A data processing method, system, device and medium of a linearly variable filter imaging spectrometer

Abstract

The application discloses a data processing method, system, device and medium of a linearly variable filter imaging spectrometer. The method comprises the following steps: performing geometric correction on an original image, generating a corrected image and recording a waveband row number, and generating a waveband index file for positioning irregularly arranged waveband pixels in the corrected image; extracting the same waveband data from the corrected image and the waveband index file frame by frame from a plurality of continuous corrected images to splice a plurality of single-waveband images; performing weighted fusion of the plurality of single-waveband images based on texture differences to obtain a synthetic waveband image; and performing multi-scale step-by-step registration of the synthetic waveband image based on an image pyramid, combining global coarse registration and local fine adjustment, and realizing accurate registration between wavebands. The application solves the problems of complicated waveband extraction, missing single-waveband data, low signal-to-noise ratio and difficult accurate registration of complex deformation between wavebands in data processing of the linearly variable filter imaging spectrometer, and realizes efficient and high-precision spectral cube reconstruction.

Description

Technical Field

[0001] This invention relates to the field of remote sensing data processing technology, and specifically to a data processing method, system, computer equipment, and computer-readable storage medium for a linear gradient filter imaging spectrometer. Background Technology

[0002] Linear Variable Filter (LVF) imaging spectrometers use wedge-shaped coated bandpass filters as the dispersive element, with the center wavelength of the passband varying linearly along a specific direction of the filter. This type of spectrometer has a simple optical structure, is easy to miniaturize and reduce weight, and has promising applications in fields such as UAV hyperspectral remote sensing.

[0003] During data acquisition, the LVF imaging spectrometer employs pushbroom imaging, imaging the spectral responses of different bands of the target scene onto different rows of the detector in a time-division manner. Theoretically, to reconstruct the spectral cube of the target scene, it is necessary to first register multiple consecutive frames of images frame by frame, then extract the corresponding image data according to the band rows and stitch them together to form images for each band. However, due to the large number of bands in this type of spectrometer, the computational load for accurate frame-by-frame registration is extremely high, making this theoretical method difficult to meet the needs of practical applications.

[0004] Currently, some research has proposed a data processing workflow for LVF imaging spectrometers. This workflow adopts a "band extraction first, then geometric correction" strategy. The specific steps are: first, extract each single-band image one by one; then, perform geometric correction on each single-band image sequentially; subsequently, perform registration and mosaicking operations on adjacent flight paths; finally, register images of different bands using a feature-based image registration algorithm. This existing technical solution has the following shortcomings:

[0005] First, the extracted single-band images generally suffer from data loss and poor signal-to-noise ratio. The data loss is mainly caused by irregular pixel resampling due to terrain undulation and mismatch between UAV flight speed and camera sampling frequency. No effective processing method has been proposed in the existing technology to address the above-mentioned single-band data quality problems.

[0006] Secondly, the strategy of "extracting bands first and then geometrically correcting" requires performing geometric correction row by row for each band. Since the LVF imaging spectrometer has a large number of bands and different rows of each single-band image correspond to different exterior orientation elements, the entire calculation process is cumbersome and the data processing efficiency is low.

[0007] Furthermore, there are complex deformation differences between images of different bands, including global deviations such as overall translation and rotation, as well as local distortion deviations caused by factors such as terrain undulation. At the same time, there are also nonlinear radiation differences between bands. Existing feature-based registration methods are difficult to adapt to the complex deformation scenarios with both global and local deviations, and cannot achieve accurate registration between bands.

[0008] Therefore, providing a computationally simple data processing method for linear gradient filter imaging spectrometers that can effectively improve the quality of single-band data and achieve accurate registration between bands has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0009] This invention addresses the problems of large computational load, poor single-band data quality, and difficulty in registering complex deformations between bands in the existing data processing workflow of linear graded filter imaging spectrometers. It provides a data processing method, system, computer equipment, and computer-readable storage medium for linear graded filter imaging spectrometers to improve data processing efficiency, improve single-band data quality, and improve registration accuracy between bands.

[0010] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0011] This invention provides a data processing method for a linear graded filter imaging spectrometer, comprising the following steps:

[0012] Geometric correction steps: Perform geometric correction on each frame of the acquired original image to generate a corrected image, and record the band row number of each pixel during the resampling process to generate a band index file for locating irregularly arranged band pixels in the corrected image.

[0013] Band extraction steps: Based on the corrected image and the band index file, extract the same band data frame by frame from multiple consecutive corrected images, and stitch them together to form multiple single-band images;

[0014] Band synthesis step: The multiple single-band images are weighted and fused based on texture differences to obtain a synthesized band image; and

[0015] Band registration steps: Based on the image pyramid, the synthetic band image is registered in a multi-scale stepwise manner, combining global coarse registration and local fine-tuning, to achieve accurate registration between bands.

[0016] The geometric correction step is implemented using the collinear equation indirect method; the nearest neighbor method is used during resampling to prevent spectral aliasing between different bands.

[0017] The band synthesis step specifically includes the following sub-steps:

[0018] Initial fusion sub-step: Perform initial fusion of multiple single-band images by averaging adjacent bands to fill in data gaps in the single-band images;

[0019] Texture detection sub-step: Detect texture features of each single-band image using the Sobel operator;

[0020] Weight calculation sub-step: Using the single-band image corresponding to the center wavelength as the reference band, calculate the texture difference between other single-band images and the reference band, and determine the fusion weight of each single-band image based on the texture difference.

