A sub-pixel level image registration method, system and readable storage medium

Abstract

The application provides a sub-pixel level image registration method, a system and a readable storage medium. According to the ground feature scale invariance principle of different scale remote sensing images of the same area, the method simulates low spatial resolution panchromatic images by using low spatial resolution multispectral images, and according to the ground feature scale invariance theory of low spatial resolution panchromatic images and high spatial resolution panchromatic images of the same area, a multi-scale image accurate registration method is constructed, and sub-pixel level spatial position registration of remote sensing images is realized. The method has clear physical meaning, clear registration operation logic, and can realize accurate spatial position registration of different scale remote sensing images, and provides high-quality image data set for remote sensing image space-spectrum fusion and application.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of spatial position registration of images, and more particularly to a sub-pixel level image registration method, system and readable storage medium. BACKGROUND

[0002] Image registration is a process of geometric calibration of images taken by the same or different sensors at the same or different times for the same scene, and is a prerequisite for the business application and collaborative processing of multi-source remote sensing data. As an indispensable preprocessing step in remote sensing image application, the accuracy of image registration directly affects subsequent applications such as image fusion, change monitoring and stereo matching. At present, multi-source image registration methods can be roughly divided into three categories: region-based methods, feature-based methods and deep learning-based methods, and there are few sub-pixel level image registration methods.

[0003] Region-based image registration methods determine the optimal matching item according to certain similarity evaluation criteria by presetting a fixed-size template window, and then take the center pixel point of the template window as the homonymic feature point, and then realize image registration through the construction of a geometric change model. Common similarity measures include Sum of Squared Differences (SSD) (Hisham et al., 2015), Normalized Cross Correlation (NCC) (Martinez et al., 1999) and Mutual Information (MI) (Suri and Reinartz, 2010) and other indicators. Given the high sensitivity of these similarity measures to nonlinear gray changes, their performance in multi-source remote sensing image registration applications is not satisfactory. Ye et al. (2019) proposed a Channel Features of Orientated Gradients (CFOG) method, which generates a pixel-by-pixel dense structural feature representation for window images, improving the accuracy of image registration. On this basis, Fan et al. (2021) introduced an angle weighting strategy to optimize the direction gradient calculation process and proposed the concept of angle-weighted direction gradient. This method improves the accuracy of the calculation results, but the complexity of the calculation also increases. Although region-based methods can resist the influence of nonlinear radiation differences to some extent, their accuracy is still highly dependent on the consistency of geometric positioning between images. In addition, this type of method is highly sensitive to rotation, scale changes and local geometric differences between multi-source images, which to some extent limits their widespread applicability in practical applications.

[0004] The main steps of feature-based image registration method include feature extraction, feature description, feature matching and geometric correction. First, we need to extract the feature points with high repeatability and significant robustness between multi-source images. Then, we determine the optimal matching pair by analyzing the correlation between feature points, and establish the transformation relationship of the image. Common features include point features (Huo et al., 2012), line features (Li and Leung, 2007) and surface features (Goncalves et al., 2011). The most commonly used feature transformation method is Scale Invariant Feature Transform (SIFT). Although SIFT algorithm can effectively deal with the rotation and scale changes between images, it has been widely used in image registration field. However, due to its dependence on pixel gradient information to construct feature description, it lacks robustness to nonlinear radiation differences, so it is difficult to stably identify and match the same feature points when dealing with multi-sensor image registration tasks, and its applicability is limited. For the registration of visible light and long-wave infrared images, Aguilera et al. (2015) proposed a Log-Gabor Histogram Descriptor (LGHD) method, which has shown good performance in multi-source remote sensing image registration. To deal with the problem of nonlinear gray difference between multi-sensor images, Dellinger et al. (2015) improved the SIFT algorithm and proposed the SAR-SIFT method, which uses ratio gradient instead of traditional gray difference gradient to weaken the interference of speckle noise to some extent. Xiang et al. (2018) proposed the OS-SIFT algorithm, which uses the multi-scale exponential weighted mean ratio ROEWA and the multi-scale Sobel operator to calculate the gradient information on SAR and optical images respectively, and detects Harris feature points in the scale space. Similarly, Li et al. (2019) proposed a radiation-variation insensitive feature transform (RIFT) based on Log-Gabor filter, which uses a ring structure feature map for feature description, and achieves further improvement in image registration accuracy compared with LGHD. Zhao Siliang et al. (2024) proposed an ICP algorithm based on line feature constraint, which has the advantages of high stability and high algorithm efficiency compared with traditional ICP algorithm. Compared with region-based image registration method, feature-based registration algorithm shows significant advantages in computational efficiency and robustness.However, due to the wide coverage, complex and diverse surface features, and rich spectral information of remote sensing images, especially the hundreds of spectral channels of hyperspectral images, it is often difficult to effectively extract features with sufficient discriminability using traditional algorithms for registration, resulting in a situation where the number of feature points is often small, error matching occurs frequently, and even effective registration cannot be achieved.

[0005] In view of the limitations of traditional registration algorithms in feature extraction and feature matching in the field of image registration, the Convolutional Neural Network (CNN) in deep learning method has attracted widespread attention due to its excellent feature extraction capability. According to the different model training methods, the image registration method based on deep learning can be divided into two categories: supervised learning-based and unsupervised learning-based registration methods. The supervised learning-based registration method relies on the ground-truth, which inputs the paired image patches containing the ground-truth into the neural network, and then calculates the matching score between the patches. Typical methods include MatchNet (Han N.X et al., 2015) and LIFT (Yi K.M et al., 2016). MatchNet uses CNN to extract features, then fuses these features and inputs them into a fully connected network for similarity measurement, and finally outputs the matching score. LIFT is an extension of MatchNet, which adds multiple layers of neural networks to regress the parameters required for image transformation. Lan Z et al. (2021) introduced the idea of D2-Net feature extraction and proposed a matching method CMM-Net based on deep learning features. The supervised learning method based on deep learning uses CNN to extract high-level semantic features with high robustness, which can support accurate image registration results. However, the significant limitation of the supervised learning method is the requirement of a large number of samples with ground-truth for network training. The core strategy of the unsupervised learning-based registration method is to use self-learning methods to realize data augmentation. It mainly realizes geometric transformation such as rotation, translation, and scaling on a single image patch to form a series of matching and non-matching image pairs, and then inputs these image pairs into the network for training and outputs the transformation parameters, including Savinov N et al. (2017) and Wang S et al. (2018). In addition, there is a special unsupervised learning method, which uses artificially synthesized datasets to pre-train the model. These synthetic datasets not only contain pseudo-labels, but also define explicit mapping relationships, so that the model can learn effective feature representations without real labels (Detone D et al., 2017). In addition, some scholars train CNN and Generative Adversarial Nets (GAN) in an unsupervised or supervised manner to directly output image displacement fields to achieve end-to-end image registration (Yang et al., 2021; Luo et al., 2021; Ni et al., 2021). Although the integration of deep learning technology into current research has become a trend, deep learning methods are still highly dependent on the accurate and real value labeling of control points in practical applications, and large-scale public training datasets have not yet been released, which seriously hinders the widespread promotion of related applications.

