A high-resolution image color uniformity method and device based on reflectance bottom map denoising

Abstract

This invention discloses a high-resolution image color balancing method and apparatus based on reflectance basemap denoising, belonging to the field of remote sensing image processing technology. The method includes: firstly, generating a white spot noise mask; then, coupling temporal auxiliary observation information with an adaptive local regression model for multi-round iterative repair; and finally, combining a fast-moving method to fill in residual isolated noise and reconstruct a high-quality reference basemap; and finally, through a high-low frequency separation strategy, fusing the real low-frequency tone constraints of the reconstructed basemap with the high-frequency texture details of the high-resolution source image to achieve macroscopic color smoothing and high-precision fidelity of microscopic texture. This invention can effectively eliminate tone distortion and spectral distortion caused by basemap noise, providing technical support for the high-quality production of large-scale high-resolution remote sensing analysis-ready data.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, specifically relating to a method and apparatus for color balancing of high-resolution images based on reflectance base map denoising. Background Technology

[0002] In recent years, high-resolution remote sensing satellites have played an increasingly important role in land surveying, agricultural yield estimation, forestry surveys, and natural resource monitoring, thanks to their rich spatial detail information. With the deepening development of quantitative remote sensing, the concept of Analysis Ready Data (ARD) has become an industry consensus. ARD requires remote sensing data to undergo rigorous standardized preprocessing, including radiometric calibration, atmospheric correction, and geometric correction, enabling users to directly conduct time-series analysis and quantitative inversion of surface parameters. However, high-resolution sensors generally have narrow observation widths. When conducting large-scale surface cover monitoring at the provincial, national, or even global scale, it is usually necessary to mosaic multiple remote sensing images acquired from different orbits and on different dates. Due to significant differences in solar altitude angle, atmospheric environment, seasonal phase, and sensor observation angle at the time of acquisition, the mosaicked images often exhibit drastic tonal differences and radiometric inconsistencies in adjacent areas, severely restricting their quantitative applications and interpretation.

[0003] To eliminate radiometric differences between multiple high-resolution images, color balancing techniques have become a crucial step in large-scale image mosaicking. Currently, the mainstream and effective approach is to introduce low-to-medium resolution satellite imagery with global radiometric consistency (such as the MODIS surface reflectance product MOD09GA) as an external absolute reference base map to constrain the macroscopic tonal range of the high-resolution images. However, this approach faces a significant drawback in practical applications: noise inherent in the base map itself can lead to radiometric errors. Although MOD09GA offers wide coverage, temporal continuity, and high radiometric consistency, under actual observation conditions, factors such as atmospheric disturbances and sensor noise can cause abnormally bright pixels (white spot noise, such as...) to appear in cloudless areas. Figure 1 (As shown). Existing conventional color balancing processes typically lack effective mechanisms for identifying and thoroughly repairing these anomalous reference pixels. If a noisy base map is directly used as a color reference, these anomalous high-radiative features will be forcibly and incorrectly transmitted to the target high-resolution image during the color balancing process. This results in local radiometric distortions such as anomalous bright patches and abrupt color changes in the post-color-balancing image, severely damaging the true spectral characteristics of dark-toned features (such as vegetation) and rendering the processed image worthless for quantitative analysis.

[0004] Therefore, how to effectively eliminate erroneous radiometric mapping caused by noise in the reference base map, and achieve high-precision spectral and texture fidelity while ensuring global color consistency of high-resolution images, has become a key problem that urgently needs to be solved in the preparation of large-scale high-resolution images for color uniformity and high-quality ARD. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a high-resolution image color balancing method and apparatus based on reflectance base map denoising. The method uses white spot noise identification and multi-round temporal iterative repair of the reference base map as a pre-constraint for high-resolution image color balancing, and combines the low-frequency tones of the reconstructed base map with the high-frequency textures of the source image for fusion and fidelity preservation.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A high-resolution image color balancing method based on reflectance undermap denoising, the method comprising:

[0008] Step 1: Convert the format and extract the bands of the reference base map product to serve as the reference image for subsequent processing; preprocess the high-resolution source image to obtain the processed high-resolution source image.