[0021] Weighted fusion sub-step: Weighted fusion of each single-band image is performed according to the fusion weight to obtain a composite band image.

[0022] Preferably, the weight calculation sub-step is performed on a pixel block basis. By calculating the average texture difference within each pixel block and comparing the average value with a multi-level threshold, a corresponding fusion weight is assigned to all pixels within the block. The smaller the texture difference, the higher the corresponding fusion weight.

[0023] In the band registration step, global coarse registration uses the Radiance Invariant Feature Transform (RIFT) algorithm to extract feature point pairs and solve for the similarity transformation matrix to eliminate global displacement deviations between bands. Local fine-tuning includes the following operations: extracting feature points and obtaining corresponding point pairs through template matching; constructing a Delaunay triangulation based on the corresponding point pairs; solving for the local affine transformation matrix for each triangular region; and eliminating local displacement deviations between bands through piecewise linear transformation. Registration parameters are passed step by step according to the inter-layer scaling factor of the image pyramid to achieve step-by-step compensation for displacement deviations.

[0024] The present invention also provides a data processing system for a linear graded filter imaging spectrometer, comprising:

[0025] The geometric correction module is used to perform geometric correction on each frame of the acquired original image, generate the corrected image, and record the band row number of each pixel during the resampling process, generating a band index file for locating irregularly arranged band pixels in the corrected image.

[0026] The band extraction module is used to extract the same band data frame by frame from multiple consecutive corrected images based on the corrected images and the band index file, and stitch them together to form multiple single-band images.

[0027] The band synthesis module is used to perform weighted fusion based on texture differences on the multiple single-band images to obtain a synthesized band image;

[0028] The band registration module is used to perform multi-scale step-by-step registration of the synthetic band image based on the image pyramid, combining global coarse registration with local fine-tuning, to achieve accurate registration between bands.

[0029] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the above-described data processing method for a linear gradient filter imaging spectrometer.

[0030] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the above-described data processing method for a linear gradient filter imaging spectrometer is implemented.

[0031] Compared with the prior art, the present invention has at least the following beneficial effects:

[0032] First, by performing geometric correction on each frame of the original image and simultaneously generating a band index file, and then extracting the same band data frame by frame from multiple corrected images based on this index file, the cumbersome process of "extracting bands first and then performing geometric correction band by band" in existing technologies is avoided. Geometric correction only needs to be performed once on the original frame image, and band extraction can directly locate irregularly arranged band pixels with the help of the index file, without having to perform row-by-row geometric correction for each single band, which significantly reduces computational complexity and improves data processing efficiency.

[0033] Second, after acquiring multiple single-band images, a weighted fusion based on texture differences is further performed. Specifically, an initial fusion using cumulative averaging first fills in the missing data areas in the single-band images. Then, using the single-band image corresponding to the center wavelength as a reference band, the texture differences between other single-band images and the reference band are calculated. The smaller the texture difference, the higher the fusion weight is assigned. This method can effectively suppress the "black line" missing problem caused by irregular pixel resampling or sampling frequency mismatch in single-band images, while improving the signal-to-noise ratio and reducing visual artifacts such as jagged edges and misalignments that may appear in the synthesized band images, ensuring that the quality of the synthesized band images meets the requirements of subsequent quantitative analysis.

[0034] Third, to address the complex deformation differences between synthesized band images, a multi-scale, step-by-step registration strategy based on an image pyramid is adopted. At the small-scale level, the RIFT algorithm is used to extract feature point pairs and solve the similarity transformation matrix to eliminate global displacement bias between bands. At the medium-scale level, based on coarse registration parameters, corresponding point pairs are obtained through template matching and a Delaunay triangulation is constructed. Each triangular region is then subjected to piecewise linear transformation to eliminate local displacement bias. Registration parameters are passed step-by-step according to the pyramid scaling factor, ultimately achieving step-by-step compensation for displacement bias at the original image scale. This strategy can handle both global and local biases simultaneously and is robust to nonlinear radiation differences between bands, making it suitable for accurate band registration in complex application scenarios such as those with large terrain undulations. Attached Figure Description

[0035] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0036] Figure 1 This is an overall flowchart of the data processing method for the linear gradient filter imaging spectrometer in this embodiment of the invention;

[0037] Figure 2 This is a schematic diagram illustrating geometric correction and band index file generation in an embodiment of the present invention;

[0038] Figure 3 This is a schematic diagram illustrating band extraction based on a band index file in an embodiment of the present invention;

[0039] Figure 4 This is a schematic diagram of the phenomenon of missing (black line) single-band image data in an embodiment of the present invention, wherein the right image is a magnified detail of the left image;

[0040] Figure 5 This is a schematic diagram illustrating the band synthesis principle in an embodiment of the present invention;

[0041] Figure 6 This is a schematic diagram of the initial fusion process in an embodiment of the present invention;

[0042] Figure 7 This is a schematic diagram of the weighted fusion process based on texture differences in an embodiment of the present invention;

[0043] Figure 8 This is a schematic diagram of the weight calculation process based on texture differences in an embodiment of the present invention;

[0044] Figure 9 This is a schematic diagram of local details of the synthesized band based on texture difference weighted fusion in an embodiment of the present invention;

[0045] Figure 10A schematic diagram showing local details of the synthesized band obtained through non-averaging operations;

[0046] Figure 11 This is a schematic diagram of the image pyramid structure in an embodiment of the present invention;

[0047] Figure 12 This is a schematic diagram of the band registration technique in an embodiment of the present invention;

[0048] Figure 13 This is a schematic diagram of RIFT feature matching results in an embodiment of the present invention;

[0049] Figure 14 This is a schematic diagram of the feature point matching results in the global coarse registration stage of this invention.