[0006] In the future of the development of image registration technology, the core difficulty and key focus of the research still focus on the continuous improvement of registration efficiency and registration accuracy. In this field, the method based on the spatial feature scale invariance of ground objects shows unique superiority compared with traditional algorithms.

[0007] For the spatial position registration of low-resolution multispectral images and high-resolution panchromatic images, there are two conventional registration methods, one is to orthorectify the multispectral images and the panchromatic images respectively, and to register the two images to the real spatial position; the other is to register the multispectral images and the panchromatic images by image-to-image registration. However, no matter which method is adopted, there are internal distortions of the images or inaccurate selection of control points during registration, which leads to inaccurate registration or even cannot meet the accuracy requirements of the relevant technical specifications that the image spatial position error is less than 1.5 pixels in the plain area and less than 3 pixels in the mountainous area. SUMMARY

[0008] In view of the above problems, the purpose of the present application is to provide a sub-pixel level image registration method, system and readable storage medium.

[0009] The first aspect of the present application provides a sub-pixel level image registration method, which comprises the following steps:

[0010] S1: inputting a remote sensing image, wherein the remote sensing image comprises a panchromatic image and a multispectral image;

[0011] S2: performing geometric initial correction on the high-resolution panchromatic image and the low-resolution multispectral image respectively, finding the same ground feature points, and confirming that the spatial positions of the panchromatic image and the multispectral image are consistent;

[0012] S3: up-sampling the multispectral image to the same resolution as the panchromatic image;

[0013] S4: simulating a low-resolution panchromatic image (i.e. average intensity image I) by using the multispectral image;

[0014] S5: taking the upper left corner point (x0, y0) of the average intensity image as the upper left corner point, and cutting the average intensity image sample block I 00 and the panchromatic image sample block P0 according to the preset fixed window (such as 128*128 pixels);

[0015] S6: taking the average intensity image sample block I 00 (x0, y0) as the center, determining the surrounding 8 pixel points, taking each of the 8 pixel points as the upper left corner point in turn, and cutting the average intensity sample block according to the preset fixed window, so as to cut the average intensity image sample block I 01 -I 08;

[0016] S7: matching the spatial position of the average intensity image sample block I 00 -I 08 to the same reference as the panchromatic image sample block P0, and calculating the correlation coefficient between the average intensity image sample block I 00 -I 08 and the panchromatic image sample block P0.

[0017] S8: judging whether the correlation coefficient between the average intensity image sample block I 00 and the panchromatic image sample block P0 is the maximum, if yes, determining (x0, y0) as the accurate registration result, and executing S13; if not, executing S9.

[0018] S9: taking the average intensity image sample block with the maximum correlation coefficient with the panchromatic image sample block P0 as the n-th registration result, and recording its top-left corner point as (x n ,y n ); wherein, n is the number of registrations; and determining the neighboring independent pixel points around (x n ,y n ) as the center.

[0019] S10: taking the center point (x n ,y n ) and the independent pixel points as the top-left corner coordinate points, and cutting the average intensity image sample block I n0 -I n8 according to the preset fixed window size in sequence.

[0020] S11: matching the spatial position of the average intensity image sample block I n0 -I n8 to the same reference as the panchromatic image sample block P0, and calculating the correlation coefficient between the average intensity image sample block I n0 -I n8 and the panchromatic image sample block P0.

[0021] S12: judging whether the correlation coefficient between the average intensity image sample block with the center point (x n ,y n ) and the panchromatic image sample block P0 is the maximum, if yes, determining the final registration scheme, and executing S13; if not, setting n = n + 1, and returning to S9.

[0022] S13: according to the final accurate registration scheme between the multi-spectral image and the panchromatic image, taking the high-resolution panchromatic image as the reference, performing accurate registration on the multi-spectral image, i.e. translating the multi-spectral image coordinate by (x0-x n ,y0-y n ) pixels, and saving the registration image result.

[0023] Preferably, the low-resolution panchromatic image is simulated by using the multi-spectral image, specifically, the low-resolution panchromatic image is simulated by using the average intensity of the multi-spectral image, and the expression is as follows:

[0024]

[0025] wherein I is the low-resolution panchromatic image, M i is the multi-spectral image, and N is the total number of bands of the multi-spectral image.

[0026] Preferably, the average intensity image sample block I 00 is determined by using the following formula: 08 The calculation formula of the correlation coefficient of the average intensity image sample block I

[0027]

[0028] wherein r(I 0i ,P0) is the correlation coefficient of the average intensity image sample block I 0i and the panchromatic image sample block P0, Cov(I 0i ,P0) is the covariance of the average intensity image sample block I 0i and the panchromatic image sample block P0, and is the standard deviation of the average intensity image sample block I 0i , and is the standard deviation of the panchromatic image sample block P0.

[0029] Preferably, the independent pixels in S9 are determined, and specifically, the independent pixels are determined by using the following formula:

[0030] (x n ,y n ) as the center, there are 8 adjacent pixel points around the center, the remaining pixel points are taken as the independent pixel points after removing the adjacent pixel points overlapping with the center (x n-1 ,y n-1 ).

[0031] Preferably, the average intensity image sample block I n0 is determined by using the following formula: n8

[0032]

[0033] wherein r(I ni ,P0) is the correlation coefficient of the average intensity image sample block I ni and the panchromatic image sample block P0, Cov(I ni ,P0) is the covariance of the average intensity image sample block I ni and the panchromatic image sample block P0.​ the standard deviation of the average intensity image sample block I 0i ; the standard deviation of the panchromatic image sample block P0.