[0009] Step 2: Analyze the quality labeling data of the reference image, generate a binary quality mask through bit operations, and mark the pixels as white spot noise regions and effective observation regions;

[0010] Step 3: Using the pixels in the effective observation area as samples, the white spot noise in the white spot noise area is repaired in multiple rounds through adaptive window and local linear regression model. After each round of repair, the successfully repaired pixels are updated to the effective observation area.

[0011] Step 4: Using the repaired or originally valid pixel information in the effective observation area, fill in the residual white spot noise in the white spot noise area that has not been repaired, and reconstruct a high-quality reference image base map.

[0012] Step 5: Superimpose the high-frequency texture information of the processed high-resolution source image onto the low-frequency tonal component of the reconstructed high-quality reference image base to obtain a uniform color result image.

[0013] Furthermore, the preprocessing of the high-resolution source image in step 1 includes: radiometric calibration of the original digital quantization values ​​of the input high-resolution source image to convert them into physically meaningful apparent radiance; then, using an automatic atmospheric correction algorithm to eliminate the effects of atmospheric scattering and absorption to obtain a true surface reflectance image; and finally, obtaining a high spatial resolution multi-band orthorectified image through orthorectification.

[0014] Furthermore, step 2 includes: resampling the quality labeling data with a resolution lower than a preset threshold using the nearest neighbor interpolation method to match its spatial resolution with the surface reflectance data; parsing the resampled binary status code through bit operations to extract abnormal white spot noise pixels and construct an abnormal status set; traversing the entire image matrix to generate a binary quality mask, marking the pixels as white spot noise areas to be repaired and effective observation areas.

[0015] Furthermore, the process of determining the adaptive window in step 3 includes: performing Euclidean distance transformation on the generated binarized quality mask, calculating the spatial distance from each white spot noise pixel to be repaired to its nearest effective observation pixel; extracting the mean and maximum values ​​of non-zero values ​​in the distance field, deriving the initial window size based on the sum of the mean and a preset initial expansion parameter, and deriving the maximum window size based on the sum of the maximum value and a preset safety buffer parameter.

[0016] Furthermore, the establishment of the local linear regression model and the multi-round iterative repair process in step 3 includes: extracting bidirectional effective observation samples between the target temporal reference image to be repaired and the auxiliary temporal image within the current adaptive window; constructing an objective function based on the least squares method and solving for the optimal local radiation gain coefficient and local radiation bias; using the solved optimal local radiation gain coefficient and local radiation bias, performing a linear transformation on the pixel values ​​at the corresponding positions in the auxiliary temporal image to obtain the predicted regression value of the white spot noise pixel to be repaired in the target temporal reference image; after each round of repair, updating the marker of the successfully repaired pixel in the binarized quality mask from the white spot noise region to the effective observation region for the next round of iteration.

[0017] Furthermore, in step 4, pixels that are still marked as white spot noise after multiple rounds of iterative repair are identified as residual white spot pixels. A fast-moving method is used to fill in the space using information from surrounding repaired or originally valid pixels.

[0018] Furthermore, step 5 includes: extracting low-frequency information from the processed high-resolution source image and the reconstructed high-quality reference image base map to obtain their respective low-frequency components; constructing a local linear model based on a block-based computing strategy to derive the gain coefficient of the local region; subtracting the low-frequency component from the processed high-resolution source image to extract high-frequency texture information; multiplying the extracted high-frequency texture information by the gain coefficient of the local region and adding it to the low-frequency component of the reconstructed high-quality reference image base map to generate a uniform color result image.

[0019] On the other hand, the present invention provides a high-resolution image color balancing device based on reflectance undermap denoising, comprising:

[0020] The processing module is used to perform format conversion and band extraction on the reference base map product, which serves as a reference image for subsequent processing; and to preprocess the high-resolution source image to obtain the processed high-resolution source image.

[0021] The identification module is used to parse the quality identification data of the reference image, generate a binary quality mask through bit operations, and mark the pixels as white spot noise areas and effective observation areas;

[0022] The repair module is used to use the pixels in the effective observation area as samples to perform multiple rounds of iterative repair on the white spot noise in the white spot noise area through an adaptive window and a local linear regression model. After each round of repair, the successfully repaired pixels are updated to the effective observation area.

[0023] The reconstruction module is used to fill in the residual white spot noise in the white spot noise area that has not been repaired by utilizing the repaired or originally valid pixel information in the effective observation area, and reconstruct a high-quality reference image base map.