[0050] Figure 15 This is a schematic diagram of the extraction results of corresponding point pairs during the local fine-tuning stage in an embodiment of the present invention;

[0051] Figure 16 This is a schematic diagram of local registration based on Delaunay triangulation in an embodiment of the present invention;

[0052] Figure 17 This is a comparison image of the color composite effect before and after L1 layer band registration in an embodiment of the present invention;

[0053] Figure 18 This is a structural block diagram of the data processing system of the linear gradient filter imaging spectrometer in an embodiment of the present invention. Detailed Implementation

[0054] The technical solutions in the embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The described embodiments are only a part of the implementation of the present invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0055] In practical applications of linear graded-filter (LVF) imaging spectrometers, the data acquisition method results in raw images characterized by "row-by-row time-division imaging and staggered band arrangement": in each frame of the raw image, pixels in different rows correspond to the responses of different rows of the target scene in different bands; push-broom imaging of each single band is achieved through the movement and continuous shooting of the UAV platform, and then complete spectral coverage of the target scene is achieved through continuous multiple frames of images. Therefore, reconstructing the spectral cube of the target scene from the raw images requires completing multiple processing steps in sequence, including geometric correction, extraction and stitching of data in the same band, and correction of inter-band radiation and geometric deformation. Currently, there is no unified standard for the data processing flow of this type of sensor, and there is strong coupling between the various processing steps. Some processing steps (such as band synthesis and band registration) lack mature technical implementation solutions.

[0056] To address the aforementioned technological limitations, the design concept of this invention is as follows: the entire data processing flow is decoupled into four sequentially executed stages: geometric correction, band extraction, band synthesis, and band registration. In the geometric correction stage, a band index file is generated, recording the association between pixel positions and band row numbers, providing a direct location basis for subsequent band extraction without requiring complex band row matching calculations after geometric correction. In the band synthesis stage, texture difference is used as the difference evaluation index between individual bands, assigning corresponding fusion weights to each individual band to ensure the quality of the fused data. In the band registration stage, an image pyramid is constructed to decompose the complex deformations between bands into global and local deviations, which are then processed using registration algorithms adapted to different scales. Registration parameters are passed step-by-step to achieve progressive registration from coarse to fine.

[0057] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0058] Example 1: A data processing method for a linear graded filter imaging spectrometer

[0059] This invention provides a data processing method for a linear graded-filter imaging spectrometer. This method sequentially performs four steps: geometric correction, band extraction, band synthesis, and band registration, ultimately achieving precise registration between bands and supporting quantitative applications of hyperspectral data. The specific process is as follows: First, geometric correction is performed on each frame of the acquired original image. During the correction process, the band row number of each pixel is recorded, and a band index file is generated. Then, based on the corrected image and the band index file, data of the same band is extracted frame by frame from multiple consecutive corrected images, and these frames are stitched together to form multiple single-band images. Next, the multiple single-band images are weighted and fused based on texture differences to obtain a synthesized band image. Finally, based on an image pyramid, multi-scale step-by-step registration combining global coarse registration and local fine-tuning is performed on the synthesized band image to achieve precise registration between bands.

[0060] Figure 1 This is an overall flowchart of the data processing method for the linear graded filter imaging spectrometer in an embodiment of the present invention. See also... Figure 1 The method specifically includes the following steps.

[0061] S1, Geometric correction step: Perform geometric correction on each frame of the acquired original image to generate a corrected image, and record the band row number of each pixel during the resampling process to generate a band index file for locating irregularly arranged band pixels in the corrected image.

[0062] The LVF imaging spectrometer uses a wedge-shaped coated bandpass filter as the spectroscopic element. The center wavelength of its passband varies linearly along a specific direction of the filter. During operation, it employs pushbroom imaging to image the spectral responses of different bands of the target scene onto different rows of the detector in a time-division manner. Therefore, in each frame of the original image, pixels in different rows correspond to responses in different bands. Before geometric correction, the row number of a pixel directly determines its band; however, after geometric correction, the pixel position shifts, and pixels originally in the same row may be scattered to different rows. This causes the correspondence between band information and pixel position to become irregular. If the band row number association information is not recorded, it is impossible to accurately extract data for the same band subsequently.

[0063] In some implementations, the geometric correction step is achieved indirectly using collinear equations, and the nearest neighbor method is used during resampling to prevent spectral aliasing between different bands.

