to obtain the overall wavelet correlation coefficient. The formula for calculating the wavelet correlation coefficient between the two band images is:

[0034] in, This represents the wavelet correlation coefficient between band image X and band image Y, where M and N represent the number of rows and columns of the image, respectively. and Indicates the position of image X and image Y Pixel value at that location, and This represents the average pixel value of image X and image Y. This represents the value of the wavelet basis function at scale a and direction b. Represents the weights of correlation coefficients at different scales and directions; t is a traversal index used to represent the 1st to Kth different "scale-direction" wavelet coefficient combinations in sequence. t itself does not represent a specific scale or direction; it is an abstract sequence number (index) whose value ranges from 1 to K. Each specific t value (such as t=1, 2, ...) corresponds to a specific "scale-direction" combination.

[0035] In a specific example, a band is selected as the reference band from the obtained multi-band imagery. The specific method is as follows: A suitable wavelet function is selected, and wavelet transforms are applied to the band images sequentially to convert them to the wavelet domain. The wavelet coefficients corresponding to each band image are calculated at different scales and frequencies. The wavelet coefficients of two band images in the wavelet domain are compared using correlation coefficients, such as the Pearson correlation coefficient, to measure their similarity. The wavelet correlation coefficient value is typically between -1 and 1, where 1 indicates a perfect positive correlation (the two signals are identical in the wavelet domain), -1 indicates a perfect negative correlation (the two signals are completely opposite in the wavelet domain), and 0 indicates no linear correlation (the two signals have no linear relationship in the wavelet domain). The sum of the wavelet correlation coefficients of each band with other bands is calculated sequentially, and the band with the largest sum of wavelet correlation coefficients with other bands is selected as the reference band.

[0036] In one possible implementation, calculating the mixed feature points of the reference band includes: calculating a first feature point of the reference band and calculating a second feature point of the reference band.

[0037] In a specific example, the method includes S130, hybrid feature point extraction; Furthermore, hybrid feature point extraction is performed on the reference band and other bands obtained in step S120. Considering matching accuracy and processing efficiency, SIFT (Scale-Invariant Feature Transformation) feature points and ORB (Oriented FAST and Rotated BRIEF) feature points are used in this embodiment.

[0038] Furthermore, SIFT and ORB are used simultaneously for both the reference band and other bands. For example, SIFT is used on the reference band A1 to obtain its SIFT feature point set, and ORB is used to obtain its ORB feature point set; SIFT is used on the non-reference band B1 to obtain its SIFT feature point set, and ORB is used to obtain its ORB feature point set. Then, the SIFT feature point set of A1 is matched with the SIFT feature point set of B1 to obtain feature point pair P1, and the ORB feature point set of A1 is matched with the ORB feature point set of B1 to obtain feature point pair P2. Next, feature point pair P1 and feature point pair P2 are merged to obtain the complete matched feature point pair for bands A1 and B1, which is used for subsequent calculation of the transformation model between A1 and B1.

[0039] In one possible implementation, calculating the first feature point of the reference band includes: filtering the reference band to obtain a scale space; calculating the gradient of the reference band at each scale in the scale space, and detecting potential key points based on the extreme points of the gradient calculation; detecting the location of potential key points based on the potential key points using the difference of Gaussian image pyramid algorithm; calculating the gradient magnitude and direction of the neighborhood of each potential key point based on the location of the potential key points, and determining the main direction of the potential key points based on the calculation results using a gradient histogram; and constructing a scale- and rotation-invariant local feature descriptor in the neighborhood of each potential key point to obtain the first feature point of the reference band.

[0040] In a specific example, such as Figure 2 As shown, the steps for SIFT feature point extraction are as follows: The image is filtered using Gaussian blur functions at different scales to generate a series of image pyramids, also known as scale spaces; in the scale space, gradient calculations are performed on the image at each scale, and extreme points in the image are searched to detect potential keypoints; the DoG (Difference of Gaussians) image pyramid is used to detect the locations of potential keypoints; after determining the locations of potential keypoints, the gradient magnitude and direction are calculated for the neighborhood around each keypoint, and the gradient histogram is used to determine the main orientation of the keypoint; a scale- and rotation-invariant local feature descriptor is constructed in the neighborhood around each keypoint. In this embodiment, SIFT feature point extraction is a scale-invariant feature extraction method. It can extract key points of an image at different scales and rotations and generate unique descriptors. It has good robustness to changes in scale, rotation, and illumination.