[0024] The overlay module is used to overlay the high-frequency texture information of the processed high-resolution source image onto the low-frequency tonal components of the reconstructed high-quality reference image base map to obtain a uniform color result image.

[0025] Thirdly, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned high-resolution image color balancing method based on reflectance undermap denoising.

[0026] Fourthly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned high-resolution image color balancing method based on reflectivity undermap denoising.

[0027] The beneficial effects of this invention are as follows:

[0028] This invention fundamentally eliminates radiation distortion caused by base map noise: It uses reference base map denoising as a key pre-color homogenization step. Through quality identifier analysis, multi-temporal adaptive local regression iterative repair, and fast-moving spatial interpolation, it effectively identifies and repairs white spot noise in reference base maps such as MOD09GA. Compared to existing technologies, it avoids abnormally bright pixels being incorrectly mapped to high-resolution images and significantly suppresses artificially high reflectance and local tone abrupt changes in the visible light band.

[0029] Achieving macroscopic color smoothing and high-precision fidelity of microscopic texture: Based on obtaining a high-quality reconstructed base map, this invention employs a high-low frequency separation strategy to fuse reliable low-frequency tonal components with the original high-frequency texture information of the high-resolution source image. Compared to traditional color balancing methods, this effectively corrects macroscopic color differences between multi-temporal and multi-track images while fully preserving the spatial edges and detailed textures of ground features, avoiding blurring or loss of details.

[0030] Ensuring Quantitative Application and High-Quality ARD Production: This invention guarantees the physical authenticity and spatial continuity of surface reflectance data from the source, providing reliable technical support for the high-quality, high-precision, large-scale production of ARD data ready for large-scale high-resolution satellite remote sensing analysis. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of white spot noise in the prior art;

[0032] Figure 2 This is a flowchart of a high-resolution image color balancing method based on reflectance undermap denoising according to the present invention;

[0033] Figure 3 The images show a comparison of the effects of removing white spot noise from MOD09GA images before and after the present invention. (a) and (d) are different original MOD09GA images, (b) and (e) are corresponding white spot noise masks, and (c) and (f) are the resulting images after white spot noise removal.

[0034] Figure 4 This is a comparison of the mosaic results before and after color balancing processing of high-resolution satellite images according to the present invention. (a) and (c) are mosaic images of different original GF-1 images, and (b) and (d) are mosaic images of the corresponding color balancing results. Detailed Implementation

[0035] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0036] This invention provides a high-resolution image color balancing method and apparatus based on reflectance basemap denoising. It uses benchmark image quality optimization as a prerequisite constraint for high-resolution image color balancing, combining low-frequency tone constraints from the reconstructed basemap with high-frequency texture preservation from the source image. This achieves high-precision fidelity preservation of micro-textures while eliminating temporal and observational color differences in large-scale images. This method ensures the physical authenticity and spatial continuity of surface reflectance data from the source, providing solid and reliable technical support for the high-quality, high-precision, large-scale production of large-scale high-resolution satellite remote sensing analysis-ready data (ARD). Specifically, for example... Figure 2 As shown, the method includes:

[0037] Step 1: Multi-source data preprocessing: The reference base map product is converted in format and its bands are extracted to serve as the reference image for subsequent processing; the high-resolution source images are then subjected to radiometric calibration, atmospheric correction and orthorectification in sequence to obtain the processed high-resolution source images;

[0038] Step 2, White Spot Noise Identification and Mask Generation of Reference Image: Analyze the quality identification data of the reference image, generate a binary quality mask through bit operations, and mark the pixels as white spot noise areas and effective observation areas;

[0039] Step 3: Multi-round iterative repair of the base map time series: Using the pixels in the effective observation area as samples, the white spot noise in the white spot noise area is repaired in multiple rounds through adaptive window and local linear regression model. After each round of repair, the successfully repaired pixels are updated to the effective observation area.

[0040] Step 4, Spatial interpolation of residual white spot noise: Using the repaired or originally valid pixel information in the effective observation area, fill in the residual white spot noise in the white spot noise area that has not been repaired, and reconstruct a high-quality reference image base map.

[0041] Step 5: Color balancing based on high and low frequency separation: The high-frequency texture information of the processed high-resolution source image is superimposed on the low-frequency tonal component of the reconstructed high-quality reference image base to obtain the color balancing result image.