[0034] The second aspect of the present application provides a sub-pixel image registration system, comprising a memory and a processor, the memory comprising a sub-pixel image registration method program, the sub-pixel image registration method program being executed by the processor to implement the following steps:

[0035] S1: inputting a remote sensing image, the remote sensing image comprising a panchromatic image and a multi-spectral image;

[0036] S2: performing geometric initial correction on the high-resolution panchromatic image and the low-resolution multi-spectral image respectively, finding homonymic ground object points, and confirming that the spatial positions of the panchromatic image and the multi-spectral image are consistent;

[0037] S3: up-sampling the multi-spectral image to the same resolution as the panchromatic image;

[0038] S4: simulating a low-resolution panchromatic image using the multi-spectral image;

[0039] S5: taking a pixel point (x0, y0) on the low-resolution panchromatic image as the upper left corner point, and cutting an average intensity image sample block I 00 and a panchromatic image sample block P0 according to a preset fixed window (such as 128*128 pixels);

[0040] S6: taking the average intensity image sample block I 00 as the center, determining 8 adjacent pixel points around it, and cutting average intensity image sample blocks I 01 -I 08 in turn according to the preset fixed window with the 8 adjacent pixel points as the upper left corner points;

[0041] S7: matching the spatial positions of the average intensity image sample block I 00 -I 08 and the panchromatic image sample block P0 to the same reference, and calculating the correlation coefficient of the average intensity image sample block I 00 -I 08 and the panchromatic image sample block P0;

[0042] S8: judging whether the correlation coefficient of the average intensity image sample block I 00 is the maximum, if yes, determining that (x0, y0) is the accurate registration result, and executing S13; if not, executing S9;

[0043] S9: taking the average intensity image sample block with the largest correlation coefficient with the panchromatic image sample block P0 as the n-th registration result, and recording the top-left corner point thereof as (x n ,y n ); wherein n is the number of registrations;

[0044] determining the neighboring independent pixel points around the center point (x n ,y n );

[0045] S10: taking the center point (x n ,y n ) and the independent pixel points as the top-left corner points, and cutting the average intensity image sample blocks I n0 -I n8 in sequence according to the preset fixed window size;

[0046] S11: matching the average intensity image sample blocks I n0 -I n8 to the same reference as the panchromatic image sample block P0 in spatial position, and calculating the correlation coefficient of the average intensity image sample blocks I n0 -I n8 with the panchromatic image sample block P0;

[0047] S12: judging whether the correlation coefficient of the average intensity image sample block I n containing the center point (x n ,y n0 ) with the panchromatic image sample block P0 is the largest, if yes, executing S13; if no, setting n=n+1 and returning to S9;

[0048] S13: according to the final multi-spectral image and panchromatic image accurate registration scheme, taking the high-resolution panchromatic image as the reference, performing accurate registration on the multi-spectral image, i.e. translating the multi-spectral image coordinates by (x0-x n ,y0-y n ) pixels, and saving the registration image result.

[0049] Preferably, the low-resolution panchromatic image is simulated by using the multi-spectral image, specifically, the low-resolution panchromatic image is simulated by using the average intensity of the multi-spectral image, and the expression is as follows:

[0050]

[0051] wherein I is the low-resolution panchromatic image, M i is the multi-spectral image, and N is the total number of bands of the multi-spectral image.

[0052] Preferably, the average intensity image sample block I 00 -I 08The formula for calculating the correlation coefficient of the panchromatic image sample block P0 is:

[0053]

[0054] Wherein, r(I 0i ,P0) is the average intensity image sample block I 0i The correlation coefficient of the panchromatic image sample block P0, Cov(I 0i ,P0) is the average intensity image sample block I 0i The covariance of the panchromatic image sample block P0, The standard deviation of the average intensity image sample block I 0i , The standard deviation of the panchromatic image sample block P0.

[0055] The average intensity image sample block I n0 -I n8 The formula for calculating the correlation coefficient of the panchromatic image sample block P0 is:

[0056]

[0057] Wherein, r(I ni ,P0) is the average intensity image sample block I ni The correlation coefficient of the panchromatic image sample block P0, Cov(I ni ,P0) is the average intensity image sample block I ni The covariance of the panchromatic image sample block P0, The standard deviation of the average intensity image sample block I 0i , The standard deviation of the panchromatic image sample block P0.

[0058] Preferably, the determination of the independent pixel in S9 is specifically:

[0059] (x n ,y n ) as the center, there are 8 adjacent pixel points around it, remove the adjacent pixel points overlapping with (x n-1 ,y n-1 ) as the center point, and the remaining pixel points are taken as independent pixel points.

[0060] The third aspect of the present application provides a computer readable storage medium, the computer readable storage medium comprises a sub-pixel level image registration method program, when the sub-pixel level image registration method program is executed by a processor, the steps of a sub-pixel level image registration method are realized.

[0061] Compared with the prior art, the beneficial effects of the technical scheme of the present application are that the present application provides a sub-pixel level image registration method, system and readable storage medium.

[0062] The method provided by the present application realizes sub-pixel level spatial position registration of remote sensing images according to the feature scale invariance principle of ground objects in remote sensing images of different scales in the same area, simulates low spatial resolution panchromatic images by using low spatial resolution multispectral images, and constructs a multiscale image accurate registration method based on the feature scale invariance theory of ground objects in low spatial resolution panchromatic images and high spatial resolution panchromatic images in the same area. The method has clear physical meaning and clear registration operation logic, and can realize accurate spatial position registration of remote sensing images of different scales, thereby providing high-quality image data sets for remote sensing image space-spectrum fusion and application.

[0063] The method provided by the present application can realize accurate spatial position registration of multispectral images and panchromatic images, multispectral images and multispectral images, and panchromatic images and panchromatic images in the same area, and has no limitation on the type, accuracy and band number of remote sensing images, and has openness and universality. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 It is a flowchart of the sub-pixel level image registration method described in Embodiment 1.

[0065] Figure 2 It is a schematic diagram of the position relationship of a center point (x0, y0) and its adjacent pixel points.

[0066] Figure 3 It is a schematic diagram of the position relationship of the upper left pixel points of the simulated image sample blocks of different registration times.

[0067] Figure 4 It is a GF1 panchromatic image and multispectral band fusion image (2-meter resolution) after accurate registration. DETAILED DESCRIPTION

[0068] In order to enable the above-mentioned objects, features and advantages of the present application to be more clearly understood, the present application will be further described below in conjunction with the drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0069] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and therefore, the scope of protection of the present application is not limited by the specific embodiments disclosed below.