[0041] In one possible implementation, calculating the second feature point of the reference band includes: setting a threshold and detecting key points in the reference band, calculating the response value of each key point; performing non-maximum suppression on the key points and retaining the key point corresponding to the maximum response value; calculating binary descriptors around the retained key points to obtain the second feature point of the reference band.

[0042] Furthermore, such as Figure 3As shown, for ORB feature point extraction, the steps include: setting an appropriate threshold, detecting key points in the image using the FAST (Features from Accelerated Segment Test) method, and calculating the response value of each key point; applying non-maximum suppression to the obtained key points, retaining the key points with the largest response; for each key point, selecting a fixed-size region, and using the BRIEF (Binary Robust Independent ElementaryFeatures) method to calculate a binary descriptor around the key point to represent the image information around the key point; The ORB feature point extraction in this embodiment has a high feature matching accuracy and low computational cost, which can effectively compensate for the shortcomings of SIFT feature point extraction.

[0043] In one possible implementation, calculating the mixed feature points of the other bands includes: calculating the third feature points of the other bands and calculating the fourth feature points of the other bands.

[0044] In one possible implementation, calculating the third feature points of the other bands includes: filtering the other bands to obtain a scale space; calculating gradients for the other bands at each scale in the scale space, and detecting potential keypoints based on the extreme points of the gradient calculations; detecting the positions of potential keypoints based on the potential keypoints using the Difference of Gaussian Image Pyramid algorithm; calculating the gradient magnitude and direction for the neighborhood of each potential keypoint based on its position, and determining the main direction of the potential keypoint based on the calculation results using a gradient histogram; and constructing a scale- and rotation-invariant local feature descriptor in the neighborhood of each potential keypoint to obtain the third feature points of the other bands.

[0045] In one possible implementation, calculating the fourth feature point of the other bands includes: setting a threshold and detecting key points in the other bands, calculating the response value of each key point; performing non-maximum suppression on the key points and retaining the key point corresponding to the maximum response value; calculating binary descriptors around the retained key points to obtain the fourth feature point of the other bands.

[0046] In one possible implementation, matching the mixed feature points of the reference band with the mixed feature points of other bands to obtain a matching feature point pair includes: matching a first feature point of the reference band with a third feature point of the other bands to obtain a first feature point pair; matching a second feature point of the reference band with a fourth feature point of the other bands to obtain a second feature point pair; and merging the first feature point pair and the second feature point pair to obtain a matching feature point pair.

[0047] In a specific example, the method includes step S140, which uses FLANN (Fast Library for Approximate Nearest Neighbors) to match feature points: Furthermore, the FLANN method is used to match the mixed feature points extracted from the reference band and other bands other than the reference band in step S130.

[0048] Furthermore, a FLANN index structure is established for the mixed feature point set. FLANN provides various index structures; this embodiment uses a KD-tree structure. For each feature point, its nearest neighbor is found using the index. FLANN uses an approximate nearest neighbor algorithm to improve matching speed. Based on the distance between the feature point and its nearest neighbor, a distance threshold is applied for filtering, retaining the matching pairs with the smallest distance. These matching pairs are the final matched feature point pairs. In this embodiment, Euclidean distance is used as the distance metric for two feature points. d 1 and d The formula for calculating the Euclidean distance between two points is:

[0049] in, The descriptor for the first feature point is in the... The value in the dimension; The descriptor for the first feature point is in the... The numerical value in the dimension; N is the feature description dimension of the feature point.

[0050] In one possible implementation, calculating the transformation model between the reference band and the other bands based on the matching feature point pairs includes: using a random sampling consensus algorithm or the least squares method to calculate the transformation model between the reference band and the other bands based on the matching feature point pairs.

[0051] In a specific example, for the matched feature point pairs, the transformation model between the reference band and other bands other than the reference band is calculated based on RANSAC (Random Sample Consensus) or the least squares method. In a specific example, the method includes: step S150, generating a transformation model; Furthermore, to complete the mixed feature point matching in step S140, a transformation model between these matching point pairs must be generated. This transformation model is used to model the geometric relationships between images. In remote sensing image processing, band matching typically uses an affine transformation model. An affine transformation is a linear transformation capable of handling translation, rotation, scaling, and shearing. The affine transformation model can be expressed as:

[0052] in, a 11 and a 22 Control the scaling factor. a 12 and a 21 Controlling the shear factor, t x and t y Control the translation factor.