[0042] Furthermore, in step 1:

[0043] The reference base map product covering the study area (MOD09GA selected here) was converted to a unified data format using a projection conversion tool, and the visible and near-infrared bands corresponding to the high-resolution image bands were extracted and synthesized into a medium-resolution reference image. The raw digital quantization values ​​of the input high-resolution source image (GF-1 wide-swath image) were radiometrically calibrated to convert them into physically meaningful apparent radiance. Subsequently, an automatic atmospheric correction algorithm (QUAAC algorithm, FLAASH, 6S radiative transfer model) was used to eliminate the effects of atmospheric scattering and absorption, obtaining a true surface reflectance image. Finally, orthorectification was performed to obtain a high spatial resolution multi-band orthorectified image, which served as the processed high-resolution source image.

[0044] Furthermore, in step 2:

[0045] By analyzing the quality label data provided with the MOD09GA product after processing, anomalous pixels affected by atmospheric or noise pollution are located. Nearest neighbor interpolation is used to resample the quality label data with a resolution below a preset threshold, matching its spatial resolution to the surface reflectance data while preserving the physical meaning of the discrete state labels. Bitwise operations are used to parse the resampled binary state code, extracting anomalous white spot noise pixels and constructing an anomalous state set. A binary quality mask is generated by traversing the entire image matrix, marking pixels as the white spot noise region to be repaired and the effective observation region, respectively.

[0046] ,

[0047] In the formula, The spatial coordinates after resampling are The pixel status code, and These represent the column coordinates and row coordinates of a pixel, respectively. This represents the noisy area of ​​vitiligo that needs to be repaired. This represents the effective observation area.

[0048] Furthermore, in step 3:

[0049] Using effective observations from auxiliary temporal image sequences, dynamic inpainting noise is performed through an adaptive spatial window and a local linear regression model. For any pixel in the target image to be repaired, the candidate temporal image sequences are searched sequentially from most recent to furthest to find the best auxiliary temporal image at that spatial location that is not contaminated by noise. A Euclidean distance transformation is performed on the generated white spot noise mask to calculate the spatial distance D from each white spot noise pixel to its nearest effective pixel.

[0050] ,

[0051] In the formula, The spatial coordinates of the white patches to be repaired; The spatial coordinates of the effective pixels; This is the set of valid observed pixels in the current target image. The mean of the non-zero values ​​in the range field is extracted. With the maximum value Derive the initial window size With the largest window :

[0052] ,

[0053] ,

[0054] In the formula, and These are the initial expansion parameter and the safety buffer parameter, respectively; the lower limit of the initial window is set to 15 pixels to meet the minimum degrees of freedom required for local regression statistical analysis. The spatial autocorrelation boundary limits the maximum expansion range of the window.

[0055] For each vitiligo pixel to be repaired, firstly... The target temporal reference image and auxiliary temporal image are extracted using the current window size. When the number of bidirectional effective samples is less than the preset sample window threshold, the initial expansion parameter is used. Increase the current window size and recount the number of valid two-way samples; stop expanding and establish a local linear regression model when the number of valid two-way samples reaches a threshold; the current window reaches... If the number of valid samples in both directions is still less than the threshold, it is determined that the pixel cannot be restored under the current auxiliary temporal image, and the next auxiliary temporal image will be investigated.

[0056] Within a defined adaptive window, extract the temporal reference image of the target to be repaired. With auxiliary temporal images Effective pixel samples between the two images. The radiance of the two images within the local window has a linear mapping relationship. The objective function E is constructed based on the least squares method:

[0057] ,

[0058] In the formula, This represents the total number of valid bidirectional samples within the window. For the first in the local window The spatial coordinates of each valid sample; and These are the optimal local radiation gain coefficient and optimal local radiation bias obtained by solving the least squares method, respectively:

[0059] ,

[0060] ,

[0061] In the formula, and Representing the local window respectively The mean local radiometric values ​​of two-way valid samples in the reference and auxiliary images. If the number of valid samples within the window meets the statistical significance threshold, the established local linear model is used to predict the vitiligo pixels to be repaired:

[0062] ,

[0063] In the formula, The predicted regression value is the pixel to be repaired in the target temporal reference image.