[0070] Embodiment 1

[0071] As Figure 1As shown, the embodiment discloses a sub-pixel level image registration method, which comprises the following steps:

[0072] S1: inputting a remote sensing image, wherein the remote sensing image comprises a panchromatic image and a multispectral image.

[0073] It should be noted that the embodiment inputs a panchromatic band P of a remote sensing image and a multispectral band M of the remote sensing image i (including near-infrared N, red R, green G, blue B, etc.), and the method does not limit the number of bands of the multispectral image.

[0074] S2: performing geometric initial correction on the high-resolution panchromatic image and the low-resolution multispectral image respectively, finding a same-named ground object point, and confirming that the spatial positions of the panchromatic image and the multispectral image are consistent.

[0075] It should be noted that the embodiment performs geometric initial correction on the high-resolution panchromatic image and the low-resolution multispectral image respectively by using satellite RPC parameters in satellite imaging, finds a same-named ground object point, and confirms that the spatial positions of the panchromatic image and the multispectral image are basically consistent.

[0076] S3: upsampling the multispectral image to the same resolution as the panchromatic image.

[0077] S4: simulating a low-resolution panchromatic image by using the multispectral image.

[0078] It should be noted that the embodiment upsamples the multispectral image to the same resolution as the panchromatic image, and simulates a low-resolution panchromatic image I by using the multispectral image, that is, by default, calculating the average intensity of the multispectral image to simulate the low-resolution panchromatic image.

[0079] wherein I is a low-resolution panchromatic image, M i is a multispectral image, and N is the total number of bands of the multispectral image.

[0080] S5: taking an upper-left corner point (x0, y0) of the low-resolution panchromatic image as a starting point, and cutting an average intensity image sample block I 00 and a panchromatic image sample block P0 according to a preset fixed window (such as 128*128 pixels).

[0081] S6: taking the average intensity image sample block I 00 and the upper-left corner point (x0, y0) as a center, determining eight adjacent pixel points around the center, taking the eight adjacent pixel points as upper-left corner points in sequence, and cutting average intensity sample blocks according to the preset fixed window, so that average intensity image sample blocks I 01 -I 08 of all adjacent pixel points can be cut.

[0082] It should be noted that the present embodiment takes the point of the upper left corner (x0, y0) and cuts the average intensity image I according to the preset fixed window (such as 128*128 pixels) 00 With the panchromatic image P0, there are 8 adjacent pixel points around the center (x0, y0), and the fixed window size image block is cut on the average intensity image I with the upper left corner point as the center, and a total of 8 new image blocks I are obtained 01 -I 08 The center point (x0, y0) and its adjacent point position relationship diagram is shown in Figure 2

[0083] S7: The average intensity image sample block I 00 -I 08 is matched to the same reference as the spatial position of the panchromatic image sample block P0, and the correlation coefficient of the average intensity image sample block I 00 -I 08 and the panchromatic image sample block P0 is calculated. The calculation formula is:

[0084]

[0085] Wherein, r(I 0i ,P0) is the correlation coefficient of the average intensity image sample block I 0i and the panchromatic image sample block P0, Cov(I 0i ,P0) is the covariance of the average intensity image sample block I 0i and the panchromatic image sample block P0, is the standard deviation of the average intensity image sample block I 0i , is the standard deviation of the panchromatic image sample block P0.

[0086] S8: Determine whether the correlation coefficient of the average intensity image sample block I 00 and the panchromatic image sample block P0 is the maximum, if it is the maximum, it is determined that (x0, y0) is the accurate registration result, and S13 is executed; if not, S9 is executed.

[0087] It should be noted that the present embodiment matches the spatial positions of the 9 average intensity image blocks I 00 -I 08 and the panchromatic image P0 to the same reference, calculates the correlation coefficients of all 9 average intensity images and the panchromatic image. If the upper left corner point of the average intensity image block with the maximum correlation coefficient with the panchromatic image is exactly (x0, y0), the multispectral image and the panchromatic image have realized accurate registration, and further registration is not required. If not, take the one with the maximum correlation coefficient as the first registration result, and record its upper left corner point as (x1, y1).

[0088] ​S9: take the average intensity image sample block with the largest correlation coefficient with the panchromatic image sample block P0 as the n-th registration result, and mark its upper left corner point as (x n ,y n ); wherein n is the number of registrations; and determine the neighboring independent pixel points around the center point (x n ,y n ).

[0089] S10: take the center point (x n ,y n ) and the independent pixel points as the upper left corner points, and cut the average intensity image sample blocks I n0 -I n8 in sequence according to the preset fixed window size.

[0090] S11: match the average intensity image sample blocks I n0 -I n8 to the same reference as the spatial position of the panchromatic image sample block P0, and calculate the correlation coefficient of the average intensity image sample blocks I n0 -I n8 with the panchromatic image sample block P0.

[0091] It should be noted that the first registration and the second registration are used to illustrate the method described in the embodiment. If the upper left corner point of the average intensity image block with the largest correlation coefficient with the panchromatic image sample block P0 is not (x0, y0), then (x0, y0) is marked as (x1, y1).

[0092] There are 8 neighboring pixel points around (x1, y1), and the remaining neighboring pixel points are taken as independent pixel points after removing the overlapping points with the neighboring pixel points of the center point (x0, y0) in the first registration. If (x1, y1) is diagonal to (x0, y0), there are 5 independent pixel points. If (x1, y1) is in the coordinate axis direction of (x0, y0), there are 3 independent pixel points. The image block I 10 -I 18 of the fixed window size is cut with the center point (x1, y1) and the independent pixel points as the upper left corner points.

[0093] S12: determine whether the correlation coefficient of the average intensity image sample block with (x n ,y n ) as the upper left corner point with the panchromatic image sample block P0 is the largest. If yes, the final registration scheme is obtained and S13 is performed. If no, n = n + 1, and return to S9.

[0094] It should be noted that the embodiment calculates the correlation coefficient of the second cutting image block and the panchromatic image. If the image block with the largest correlation coefficient with the panchromatic image block is still smaller than the image block where (x1, y1) is located, the multispectral image coordinates are translated by (x0-x1, y0-y1) to obtain the accurate registration result (in the header file, the multispectral image space position is registered to the panchromatic image, and the accurate registration of the two images is realized). If the image block with the largest correlation coefficient is larger than the image block where (x1, y1) is located, the upper left corner point thereof is recorded as (x2, y2) as the second registration result.