[0053] Furthermore, in this embodiment, the least squares method is used to estimate the affine transformation parameters. The general steps for estimating the affine transformation parameters using the least squares method are as follows: Furthermore, consider two sets of matching feature points: feature points in the reference band image and corresponding feature points in the image to be registered. The affine transformation process is expressed as follows:

[0054] in,( u,v ) are feature points in the affine transformed band image to be registered. x,y ) are the corresponding feature points in the reference band image.

[0055] Furthermore, the least squares problem of affine transformation can be expressed as:

[0056] Where T is the affine transformation parameter matrix; This is the summation operation for all matching feature point pairs i; Let L be the square of the L2 norm (Euclidean distance) of the vector; Let x be the x-coordinate of the feature point in the reference band image in the i-th pair of feature points; Let y be the y-coordinate of the feature point in the reference band image in the i-th pair of feature points; Let x be the translation parameter along the x-direction in the affine transformation matrix; Let be the translation parameter along the y-direction in the affine transformation matrix.

[0057] Furthermore, the least squares problem is solved through the following steps: expand the affine transformation model to form a model matrix and parameter vector; calculate the partial derivatives of the objective function with respect to the parameters, set the derivatives to zero, and solve for the estimated values ​​of the parameters; construct the matrix equations, and use the known point pairs to establish a system of linear equations; use the system of linear equations to solve the least squares problem.

[0058] In a specific example, the method includes: step S160, coordinate transformation and pixel interpolation: Furthermore, based on the affine transformation parameters calculated in step S150, coordinate transformation and pixel interpolation are performed on the band image to be registered.

[0059] Furthermore, a coordinate transformation is performed on the image to be registered, which gives the coordinates of each pixel in the image to be registered. , Apply affine transformation to obtain the new coordinates after the transformation. , The calculation formula is:

[0060] Where T is the affine transformation parameter matrix obtained in step S150.

[0061] Furthermore, since the coordinates obtained after coordinate transformation are usually floating-point numbers, pixel interpolation is required to obtain the interpolated pixel values. Common interpolation methods include nearest neighbor interpolation and bilinear interpolation. This embodiment uses bilinear interpolation, and the calculation formula is as follows:

[0062] in, , , , These are the coordinates of the four nearest pixels. α and β The weights depend on the relative positions of the target coordinates within the original image coordinates. The floating-point coordinates of the target pixel to be interpolated after the affine transformation in the original image; The interpolated pixel value is calculated using bilinear interpolation based on this floating-point coordinate.

[0063] The advantages of this embodiment compared to the prior art are as follows: (1) For multi-band images of remote sensing satellites, the reference band is selected from them using wavelet correlation coefficient. The wavelet correlation coefficient is used to measure the similarity between two images in the wavelet transform domain. It is not limited to spatial similarity, but also involves the similarity of images in scale and frequency, thereby reducing the randomness of band matching and improving the accuracy of band matching.

[0064] (2) The matching method based on hybrid feature points is adopted. On the one hand, it avoids the problem of matching failure or low matching accuracy due to insufficient number of feature point pairs. On the other hand, it combines the advantages of high accuracy of SIFT feature points and fast calculation speed of ORB feature points, which greatly improves the efficiency of remote sensing image band matching.

[0065] (3) The FLANN matching method is adopted, which improves the matching speed through efficient approximate nearest neighbor search, thus making it suitable for processing large-scale satellite image data.

[0066] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. For those skilled in the art, other variations or modifications can be made based on the above description. It is impossible to exhaustively list all the implementation methods here. All obvious variations or modifications derived from the technical solutions of the present invention are still within the protection scope of the present invention.

Claims

1. A method for remote sensing image band matching based on hybrid feature points, characterized in that, The method includes: Acquire multi-band remote sensing images and perform noise reduction processing on the multi-band remote sensing images; A reference band is selected from the denoised multi-band remote sensing image to obtain the reference band and other bands besides the reference band; Calculate the mixed feature points of the reference band, and calculate the mixed feature points of the other bands; The mixed feature points of the reference band are matched with the mixed feature points of other bands to obtain matched feature point pairs; Calculate the transformation model between the reference band and the other bands based on the matching feature point pairs; The transformation model and interpolation algorithm are used to perform coordinate transformation and pixel interpolation on the other bands to achieve remote sensing image band matching.