[0064] To maximize the use of restored surface information, an iterative mechanism for dynamic mask updates is introduced. The mask state of pixels successfully restored in the current round... It is immediately updated from 1 to 0 and participates in the establishment of the local linear regression model as an effective prior pixel in the next round of regression for adjacent continuous white spot noise.

[0065] After processing all white spot noise pixels to be repaired in the current round, the number of newly successfully repaired pixels in this round is counted. If no new successfully repaired pixels are found in this round and there are no more pixels with a state of 1 in the quality mask, the multi-round iterative repair is terminated; otherwise, the next round of iteration begins based on the updated quality mask. After the iteration terminates, the pixels with a state of 1 in the quality mask constitute the residual white spot noise set.

[0066] Furthermore, in step 4:

[0067] After multiple rounds of temporal iterative restoration, pixels whose state is still 1 in the quality mask at the termination time are identified as residual white spots. These white spots may not be able to establish an effective regression model because their candidate temporal sequences are all contaminated. The Fast Marching Method (FMM) is introduced to perform smooth interpolation in the pure spatial domain (Kriging, Inverse Distance Weighting (IDW), or image inpainting algorithms based on partial differential equations (PDEs) can also be used). Utilizing the restored or originally effective real pixel information around the residual white spots, numerical filling is performed by pushing inward, ultimately reconstructing a high-quality reference image base map with continuous radiometric features and no obvious white spot noise.

[0068] Figure 3 These are examples of noise removal results for two typical local areas in a MOD09GA image. Figure 3 Images (a)-(c) demonstrate the restoration performance against a complex mountainous background. In image (a), the topographic relief and vegetation features of the area before processing are largely obscured by white spot noise. After processing (image (b) shows the white spot noise mask, and image (c) shows the white spot noise removal results), the spatial texture of the mountains and the original radiation characteristics of the vegetation are restored, and the outlines of the land features are clearly distinguishable. Figure 3 Images (d)-(f) demonstrate the treatment effect in the water-land interface area. In particular, (d) shows the strong stretching interference of abnormally bright noise on the low background reflectivity of the water body before restoration; after restoration ((e) shows the white spot noise mask, and (f) shows the white spot noise removal result), the abnormally bright noise is effectively suppressed, and the radiation state and geometric boundary of the water body and coastal features are reconstructed.

[0069] Furthermore, in step 5:

[0070] By employing a frequency domain separation strategy, the macroscopic tonal information of the reconstructed high-quality reference basemap is fused with the microscopic detail information of the high-resolution source image. From a spectral perspective, the image is decomposed into high-frequency information representing microscopic spatial edges and textures, and low-frequency information representing macroscopic color distribution trends.

[0071] ,

[0072] In the formula, The image to be separated in the frequency domain is either a processed high-resolution source image or a reconstructed high-quality reference image base map obtained in step 4. and These represent the low-frequency and high-frequency information components of the corresponding image, respectively.

[0073] Gaussian low-pass filtering (mean filtering or wavelet transform can also be used to extract high and low frequency components of the image) and downsampling are applied to the preprocessed high-resolution source image and the restored reference base map, respectively, to extract their respective low-frequency components. The downsampled source image and the reference color base map of the corresponding region can be represented as follows:

[0074] ,

[0075] ,

[0076] In the formula, This represents the high-resolution source image after downsampling. and These represent the low-frequency and high-frequency information components of the sampled high-resolution source image, respectively. This refers to a reference image after vitiligo repair. and These represent the low-frequency and high-frequency information components of the reference base image after vitiligo repair, respectively.

[0077] Extracted low-frequency information components and Upsampling to the original image size is performed using bilinear interpolation. A large-interval block averaging sampling method is employed, resulting in a smoother interpolated image with good fit to the source image. These two images can be considered to represent the overall low-frequency information of the source image and the target reference image, respectively, and are denoted as follows: and .

[0078] A block-based computation strategy is employed to construct local, independent linear models, deriving the optimal radiometric gain coefficient for each local region. In relative radiometric correction, it is typically assumed that the pixel brightness across images from different time phases follows a linear relationship.

[0079] ,

[0080] In the formula, and These represent the target image and the source image respectively, located in the source image. Pixel value at; and These are the row and column numbers of the pixel, respectively. and The parameters in the linear model represent the gain coefficient and the offset, respectively.