[0095] Average intensity image sample block I 10 -I 18 The correlation coefficient of the average intensity image sample block I

[0096]

[0097] The correlation coefficient of the average intensity image sample block I 1i and the panchromatic image sample block P0, 1i The correlation coefficient of the average intensity image sample block I 1i and the panchromatic image sample block P0, The standard deviation of the average intensity image sample block I 1i , The standard deviation of the panchromatic image sample block P0.

[0098] S13: According to the final multispectral image and panchromatic image accurate registration scheme, the multispectral image is accurately registered based on the high-resolution panchromatic image, that is, the multispectral image coordinates are translated by (x0-x n , y0-y n ) pixels, and the registration image result is saved.

[0099] The embodiment is based on the scale invariance principle of the feature of the same region in different spatial resolution images, and the low-resolution image is up-sampled based on the high-resolution image. The multispectral image is processed by using the panchromatic image simulating the resolution of the multispectral image to realize the waveband reduction of the multispectral image. The spatial position accurate registration scheme of the multiscale image is determined by the principle that the correlation between the same region in different resolution panchromatic images is the largest during accurate registration, so as to achieve the sub-pixel registration of the multiscale remote sensing image.

[0100] The accurate registration of the spatial position of high-resolution panchromatic image and low-resolution multispectral image is the prerequisite for carrying out high-quality space-spectrum fusion and fine remote sensing analysis. Generally, the high-resolution panchromatic image and the low-resolution multispectral image are respectively geometrically pre-corrected according to the on-satellite RPC parameters during satellite imaging, and then the pre-corrected results are registered in spatial position to form a set of space-spectrum data (which can be further fine-corrected according to ground control points) corresponding to each pixel (ground object) for the basis of image fusion. Among them, the registration accuracy of the panchromatic image and the multispectral image directly affects the effect of space-spectrum fusion of remote sensing image, and the higher the registration accuracy is, the better the fusion effect is, and vice versa. The technical solution of the embodiment to solve the problem of accurate registration of spatial position between images of different resolutions is as follows:

[0101] Firstly, the average intensity of the low-resolution multispectral image can be used to simulate the low-resolution panchromatic image. According to the principle of image scale invariance, the correlation coefficient between the simulated low-resolution panchromatic image and the high-resolution panchromatic image is the largest when the low-resolution multispectral image and the high-resolution panchromatic image are accurately registered, and the correlation coefficient becomes smaller as the image registration accuracy decreases. Therefore, the accuracy of image registration can be measured by the correlation coefficient between the two.

[0102] Secondly, it is assumed that the image results corrected according to the on-satellite RPC parameters do not have internal distortion, and any small window image can be accurately registered according to the translation mode after correction. Therefore, it is not necessary to calculate the correlation coefficient between the whole image, but only to select a small window image for correlation calculation to find a homonymic ground object point between the simulated low-resolution panchromatic image and the high-resolution panchromatic image, so as to realize the accurate registration of the whole image, reduce the calculation amount and improve the registration efficiency. That is, for the bundled data obtained simultaneously by the multispectral image and the panchromatic image, it is assumed that the image results corrected according to the on-satellite RPC parameters have no internal distortion, and the two can be one-to-one registered, but it is difficult to accurately determine the homonymic ground object point by naked eye when the multispectral image is up-sampled to the same resolution as the panchromatic image, and it is necessary to determine it by calculating the correlation coefficient of the average intensity of the panchromatic image and the multispectral image. This is conducive to eliminating the influence of human factors on the registration accuracy and obtaining more accurate registration results with the help of computer algorithm.

[0103] Finally, the registration of the multispectral image and the panchromatic image in this embodiment is essentially to find a same-named ground feature point between a low-resolution image and a high-resolution image under high-resolution conditions. In general, it is difficult to find such a same-named ground feature point with naked eyes. Although the RPC parameters on the satellite do not cause distortion between the two images, which makes it impossible to achieve one-to-one correspondence between the pixels of different images, it is not possible to achieve accurate registration between the two images after geometric correction by the RPC parameters on the satellite. Therefore, there is room for further accurate registration between the multispectral image and the panchromatic image. The registration scheme can be determined by the principle of maximizing the correlation coefficient between the simulated low-resolution panchromatic image and the original high-resolution panchromatic image.

[0104] The method described in this embodiment is suitable for spatially accurately registering a high-resolution panchromatic image and a low-resolution multispectral image of a remote sensing image, and is particularly suitable for sub-pixel level accurate registration of a high-resolution panchromatic image and a low-resolution multispectral image taken by a satellite sensor under the same imaging conditions, thereby providing a high-precision image registration result for remote sensing image space-spectral fusion. At the same time, the method described in this embodiment is also suitable for accurate registration between multispectral and panchromatic images, multispectral and multispectral images, and panchromatic and panchromatic images of different spatial resolutions.

[0105] As a specific embodiment, the method described in this embodiment will be described below in combination with a specific example:

[0106] To achieve the purpose of sub-pixel level image accurate registration based on scale invariance, this embodiment mainly uses ENVI remote sensing image processing software to achieve the purpose. A GF1 satellite remote sensing image (GF1_PMS1_E116.7_N40.3_20160508_L1A0001570385) of Beijing Huairou area on May 8, 2016, which has a panchromatic band (P), a blue band (B), a green band (G), a red band (R), and a near-infrared band (N), is further described. The resolution of the panchromatic band of this image is 2 meters, and the resolution of the multispectral band is 8 meters.

[0107] S1: input a GF1 satellite remote sensing image (GF1_PMS1_E116.7_N40.3_20160508_L1A0001570385) of Beijing Huairou area on May 8, 2016, which has a panchromatic band (P), a blue band (B), a green band (G), a red band (R), and a near-infrared band (N).

[0108] S2: apply the RPC parameters provided by the GF1 image to geometrically correct the panchromatic and multispectral data.

[0109] S3: upsample the multispectral data to the same resolution as the panchromatic band, i.e. 2 meter resolution. Check the positional consistency of the co-located points of the panchromatic band and the multispectral band images. If there is a significant deviation, then translate the multispectral data to the panchromatic band.

[0110] S4: simulate a low resolution panchromatic image from the multispectral image; calculate the multispectral average intensity, the formula is fix((b1+b2+b3+b4) / 4.0).