[0064] The collinearity equation is established based on the principle that the original image point, the ground point, and the corrected image point are collinear. It solves for the coordinates of the original image pixel corresponding to the corrected image pixel, which can accurately describe the geometric relationship between the image point, the projection center, and the ground point. Through the solution, the geometric deformation caused by platform attitude changes and terrain undulations can be effectively eliminated. The specific formula is as shown in equation (1):

[0065] (1)

[0066] In equation (1), , , For interior orientation elements, , , For exterior orientation line elements, , , ( ) represents the elements of a rotation matrix composed of external orientation angle elements.

[0067] During resampling, the nearest neighbor method is used, directly taking the gray value of the nearest original pixel without interpolation, in order to avoid mixing of gray values ​​between pixels of different bands, prevent spectral aliasing between different bands, ensure the integrity of band spectral information, and make the geometric accuracy of the corrected image meet the processing requirements of subsequent band extraction, synthesis and registration.

[0068] During the resampling process, the original pixel band row number corresponding to each corrected pixel is recorded synchronously, generating a band index file with the same size as the corrected image. For example... Figure 2 As shown, the value of each pixel in the index file represents the band row number to which that location belongs in the corrected image, thereby associating the irregularly arranged band pixel locations with band information, providing a direct location basis for subsequent band extraction, and solving the problem of disordered and difficult-to-locate band pixels after geometric correction.

[0069] S2, Band extraction step: Based on the corrected image and the band index file, extract the same band data frame by frame from multiple consecutive corrected images, and stitch them together to form multiple single-band images.

[0070] Band extraction involves extracting data of the same band from multiple consecutively captured images and stitching them together to obtain the band image data corresponding to that wavelength (i.e., a single-band image). Based on geometric correction, and by combining the corrected image data with the band row number index file, accurate extraction of a single band can be achieved.

[0071] LVF imaging spectrometers acquire consecutive multi-frame images using pushbroom imaging, covering the entire target scene. For a specific wavelength band, the data is distributed across specific rows in each frame of the corrected image. However, because pixels in the same band are no longer strictly aligned in the same row after geometric correction, direct extraction by row number is impossible. Furthermore, data loss can easily occur during extraction due to factors such as terrain undulations, mismatch between UAV flight speed and camera sampling frequency. A band index file is used to locate irregularly arranged band pixels in the corrected image. With the help of the band index file, the row number of each corrected pixel can be accurately identified, allowing skipping of areas with no values ​​during extraction. This ensures complete ground cover in the stitched single-band image and resolves the pixel arrangement disorder caused by terrain undulations and sampling frequency mismatch.

[0072] like Figure 3As shown, this step utilizes the band index file generated in S1 to traverse each frame of the corrected image and its corresponding band index file. When the value of a pixel in the index file equals the target band row number, the grayscale value of that pixel is extracted and stitched into the target band image according to its geographic coordinates. This process is repeated for each frame to ultimately form a complete single-band image. The extraction operation is then performed on all bands to obtain multiple single-band images. Combining band indexing with band extraction effectively addresses the issue of irregular pixel arrangement after correction, ensuring the integrity of the single-band image.

[0073] It should be noted that during single-band extraction, areas with no values ​​may appear between adjacent rows, i.e., "black lines." This is mainly due to two factors: First, topographical undulations cause the pixel arrangement of the same band in a single frame image to become irregular after image correction, with different areas experiencing varying degrees of shrinkage and stretching, resulting in pixels of the same band not being in the same row after correction (e.g., ...). Figure 3 As shown), during extraction, there may be areas with no values ​​filled between adjacent rows; secondly, the drone's flight speed and the sampling frequency of the hyperspectral camera are mismatched. Ideally, the drone's flight speed is stable and matches the sampling and shooting frequency of the LVF hyperspectral camera, that is, the same ground object moves one line relative to the sensor, and the camera takes one frame. The adjacent images captured in succession differ by only one line of ground area. However, in reality, this condition is difficult to maintain. If the drone's flight speed is faster than the camera's sampling speed, it will result in adjacent images differing by more than one line of ground area. This is reflected in single-band images as multiple "black lines" (such as...). Figure 4 (As shown). The aforementioned data gap issue will be addressed in subsequent band synthesis steps.

[0074] S3, Band synthesis step: The multiple single-band images are weighted and fused based on texture differences to obtain a synthesized band image.

[0075] The extracted single-band images generally suffer from data gaps ("black line" areas) and low signal-to-noise ratios, and the local quality varies between different single bands, making direct stitching insufficient for subsequent quantitative analysis. This step improves the aforementioned data quality issues by performing weighted fusion of multiple single bands based on texture differences, integrating the complementary information of each single band.

[0076] The choice of texture difference as the basis for weighted fusion is mainly based on the deformation differences between bands caused by geometric correction errors: between adjacent single bands or single bands with close wavelength distances, the geometric error is relatively small, and the initial fusion can use the cumulative averaging method to fill in the missing data, and it is not easy to produce "blurry ghosting"; while between single bands with far wavelength distances, the geometric error is large, and if the cumulative averaging fusion is used directly, the synthesized band will have "blurry ghosting" or "jagged misalignment" due to ignoring the geometric differences.

[0077] Based on this, the texture difference between other single bands and the reference band is calculated using the single-band image corresponding to the center wavelength as the reference band. The greater the texture difference, the greater the difference in geometric deformation between the corresponding band and the reference band, and the lower the weight is assigned during fusion. The smaller the texture difference, the higher the weight is assigned, thereby effectively suppressing the above-mentioned synthesis defects.