[0081] Specifically, for each local image block with a width of w pixels and a height of h pixels, it is assumed that all pixels within the block satisfy a linear model:

[0082] ,

[0083] ,

[0084] In the formula, and These represent the average brightness of image blocks in the target image and the source image, respectively. When... Sometimes, ,Will Substituting, we can obtain Therefore, for low reflectivity regions (such as dark pixels), the following condition must be met:

[0085] ,

[0086] because It was obtained by downsampling high-resolution source images. The gain coefficient in the linear model is obtained by downsampling the target image. It can be done and The corresponding pixel is calculated, and its gain coefficient mapping is obtained through bilinear interpolation upsampling, denoted as . The derivation leads to:

[0087] ,

[0088] Compared to In other words, It is a very small quantity, so we get the equation:

[0089] ,

[0090] Finally, the color-balanced image can be obtained through calculation. The resulting image consists of two parts: high-frequency information from the stretched source image. Low-frequency information of the color-corrected reference image While effectively utilizing the base map to correct macroscopic color deviations, it avoids the loss of any high-frequency details, achieving the dual goals of color smoothness and texture fidelity.

[0091] Figure 4 This is a comparison of the mosaic effect before and after color uniformity in two sets of typical color difference images. The original information of the experimental images is shown in Table 1.

[0092] Table 1

[0093]

[0094] The inconsistencies in the images from the first set of experiments were mainly due to differences in the field of view and load of the different wide-angle cameras mounted on the GF-1, such as... Figure 4 In (a), the left half of the first image is greenish and the right half of the second image is purple. After processing by the method of this invention, the overall color tone area is consistent, as shown in the figure. Figure 4 As shown in (b), there is no obvious tonal difference; the tonal difference in the images of the second group of experiments is due to seasonal factors, such as... Figure 4 (c) The left half of the third image was taken in April when vegetation was lush, and the overall color is dark green. The right half of the fourth image was taken in September after the crops were harvested, and the overall color is brown. The two images show obvious differences in land cover types. After processing with the method of this invention, the overall color tone tends to be consistent. Figure 4 As shown in (d), the image on the right transitions naturally from the original brown to a greenish hue.

[0095] On the other hand, the present invention provides a high-resolution image color balancing device based on reflectance undermap denoising, which includes modules capable of implementing the steps of the aforementioned method, specifically including:

[0096] The processing module is used to perform format conversion and band extraction on the reference base map product, which serves as a reference image for subsequent processing; and to preprocess the high-resolution source image to obtain the processed high-resolution source image.

[0097] The identification module is used to parse the quality identification data of the reference image, generate a binary quality mask through bit operations, and mark the pixels as white spot noise areas and effective observation areas;

[0098] The repair module is used to use the pixels in the effective observation area as samples to perform multiple rounds of iterative repair on the white spot noise in the white spot noise area through an adaptive window and a local linear regression model. After each round of repair, the successfully repaired pixels are updated to the effective observation area.

[0099] The reconstruction module is used to fill in the residual white spot noise in the white spot noise area that has not been repaired by utilizing the repaired or originally valid pixel information in the effective observation area, and reconstruct a high-quality reference image base map.

[0100] The overlay module is used to overlay the high-frequency texture information of the processed high-resolution source image onto the low-frequency tonal components of the reconstructed high-quality reference image base map to obtain a uniform color result image.

[0101] Thirdly, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned high-resolution image color balancing method based on reflectance undermap denoising.

[0102] Fourthly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned high-resolution image color balancing method based on reflectivity undermap denoising.

[0103] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A high-resolution image color uniformization method based on reflectance map denoising, characterized in that, The method includes: Step 1: Convert the format and extract the bands of the reference base map product to serve as the reference image for subsequent processing; preprocess the high-resolution source image to obtain the processed high-resolution source image. Step 2: Analyze the quality labeling data of the reference image, generate a binary quality mask through bit operations, and mark the pixels as white spot noise areas and effective observation areas; Step 3: Using the pixels in the effective observation area as samples, the white spot noise in the white spot noise area is repaired in multiple rounds through adaptive window and local linear regression model. After each round of repair, the successfully repaired pixels are updated to the effective observation area. Step 4: Using the repaired or originally valid pixel information in the effective observation area, fill in the residual white spot noise in the white spot noise area that has not been repaired, and reconstruct a high-quality reference image base map. Step 5: Superimpose the high-frequency texture information of the processed high-resolution source image onto the low-frequency tonal component of the reconstructed high-quality reference image base to obtain a uniform color result image.