[0111] S5: select the pixel point (9000, 9000) as the upper left corner point on the average intensity image, and crop the average intensity image and the panchromatic image sample block according to the window size of 256*256 pixels.

[0112] S6: take (9000, 9000) as the center, and its 8 adjacent pixel points around it, in turn, take them as the upper left corner point, crop the same window size image on the average intensity image, a total of 10 sample block images are obtained.

[0113] S7: match the spatial positions of the 9 average intensity image sample blocks obtained to the same reference as the panchromatic image sample block, and calculate the correlation coefficients of all 9 images and the panchromatic image. The calculation results are shown in Table 1.

[0114] Table 1 Correlation statistics table of multispectral average intensity image block I 0i and panchromatic image block P0

[0115] Correlation I 08 ]] I 07 ]] I 06 ]]> I 05 ]] I 04 ]] I 03 ]] I 02 ]] I 01 ]] I 00 ]]> P0 0.645 0.640 0.628 0.617 0.599 0.608 0.613 0.633 0.628

[0116] S8: according to the calculation results of S7, the average intensity image I 08 is obtained, and the correlation coefficient of the sample block and the panchromatic sample block is the largest, which is 0.645.

[0117] S9: take the upper left pixel (8999, 8999) of the I 08 sample block as the center, and there are 8 pixel points around it, remove the overlapping points with the first registration, and take the center point (8999, 8999) and the independent pixel points as the upper left corner points, and in turn crop the same window size of the average intensity image block I 10 , I 11 , I 12 , I 16 , I 17 , I 18 .

[0118] S11: match the spatial positions of the average intensity image block and the panchromatic image sample block to the same reference, and calculate the correlation coefficients of the above samples and the panchromatic image sample block. The multispectral average intensity image block I 1i and the panchromatic image block P0 correlation statistics are shown in Table 2.

[0119] Table 2 Multi-spectral average intensity image block I 1i Correlation statistics table of panchromatic image block P0

[0120] Correlation I 18 ]]> I 17 ]]> I 16 ]]> I 12 ]]> I 11 ]] I 10 ]]> [P0] 0.647 0.649 0.644 0.632 0.644 0.645

[0121] S12: According to Table 2, the average intensity image I can be obtained 17 The correlation coefficient of the sample block and the panchromatic sample block is the largest, which is 0.649.

[0122] I 17 The upper left corner pixel (8998, 8999) of the sample block is taken as the center, and there are 8 adjacent pixel points around it. The image block I is cropped in the same window size, taking the center point (8998, 8999) and the independent pixel point as the upper left corner point, and the same window size is cropped in sequence 20 , I 26 , I 27 , I 28 The upper left corner pixel position of each simulation image sample block is shown in Figure 3 , and the black, red and blue frames and numbers in the figure are the positions and numbers of the center points and their adjacent pixel points in the first, second and third registration.

[0123] S13: Similarly, the correlation coefficient of the above sample and the panchromatic image sample block is still obtained, and the average intensity image I 20 (i.e. I 17 ) sample block and the panchromatic sample block have the largest correlation coefficient. The multi-spectral average intensity image block I 2i The correlation statistics of the panchromatic image block P0 are shown in Table 3.

[0124] Table 3 Multi-spectral average intensity image block I 2i Correlation statistics table of panchromatic image block P0

[0125]

[0126]

[0127] It can be seen that the multi-spectral image pixel (8998, 8999) and the panchromatic image pixel (9000, 9000) are the same named ground object points, so that the multi-spectral image and the panchromatic image are accurately registered, and sub-pixel level accurate registration data is provided for spectral fusion of remote sensing images. The GF1 accurate registration panchromatic and multi-spectral band fusion image map of this embodiment is shown in Figure 4 .

[0128] The embodiment is based on the scale invariance principle of remote sensing image, and precisely registers the 8-meter resolution multispectral image and 2-meter resolution panchromatic image of the domestic GF1 to realize the registration accuracy of multiscale image sub-pixel level error, and theoretically achieve zero error accuracy, which is better than the registration accuracy of 1.5-3 pixels required in the conventional geometric orthographic correction. Thus, high-quality multispectral and panchromatic image sets are provided for subsequent multiscale remote sensing image space-spectrum fusion, which can greatly improve the spatial geometric texture detail information and spectral fidelity of the image space-spectrum fusion results, and improve the image application effect.

[0129] The embodiment mainly aims at the spatial position precise registration of high-resolution panchromatic image and low-resolution multispectral image, and is especially suitable for the sub-pixel level precise registration of high-resolution panchromatic image and low-resolution multispectral image shot by a satellite sensor under the same imaging condition. Meanwhile, the method described in the embodiment is also suitable for the precise registration between different multispectral images and panchromatic images, multispectral images and multispectral images, and panchromatic images and panchromatic images in the same area. First, the multispectral image and the panchromatic image are respectively subjected to initial correction according to the RPC parameters on the satellite, and the multispectral image is subjected to up-sampling, and the average intensity of the low-resolution multispectral image is used to simulate the low-resolution panchromatic image. Second, the upper left corner point and the adjacent pixels are selected, a fixed window is selected, the simulated image and the panchromatic image block are respectively cut, the correlation coefficient of the simulated image block and the panchromatic image block is calculated, and the maximum correlation coefficient is found, so as to determine the precise registration scheme of the multispectral image and the panchromatic image. Finally, the low-resolution multispectral image is subjected to precise registration according to the registration scheme, with the high-resolution panchromatic image as the reference. The technical method has solid theoretical basis, clear physical meaning, wide application object, simple operation and high operation efficiency. The spatial positions of the multispectral image and the panchromatic image after the precise registration are accurately corresponding, and the position error reaches the sub-pixel level, which provides high-quality image sets for subsequent image space-spectrum fusion. Under the background of the current rapid development of high-resolution satellite remote sensing, the method has a great promoting effect on the popularization and application of domestic high-resolution images in various industries.

[0130] Embodiment 2

[0131] The embodiment discloses a sub-pixel level image registration system, which comprises a memory and a processor, and the memory comprises a sub-pixel level image registration method program. When the sub-pixel level image registration method program is executed by the processor, the following steps are realized.