[0078] like Figure 5 As shown in the diagram (illustration of the band synthesis principle), this step assigns corresponding fusion weights to each single band based on the difference between the texture of each single band image and the texture of the reference band, and then calculates the pixel gray value of the synthesized band according to the weighted fusion formula to obtain the synthesized band image.

[0079] In some implementations, the band synthesis step specifically includes the following sub-steps:

[0080] S31, Initial fusion sub-step: Perform initial fusion of the multiple single-band images by averaging adjacent bands to fill in the data missing areas in the single-band images.

[0081] Since single-band images often have significant data gaps, making direct texture detection impossible, several rounds of averaging and fusion of adjacent bands (i.e., initial fusion) are required before texture detection to fill in the missing data in a single band. In actual data processing, the number of initial fusion rounds needs to be determined based on the actual data conditions. After initial fusion, all single bands will no longer have missing data or will have only a small amount of missing data, providing a foundation for subsequent texture detection.

[0082] like Figure 6 As shown in the diagram (initial fusion process), the initial fusion adopts the cumulative averaging method of adjacent bands, which can quickly fill the "black line" area without introducing additional texture distortion, so that each single band image has complete pixel coverage, providing a foundation for subsequent texture detection and weighted fusion.

[0083] S32, Texture detection sub-step: Detect texture features of each single-band image using the Sobel operator.

[0084] The Sobel operator extracts texture edge information by calculating the gradient magnitude of an image. It is insensitive to radiation differences and is suitable as a basis for cross-band texture comparison, accurately reflecting the local quality differences between different single bands.

[0085] S33, Weight Calculation Sub-step: Using the single-band image corresponding to the center wavelength as the reference band, calculate the texture difference between other single-band images and the reference band, and determine the fusion weight of each single-band image based on the texture difference.

[0086] like Figure 7As shown in the diagram (a schematic of the weighted fusion process based on texture differences), the center wavelength single band is used as the reference band. The texture detection results of other bands are subtracted from the texture detection results of the reference band to obtain the texture difference images between each band and the reference band. Based on the texture difference images, the fusion weight of each band is further calculated, with the fusion weight of the reference band set to 1, ultimately achieving the fusion of all single bands. Through texture difference analysis, reasonable weights can be assigned to single bands in different regions.

[0087] In some implementations, the weight calculation sub-step is performed on a pixel block basis. By calculating the average texture difference within each pixel block and comparing the average value with a multi-level threshold, a corresponding fusion weight is assigned to all pixels within the block. The smaller the texture difference, the higher the corresponding fusion weight.

[0088] Considering that single-pixel texture differences are easily affected by noise, direct weighting may lead to abrupt changes in weights, resulting in unnatural patches in the fusion result; calculating the mean by using a block-based method can smooth out local noise and make the weight distribution more stable.

[0089] Specifically, such as Figure 8 As shown in the diagram (a schematic diagram of the weight calculation process based on texture differences), the band texture differences are first normalized to avoid the influence of extreme values ​​on the normalization results. The normalization formula is as shown in equation (2):

[0090] (2)

[0091] In equation (2), Indicates differences in band texture; , These represent the pixel values ​​corresponding to the lower and upper percentage points of the band texture difference image, respectively, and are set to 5% and 95% to avoid the influence of extreme values ​​on the normalization results. This represents the normalized band difference image.

[0092] Then with ksize A block of size ksize is used to traverse the normalized texture difference image, with a step size of ksize. The mean texture difference is calculated within each block, and this mean is compared to a multi-level threshold. Then, each pixel within the block is assigned a corresponding weight to complete the weight allocation. Through multiple experiments, empirical values ​​of 0.1, 0.2, and 0.5 were used for the multi-level thresholds, with corresponding weights of 1.0, 0.8, 0.4, and 0.1. The fusion weight for the reference band was set to 1, while the weights for other bands were determined based on texture differences. This block-based weighting method ensures a reasonable weight transition in both texture-rich and uniform regions of the fused image, avoiding abrupt fusion boundaries.

[0093] S34, Weighted Fusion Sub-step: Weighted fusion of each single-band image is performed according to the fusion weight to obtain a composite band image.

[0094] After calculating the fusion weights for each band in step S33 above, band synthesis is performed according to the weighted fusion formula, which is shown in equation (3):

[0095] (3)

[0096] In equation (3), Geographic coordinates; This represents the number of individual bands involved in band synthesis; Indicates the first A single band at geographic coordinates The fusion weight corresponding to the pixel; Indicates the first A single-band image at geographic coordinates The pixel gray value at that location; Indicates the composite band in geographic coordinates The pixel grayscale value at that location.

[0097] like Figure 9 As shown (a schematic diagram of local details in the synthesized band based on texture difference weighted fusion), the problems of blurring, ghosting, and jagged edges in the synthesized band are significantly improved after texture difference weighted fusion. (Comparison) Figure 9 and Figure 10 (Schematic diagram of local details of the synthesized band based on non-average operation) The band synthesis result based on texture difference weighted fusion improves the signal-to-noise ratio by an average of 36.8% compared with the synthesis result based on non-average operation. It can achieve a good balance between image sharpness and signal-to-noise ratio, and significantly improves the problems of "misalignment" and "jagged edges" in the synthesized band, thus improving the image quality of the synthesized band.