2. The method according to claim 1, wherein, The preprocessing of the high-resolution source image in step 1 includes: radiometric calibration of the raw digital quantization values ​​of the input high-resolution source image to convert them into physically meaningful apparent radiance; then, using an automatic atmospheric correction algorithm to eliminate the effects of atmospheric scattering and absorption to obtain a true surface reflectance image; and finally, obtaining a high spatial resolution multi-band orthorectified image through orthorectification.

3. The method of claim 1, wherein the method further comprises: Step 2 includes: resampling the quality labeling data with a resolution lower than a preset threshold using the nearest neighbor interpolation method to match its spatial resolution with the surface reflectance data; parsing the resampled binary status code through bit operations to extract abnormal white spot noise pixels and construct an abnormal status set; traversing the entire image matrix to generate a binary quality mask and marking the pixels as white spot noise areas to be repaired and effective observation areas.

4. The method of claim 1, wherein the method further comprises: The process of determining the adaptive window in step 3 includes: performing Euclidean distance transformation on the generated binarized quality mask, calculating the spatial distance from each white spot noise pixel to be repaired to its nearest effective observation pixel; extracting the mean and maximum values ​​of non-zero values ​​in the distance field, deriving the initial window size based on the sum of the mean and the preset initial expansion parameter, and deriving the maximum window size based on the sum of the maximum value and the preset safety buffer parameter.

5. The method of claim 1, wherein the method further comprises: The establishment of the local linear regression model and the multi-round iterative repair process in step 3 include: extracting bidirectional effective observation samples between the target temporal reference image to be repaired and the auxiliary temporal image within the current adaptive window; constructing an objective function based on the least squares method and solving for the optimal local radiation gain coefficient and local radiation bias; using the solved optimal local radiation gain coefficient and local radiation bias, performing a linear transformation on the pixel values ​​at the corresponding positions in the auxiliary temporal image to obtain the predicted regression value of the white spot noise pixel to be repaired in the target temporal reference image; after each round of repair, updating the marker of the successfully repaired pixel in the binarized quality mask from the white spot noise region to the effective observation region for the next round of iteration.

6. The high-resolution image color balancing method based on reflectance undermap denoising according to claim 1, characterized in that, In step 4, pixels that are still marked as white spot noise after multiple rounds of iterative repair are identified as residual white spot pixels. A fast-moving method is used to fill in the spatial gaps by using information from surrounding repaired or originally valid pixels.

7. The method of claim 1, wherein the method further comprises: Step 5 includes: extracting low-frequency information from the processed high-resolution source image and the reconstructed high-quality reference image base map to obtain their respective low-frequency components; constructing a local linear model based on a block-based computing strategy to derive the gain coefficient of the local region; subtracting the low-frequency component from the processed high-resolution source image to extract high-frequency texture information; multiplying the extracted high-frequency texture information by the gain coefficient of the local region and adding it to the low-frequency component of the reconstructed high-quality reference image base map to generate a uniform color result image.

8. A high-resolution image color uniformity device based on reflectance map denoising, characterized in that, include: The processing module is used to perform format conversion and band extraction on the reference base map product, which serves as a reference image for subsequent processing. Preprocessing is performed on the high-resolution source image to obtain the processed high-resolution source image; The identification module is used to parse the quality identification data of the reference image, generate a binary quality mask through bit operations, and mark the pixels as white spot noise areas and effective observation areas; The repair module is used to use the pixels in the effective observation area as samples to perform multiple rounds of iterative repair on the white spot noise in the white spot noise area through an adaptive window and a local linear regression model. After each round of repair, the successfully repaired pixels are updated to the effective observation area. The reconstruction module is used to fill in the residual white spot noise in the white spot noise area that has not been repaired by utilizing the repaired or originally valid pixel information in the effective observation area, and reconstruct a high-quality reference image base map. The overlay module is used to overlay the high-frequency texture information of the processed high-resolution source image onto the low-frequency tonal components of the reconstructed high-quality reference image base map to obtain a uniform color result image.

9. An electronic device, comprising: include: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the high-resolution image color balancing method based on reflectance undermap denoising as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, enable the processor to implement the high-resolution image color balancing method based on reflectivity undermap denoising as described in any one of claims 1-7.

Images