[0132] S1: inputting a remote sensing image, wherein the remote sensing image comprises a panchromatic image and a multispectral image;

[0133] S2: performing geometric initial correction on the high-resolution panchromatic image and the low-resolution multispectral image respectively, finding the same named place, and confirming that the spatial positions of the panchromatic image and the multispectral image are consistent;

[0134] S3: upsample the multispectral image to the same resolution as the panchromatic image;

[0135] S4: simulate a low resolution panchromatic image using the multispectral image;

[0136] S5: take the upper left corner point (x0, y0) of the low resolution panchromatic image, and cut out an average intensity image sample block I 00 from the panchromatic image sample block P0 according to a preset fixed window size;

[0137] S6: take the average intensity image sample block I 00 as the center, determine 8 adjacent pixel points around it, and take them as the upper left corner points in turn to cut out 8 average intensity image sample blocks I 01 -I 08 around them according to a preset fixed window size;

[0138] S7: match the spatial positions of the average intensity image sample block I 00 -I 08 and the panchromatic image sample block P0 to the same reference, and calculate the correlation coefficient of the average intensity image sample block I 00 -I 08 and the panchromatic image sample block P0;

[0139] S8: determine whether the correlation coefficient of the average intensity image sample block I 00 is the largest, if yes, determine (x0, y0) as the accurate registration result, and execute S13; if not, execute S9;

[0140] S9: take the average intensity image sample block with the largest correlation coefficient as the n-th registration result, and take its upper left corner point as (x n ,y n ); wherein n is the number of registrations; take the independent pixel points around (x n ,y n ) as the center;

[0141] S10: take the center point (x n ,y n ) and the independent pixel points as the upper left corner points, and cut out 8 average intensity image sample blocks I n0 -I n8 around them according to a preset fixed window size in turn;

[0142] S11: match the spatial positions of the average intensity image sample block I n0 -I n8 and the panchromatic image sample block P0 to the same reference, and calculate the correlation coefficient of the average intensity image sample block I n0I n8 the correlation coefficient of the average intensity image sample block I n and the panchromatic image sample block P0;

[0143] S12: judging whether the correlation coefficient of the average intensity image sample block I n where the center point (x n0 ,y n ) is located is the maximum, if yes, determining the accurate registration scheme, and performing S13; if no, letting n=n+1, and returning to S9;

[0144] S13: according to the final multispectral image and panchromatic image accurate registration scheme, taking the high resolution panchromatic image as the reference, performing accurate registration on the multispectral image, i.e. translating the multispectral image coordinates by (x0-x n ,y0-y n ) pixels, and saving the registration image result.

[0145] According to the embodiment of the present application, taking the high resolution panchromatic image as the reference, performing accurate registration on the multispectral image, i.e. translating the multispectral image coordinates by (x0-x n ,y0-y i ) pixels, and saving the registration image result.

[0146] According to the embodiment of the present application, the low resolution panchromatic image is simulated by using the multispectral image, specifically, the low resolution panchromatic image is simulated by using the average intensity of the multispectral image, and the expression is as follows:

[0147]

[0148] wherein I is the low resolution panchromatic image, M 00 is the multispectral image, and N is the total number of wave bands of the multispectral image.

[0149] According to the embodiment of the present application, the calculation formula of the correlation coefficient of the average intensity image sample block I 00 -I 08 and the panchromatic image sample block P0 is as follows:

[0150]

[0151] wherein r(I 0i ,P0) is the correlation coefficient of the average intensity image sample block I 0i and the panchromatic image sample block P0, is the standard deviation of the average intensity image sample block I 0i , is the standard deviation of the panchromatic image sample block P0.

[0152] The average intensity image sample block I n0 -I n8The formula for calculating the correlation coefficient of the panchromatic image sample block P0 is:

[0153]

[0154] Wherein, r(I ni ,P0) is the average intensity image sample block I ni The correlation coefficient of the panchromatic image sample block P0, is the standard deviation of the average intensity image sample block I 0i , is the standard deviation of the panchromatic image sample block P0.

[0155] According to the embodiment of the present application, the determination of the independent pixel in S9 is specifically:

[0156] (x n ,y n ) as the center, there are 8 adjacent pixel points around it, remove the adjacent pixel points overlapping with (x n-1 ,y n-1 ) as the center, and the remaining pixel points are taken as the independent pixel points. If (x n ,y n ) and (x n-1 ,y n-1 ) are diagonal, then there are 5 independent pixel points; if (x n ,y n ) and (x n-1 ,y n-1 ) are in the direction of the coordinate axis, then there are 3 independent pixel points.

[0157] Embodiment 3

[0158] The embodiment discloses a computer readable storage medium, wherein a sub-pixel level image registration method program is included, and when the sub-pixel level image registration method program is executed by a processor, the steps of the sub-pixel level image registration method in the embodiment 1 are realized.

[0159] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division, and actual implementation can have another division manner, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed components can be through some interfaces, indirect coupling or communication connection between devices or units, which can be electrical, mechanical or other forms.

[0160] The units described as separate components above can or can not be physically separate, and the components shown as units can or can not be physical units; they can be located in one place or distributed on multiple network units; and part or all of the units can be selected according to actual needs to achieve the purposes of the embodiments.

[0161] In addition, each functional unit in each embodiment of the present application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in the form of hardware or in the form of hardware plus software functional units.

[0162] Those of ordinary skill in the art can understand that all or part of the steps of the above method embodiments can be completed by relevant hardware instructed by programs, and the foregoing programs can be stored in a computer readable storage medium, and when the programs are executed, steps including the above method embodiments are executed; and the foregoing storage medium includes mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic discs or optical discs, and various media that can store program codes.

[0163] Alternatively, the integrated units of the present application, if implemented in the form of software functional modules and sold or used as independent products, can also be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the embodiments of the present application can be embodied in the form of software products, and the computer software products are stored in a storage medium, including a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the embodiments of the present application. The foregoing storage medium includes mobile storage devices, ROMs, RAMs, magnetic discs or optical discs, and various media that can store program codes.