[0098] S4, Band registration step: Based on the image pyramid, the synthetic band image is registered in a multi-scale stepwise manner by combining global coarse registration and local fine-tuning to achieve accurate registration between bands.

[0099] The composite images of different bands have complex deformation differences, including global deviations such as overall translation, rotation, and scaling, as well as local distortion deviations caused by factors such as terrain undulation. Furthermore, there are nonlinear differences in the radiometric response between bands. Single-scale registration methods cannot handle the above complex deformations at the same time, making it difficult to achieve accurate registration.

[0100] To address the challenges of registration due to deformation and radiometric differences between imaging spectrometer images across different bands, this step employs an image registration method that combines region and feature-based approaches. Figure 11The image pyramid shown employs a "coarse-to-fine" registration strategy to achieve precise registration between LVF band images.

[0101] The registration parameter transfer between layers of the image pyramid is as shown in equation (4):

[0102] (4)

[0103] In equation (4), Indicates in Layer to be registered band The displacement of a pixel relative to a reference band; The scaling factor between layers is used to consider the scaling relationship between layers when transferring displacement parameters.

[0104] The band registration technique in this step is as follows: Figure 12 As shown, the original scale image (L0 layer) is sequentially subjected to Gaussian filtering and downsampling to obtain a medium-scale local fine-tuning image (L1 layer) and a small-scale coarse registration image (L2 layer). Through multi-scale layer processing, the complex deformation is decomposed into global deviation and local deviation, achieving progressive registration from coarse to fine.

[0105] In some implementations, global coarse registration extracts feature point pairs using a radiation-invariant eigenvalue transformation algorithm, solves for the similarity transformation matrix, eliminates global displacement deviations between bands, and obtains coarse registration parameters.

[0106] Specifically, in the small-scale L2 layer, the band images have undergone several Gaussian filtering and downsampling processes, and the local deformation differences between the band images have been sufficiently smoothed, while the global deformation differences are still preserved. At the same time, since the images have been orthorectified, there are no viewing angle differences between the band images. Therefore, there is a simple similarity transformation relationship between the band images. The registration goal of this layer is to eliminate the global displacement differences between the bands and obtain coarse registration displacement parameters.

[0107] like Figure 13 As shown in the diagram (illustration of RIFT feature matching results), based on the aforementioned image characteristics of the small-scale L2 layer, the RIFT algorithm is used to detect feature point pairs. This algorithm utilizes phase consistency information to construct feature descriptors, exhibiting robustness to nonlinear radiometric differences between images and enabling stable extraction of corresponding feature points between images in different bands. By solving the similarity transformation matrix (including translation, rotation, and scaling parameters) for feature point pairs, the mapping relationship between feature point pairs is fitted to obtain coarse registration parameters, thereby eliminating global displacement deviations between bands and providing a good initial position for subsequent local fine-tuning. The effects before and after L2 layer coarse registration are shown in the diagram. Figure 14 The diagram shows the feature point matching results in the global coarse registration stage.

[0108] In some implementations, local fine-tuning includes: extracting feature points and obtaining corresponding point pairs through template matching; constructing a Delaunay triangulation based on the corresponding point pairs; solving the local affine transformation matrix for each triangular region; and performing piecewise linear transformation to eliminate local displacement deviations between bands.

[0109] Based on the coarse registration displacement parameters, the mesoscale L1 layer uses the Harris corner detector and HOPC descriptor to extract control point pairs using a template matching method. Then, a Delaunay triangulation is constructed using these control point pairs to achieve fine-tuning of local registration. The updated registration parameters are then transferred to the original image's L0 layer, ultimately achieving accurate registration between band images. The specific process is as follows:

[0110] First, the Harris corner detector is used to detect corners in the initial registration bands, extracting a sufficient number of evenly distributed feature points. Then, based on the template matching method, the HOPC descriptor combined with the Normalized Cross-Correlation of HOPC (HOPCncc) similarity index is used to extract corresponding points in the reference bands, obtaining corresponding point pairs, such as... Figure 15 As shown in the figure (a schematic diagram of the extraction results of corresponding point pairs in the local fine-tuning stage), this figure is... Figure 14 The feature point matching results shown are the results of extracting corresponding point pairs in the local fine-tuning stage; then, based on the extracted corresponding point pairs, a Delaunay triangulation is constructed on the reference band, and the local affine transformation matrix is ​​solved based on the triangle vertex coordinates to perform local registration and achieve fine-tuning of local offset between bands; finally, the fine-tuned registration displacement parameters are transferred to the original image scale L0 layer to achieve accurate registration between bands.

[0111] Based on the extracted pairs of corresponding points, a Delaunay triangulation is constructed on the reference band, and local registration is performed based on the piecewise linear (PL) transform. Specifically, for each triangle in the reference band triangulation, it is assumed that the coordinates of its three vertices in the reference band are... The coordinates of the corresponding points in the band to be registered are: Considering the affine transformation relationship between the local triangular regions of the two bands. Then a local affine transformation relation exists. Solving for the affine transformation matrix yields the solution. ,in .