2. The remote sensing image band matching method based on hybrid feature points according to claim 1, characterized in that, The selection of reference bands from the noise-reduced multi-band remote sensing image includes: The denoised multi-band remote sensing image is divided into multiple bands; Calculate the wavelet coefficients for each band to obtain multiple wavelet coefficients; Multiple wavelet correlation coefficients are obtained by calculating the wavelet correlation coefficients between the wavelet coefficients of each band and the wavelet coefficients of the remaining bands. Calculate the sum of wavelet correlation coefficients for each band based on multiple wavelet correlation coefficients; The highest wavelet correlation coefficient and its corresponding band are selected as the reference band.

3. The remote sensing image band matching method based on hybrid feature points according to claim 2, characterized in that, The calculation of the mixed feature points of the reference band includes: Calculate the first feature point of the reference band and calculate the second feature point of the reference band.

4. The remote sensing image band matching method based on hybrid feature points according to claim 3, characterized in that, The calculation of the first feature point of the reference band includes: The reference band is filtered to obtain the scale space; Gradient calculation is performed on the reference band at each scale of the scale space, and potential key points are detected based on the extreme points of the gradient calculation. The location of potential key points is detected based on the potential key points using the difference in Gaussian image pyramid algorithm; Based on the location of the potential key points, the gradient magnitude and direction are calculated for the neighborhood of each potential key point, and the main direction of the potential key points is determined by using the gradient histogram based on the calculation results. A scale- and rotation-invariant local feature descriptor is constructed in the neighborhood of each potential key point to obtain the first feature point of the reference band.

5. The remote sensing image band matching method based on hybrid feature points according to claim 4, characterized in that, The calculation of the second feature point of the reference band includes: Set a threshold and detect key points in the reference band, and calculate the response value of each key point; Non-maximum suppression is applied to keypoints, and keypoints corresponding to the maximum response values ​​are retained. A binary descriptor is calculated around the retained key points to obtain the second feature points of the reference band.

6. The remote sensing image band matching method based on hybrid feature points according to claim 5, characterized in that, The calculation of the mixed feature points of the other bands includes: Calculate the third feature point of the other bands and calculate the fourth feature point of the other bands.

7. The remote sensing image band matching method based on hybrid feature points according to claim 6, characterized in that, The calculation of the third feature point of the other bands includes: Filtering the other bands yields the scale space; Gradient calculations are performed on the other bands at each scale of the scale space, and potential key points are detected based on the extreme points of the gradient calculations. The location of potential key points is detected based on the potential key points using the difference in Gaussian image pyramid algorithm; Based on the location of the potential key points, the gradient magnitude and direction are calculated for the neighborhood of each potential key point, and the main direction of the potential key points is determined by using the gradient histogram based on the calculation results. A scale- and rotation-invariant local feature descriptor is constructed in the neighborhood of each potential key point to obtain the third feature point of the other bands.

8. The remote sensing image band matching method based on hybrid feature points according to claim 7, characterized in that, The calculation of the fourth feature point of the other bands includes: Set a threshold and detect key points in the other bands, and calculate the response value for each key point; Non-maximum suppression is applied to keypoints, and keypoints corresponding to the maximum response values ​​are retained. Binary descriptors are calculated around the retained key points to obtain the fourth feature points of the other bands.

9. The remote sensing image band matching method based on hybrid feature points according to claim 8, characterized in that, The process of matching the mixed feature points of the reference band with the mixed feature points of other bands to obtain matched feature point pairs includes: The first feature point pair is obtained by matching the first feature point of the reference band with the third feature point of the other bands; The second feature point is matched with the fourth feature point of the other bands to obtain the second feature point pair; The first feature point pair and the second feature point pair are merged to obtain a matching feature point pair.

10. The remote sensing image band matching method based on hybrid feature points according to claim 9, characterized in that, The step of calculating the transformation model between the reference band and the other bands based on the matching feature point pairs includes: The transformation model between the reference band and the other bands is calculated based on the matching feature point pairs using a random sampling consensus algorithm or the least squares method.

Images