Claims

1. A sub-pixel-level image registration method, characterized in that, The method includes the following steps: S1: Input remote sensing images, including panchromatic images and multispectral images; S2: Perform initial geometric corrections on the high-resolution panchromatic image and the low-resolution multispectral image respectively, find corresponding ground features, and confirm that the spatial positions of the panchromatic image and the multispectral image are consistent. S3: Upsample the multispectral image to the same resolution as the panchromatic image; S4: Low-resolution panchromatic image simulation using the average intensity of multispectral images; S5: Take the top left pixel of the low-resolution panchromatic image as ( , The average intensity image sample block is cropped from the points according to the preset fixed window. With panchromatic image sample blocks ; S6: Image sample blocks with average intensity Top left corner ( , Using a fixed window as the center, determine the eight surrounding pixels, and then, taking that pixel as the top left corner, crop an average intensity image sample block of all its adjacent pixels. - ; S7: Average intensity image sample block - With panchromatic image sample blocks Spatial location matching to the same reference, calculating average intensity image sample blocks - With panchromatic image sample blocks The correlation coefficient; S8: Determine the average intensity image sample block With panchromatic image sample blocks If the correlation coefficient is the largest, then determine ( , If the registration result is accurate, proceed to S13; otherwise, proceed to S9. S9: Extract a sample block from the panchromatic image. The image sample block with the highest correlation coefficient is the first one. The result of the second registration is denoted as the top left corner point ( ). , );in, The number of registrations; with ( , Using as the center, determine the surrounding independent pixels; S10: With the center point ( , Using the top-left corner point as the unit and the independent pixel point, average intensity image sample blocks are sequentially cropped according to a preset fixed window size, which can be denoted as... - ; S11: Average intensity image sample block - With panchromatic image sample blocks Spatial locations are matched to the same reference, and the average intensity of the image sample block is calculated. - With panchromatic image sample blocks The correlation coefficient; S12: Determine ( , The average intensity image sample block where ) is located With panchromatic image sample blocks If the correlation coefficient is the maximum, then determine the final registration scheme and execute S13; otherwise, let and return to S9; S13: Based on the final accurate registration scheme of multispectral and panchromatic images, and using the high-resolution panchromatic image as a reference, the multispectral image is accurately registered, that is, the multispectral image coordinates are translated ( - , ( ) pixels, to save the registered image results.

2. The sub-pixel-level image registration method according to claim 1, characterized in that, Low-resolution panchromatic image simulation using the average intensity of multispectral images is expressed as follows: ； in, It is a low-resolution panchromatic image. For multispectral images, N is the number of bands in the multispectral image.

3. The sub-pixel-level image registration method according to claim 2, characterized in that, The average intensity image sample block - With panchromatic image sample blocks The formula for calculating the correlation coefficient is: ； in, For average intensity image sample blocks With panchromatic image sample blocks The correlation coefficient, For average intensity image sample blocks With panchromatic image sample blocks covariance, For average intensity image sample blocks standard deviation For panchromatic image sample blocks The standard deviation.

4. The sub-pixel-level image registration method according to claim 3, characterized in that, The determination of independent pixel points in S9 is specifically as follows: by( , Centered on a point, there are 8 adjacent pixels around it. Remove pixels that are adjacent to the point ( ). , When the center point is an adjacent pixel point that overlaps, the remaining pixel point is treated as an independent pixel point.

5. A sub-pixel-level image registration system, characterized in that, The system includes a memory and a processor. The memory contains a sub-pixel-level image registration method program, which, when executed by the processor, performs the following steps: S1: Input remote sensing images, including panchromatic images and multispectral images; S2: Perform initial geometric corrections on the high-resolution panchromatic image and the low-resolution multispectral image respectively, find corresponding ground features, and confirm that the spatial positions of the panchromatic image and the multispectral image are consistent. S3: Upsample the multispectral image to the same resolution as the panchromatic image; S4: Low-resolution panchromatic image simulation using the average intensity of multispectral images; S5: Extracting pixel points from a low-resolution panchromatic image ( , Using the top-left corner as the reference point, average intensity image sample blocks are cropped according to a preset fixed window. With panchromatic image sample blocks ; S6: Image sample blocks with average intensity Top left corner ( , Using a fixed window as the center, determine the eight surrounding pixels, and then, taking that pixel as the top left corner, crop an average intensity image sample block of all its adjacent pixels. - ; S7: Average intensity image sample block - With panchromatic image sample blocks Spatial location matching to the same reference, calculating average intensity image sample blocks - With panchromatic image sample blocks The correlation coefficient; S8: Determine the average intensity image sample block With panchromatic image sample blocks If the correlation coefficient is the largest, then determine ( , To ensure accurate registration, execute S13; If not, then execute S9; S9: Extract a sample block from the panchromatic image. The image sample block with the highest correlation coefficient is the first one. The result of the second registration is recorded as the pixel point in the upper left corner ( ). , );in, The number of registrations; with ( , Using as the center, determine the surrounding independent pixels; S10: With the center point ( , Using individual pixels as the top-left corner points, average intensity image sample blocks are sequentially cropped according to a preset fixed window size. - ; S11: Average intensity image sample block - With panchromatic image sample blocks Spatial location matching to the same reference, calculating average intensity image sample blocks - With panchromatic image sample blocks The correlation coefficient; S12: Determine the center point ( , The average intensity image sample block and the panchromatic image sample block where ) are located If the correlation coefficient is the largest, then execute S13; If not, then let and return to S9; S13: Based on the final accurate registration scheme of multispectral and panchromatic images, and using the high-resolution panchromatic image as a reference, the multispectral image is accurately registered, that is, the multispectral image coordinates are translated ( - , ( ) pixels, to save the registered image results.

6. The sub-pixel-level image registration system according to claim 5, characterized in that, Low-resolution panchromatic image simulation using the average intensity of multispectral images is expressed as follows: ； in, It is a low-resolution panchromatic image. For multispectral images, N is the total number of bands in the multispectral image.

7. The sub-pixel-level image registration method according to claim 6, characterized in that, The average intensity image sample block - With panchromatic image sample blocks The formula for calculating the correlation coefficient is: ； in, For average intensity image sample blocks With panchromatic image sample blocks The correlation coefficient, For average intensity image sample blocks With panchromatic image sample blocks covariance, For average intensity image sample blocks standard deviation For panchromatic image sample blocks The standard deviation.

8. The sub-pixel-level image registration method according to claim 6, characterized in that, The determination of independent pixels in S9 is specifically as follows: by( , Centered on a point, there are 8 adjacent pixels around it. Remove pixels that are adjacent to the point ( ). , When the center point is an adjacent pixel point that overlaps, the remaining pixel point is treated as an independent pixel point.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a sub-pixel-level image registration method program, which, when executed by a processor, implements the steps of a sub-pixel-level image registration method as described in any one of claims 1 to 4.

Images