[0112] Taking a specific Delaunay triangle region as an example, for any pixel position within the reference band triangle... The corresponding position of the triangle in the band to be registered is calculated based on the affine transformation model corresponding to the triangle. Next, a reverse mapping method is used to sample positions from the band to be registered. The pixel value at that location is assigned to the position in the reference band image coordinate system. At this point, the result of local registration of the band to be registered is obtained, such as... Figure 16 As shown (a schematic diagram of local registration based on Delaunay triangulation), each triangular region is traversed sequentially, and local transformations are performed one by one to complete the local registration fine-tuning of the mesoscale L1 layer, thereby achieving accurate registration between L1 layer bands.

[0113] In some implementations, in multi-scale stepwise registration, registration parameters are passed stepwise according to the inter-layer scaling factor of the image pyramid to compensate for displacement deviations stepwise, ultimately achieving band registration with sub-pixel accuracy.

[0114] Let the scaling factor between layer L2 and layer L1 be... The scaling factor between L1 and L0 layers is After calculating the local displacement deviation in the L1 layer, it is necessary to update the inter-band displacement deviation in this layer. The registration parameters of the L1 layer can be expressed as shown in equation (5):

[0115] (5)

[0116] In equation (5), This represents the displacement parameters for local fine-tuning of layer L1. This represents the global coarse registration displacement parameters for layer L2. This refers to the new displacement obtained after local fine-tuning of the L1 layer. For example, in a certain sample dataset, This means that a one-pixel offset in the L2 layer will become a two-pixel offset when passed to the L1 layer. The L1 layer registration parameters are then proportionally... By transferring the data to layer L0, the registration displacement field at the original scale can be obtained. Through the stepwise transfer of the scale factor between layers, the problems of large computational cost and easy getting trapped in local optima caused by directly solving complex deformations on large-scale images are avoided, ensuring the stability of the registration process.

[0117] like Figure 17 The image shows a comparison of the color composite effect before and after L1 layer band registration. The red channels represent the initial registered band (based on the L2 layer coarse registration parameters) and the result after local fine-tuning, respectively; the blue and green channels represent the reference band. The comparison clearly shows that after local fine-tuning, the local displacement deviation between the band and the reference band is significantly improved, resulting in a clearer, more accurate color composite with sub-pixel registration precision, fully meeting the needs of quantitative applications of hyperspectral data.

[0118] By employing the above method, geometric correction of the original image is performed first, and a band index file is generated simultaneously. This avoids the tedious calculations of "extracting bands first and then correcting them band by band" in existing technologies, significantly reducing computational complexity and improving data processing efficiency. Through weighted fusion based on texture differences, data gaps in single-band images are effectively filled and the signal-to-noise ratio is improved, thus enhancing the visual quality of the synthesized bands. By combining image pyramids with a multi-scale strategy of global coarse registration and local fine-tuning, global and local biases between bands can be eliminated simultaneously, achieving high registration accuracy even under complex terrain conditions.

[0119] Example 2: A data processing system for a linear graded filter imaging spectrometer

[0120] This invention provides a data processing system for a linear graded-filter imaging spectrometer. See also... Figure 18 The system structure diagram shown includes a geometric correction module 21, a band extraction module 22, a band synthesis module 23, and a band registration module 24. These modules work together to realize the aforementioned data processing method for the linear gradient filter imaging spectrometer.

[0121] The geometric correction module 21 performs geometric correction on each frame of the acquired original image to generate a corrected image. During resampling, it records the band row number of each pixel and generates a band index file for locating irregularly arranged band pixels in the corrected image. The band extraction module 22 extracts data from the same band frame by frame from multiple consecutive corrected images based on the corrected image and the band index file, and stitches them together to form multiple single-band images. The band synthesis module 23 performs weighted fusion based on texture differences on the multiple single-band images to obtain a synthesized band image. The band registration module 24 performs multi-scale, step-by-step registration of the synthesized band image based on an image pyramid, combining global coarse registration with local fine-tuning, to achieve accurate registration between bands. The specific functions of each module are as follows:

[0122] The geometric correction module 21 receives the original image data and integrates a collinearity equation solving unit and a resampling unit. In each frame of the original image acquired by the linear gradient filter imaging spectrometer, pixels in different rows correspond to responses in different bands. After geometric correction, the pixel positions shift, causing the correspondence between bands and pixel positions to become irregular. To address this, the geometric correction module 21 uses the collinearity equation indirect method to perform orthorectified geometric correction on each frame of the original image. By solving for the original image coordinates corresponding to the pixels in the corrected image, it eliminates geometric deformation caused by platform attitude changes and terrain undulations. During resampling, the nearest neighbor method is used, directly taking the gray value of the nearest original pixel without interpolation, avoiding mixing of gray values ​​between pixels in different bands and preventing spectral aliasing. Simultaneously, the geometric correction module 21 records the original pixel band row number corresponding to each corrected pixel, generating a band index file of the same size as the corrected image. The value of each pixel in this index file represents the band row number to which that position belongs in the corrected image. The output of the geometric correction module 21 includes the corrected image and the corresponding band index file, which are simultaneously transmitted to the band extraction module 22.

[0123] The band extraction module 22 receives the corrected image and band index file output by the geometric correction module 21. The image sequence obtained by continuous push-broom scanning covers the entire area of ​​the target scene. For a specific wavelength corresponding to a band, its data is distributed in a specific row position of each frame of the corrected image. However, after geometric correction, the pixels of the same band are no longer strictly arranged in the same row and cannot be directly extracted by row number. The band extraction module 22 traverses each frame of the corrected image and its corresponding band index file. When the value of a pixel in the index file is equal to the target band row number, the gray value of that pixel is extracted and stitched into the image of the target band according to its geographic coordinates. The above operation is repeated for each frame to form a complete single-band image. The above extraction operation is performed on all bands to obtain multiple single-band images, and these single-band images are transmitted to the band synthesis module 23.

[0124] The band synthesis module 23 receives multiple single-band images output by the band extraction module 22. These single-band images have two quality defects: First, the terrain undulation causes the corrected pixels to shrink or stretch, and the mismatch between the UAV's flight speed and the camera's sampling frequency results in "black line" areas with no fill value in the single-band images; Second, the single-band images are obtained by imaging from a row of sensor arrays, resulting in a relatively low signal-to-noise ratio. The band synthesis module 23 performs weighted fusion of each single band based on texture differences. The specific process is as follows: First, the adjacent bands of multiple single band images are cumulatively averaged for initial fusion to fill in the missing data areas. Second, the single band image corresponding to the center wavelength is used as the reference band, and the texture features of each single band image are detected using the Sobel operator. The texture differences between other single bands and the reference band are calculated. The greater the texture difference, the lower the weight assigned during fusion, and the smaller the texture difference, the higher the weight assigned. Finally, the pixel gray values ​​of the synthesized band are calculated according to the weighted fusion formula to obtain the synthesized band image, and the synthesized band image is transmitted to the band registration module 24.

[0125] The band registration module 24 receives the synthesized band image output by the band synthesis module 23. Complex deformation differences exist between the synthesized band images, including global translation, rotation, and scaling deviations, as well as local distortion deviations. Furthermore, there are nonlinear differences in the radiometric response between bands. To address these issues, the band registration module 24 constructs an image pyramid, sequentially performing Gaussian filtering and downsampling on the original-scale image (L0 layer) to obtain mesoscale (L1 layer) and small-scale (L2 layer) images. In the L2 layer, the Radiometric Invariant Feature Transform (RIFT) algorithm is used to extract feature point pairs, solve for the similarity transformation matrix, and obtain coarse registration parameters to eliminate global displacement deviations between bands. After transferring the coarse registration parameters to the L1 layer according to the inter-layer scaling factor, feature points are extracted using a Harris corner detector. Corresponding point pairs are obtained based on template matching, a Delaunay triangulation is constructed, and piecewise linear transformation is performed to achieve local fine-tuning and eliminate local displacement deviations. The updated registration parameters from the L1 layer are then transferred to the L0 layer, ultimately achieving accurate registration between bands at the original image scale.

[0126] Using the above system, the band index file is generated synchronously through the geometric correction module 21, avoiding the tedious calculation of "extracting bands first and then correcting them band by band" in the prior art; the texture difference weighted fusion of the band synthesis module 23 effectively fills the data gaps in single-band images and improves the signal-to-noise ratio; the image pyramid multi-scale registration strategy of the band registration module 24 can simultaneously eliminate global and local biases between bands.

[0127] It should be noted that the system embodiments of the present invention are based on the same inventive concept as the aforementioned method embodiments and can achieve the corresponding technical effects. Each functional module in the system corresponds one-to-one with each step in the aforementioned method embodiments. For the specific implementation method, please refer to the detailed description of the corresponding steps in the aforementioned method embodiments, which will not be repeated here.

[0128] Example 3: A computer device

[0129] This invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned data processing method for a linear graded-filter imaging spectrometer. In other words, by loading and running the computer program, this computer device can sequentially perform geometric correction, band extraction, band synthesis, and band registration on the raw images acquired by the linear graded-filter imaging spectrometer, ultimately outputting precisely registered spectral cube data.

[0130] The memory is used to store computer programs and intermediate data generated during processing. In practical applications, the memory can be read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or other forms of non-volatile storage media. The memory stores at least a computer program in the form of executable code, which includes a set of instructions for implementing the steps in the foregoing method embodiments. Furthermore, the memory can also be used to temporarily store raw image data, corrected images, band index files, single-band images, composite band images, and registration parameters for each stage, so that the processor can quickly read and write them during operation.

[0131] The processor is the core of a computer device, responsible for execution of computer programs stored in memory. A processor can be a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Digital Signal Processor (DSP), Application-Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA), or other programmable logic devices with data processing capabilities. When the processor runs a computer program, it sequentially executes four core steps according to program instructions: geometric correction, band extraction, band synthesis, and band registration. First, each frame of the original image is geometrically corrected to generate a band index file. Then, data of the same band is extracted from the corrected image and the index file and stitched together to form multiple single-band images. Next, weighted fusion based on texture differences is performed on each single-band image to obtain a synthesized band image. Finally, multi-scale hierarchical registration combining global coarse registration and local fine-tuning is performed based on the image pyramid to achieve accurate registration between bands.

[0132] A computer program is a sequence of instructions stored in memory. When loaded and executed by a processor, it enables the processor to perform the aforementioned methods. Computer programs can be written in programming languages ​​such as C++, Python, and MATLAB, and compiled into processor-executable machine code or interpreted bytecode. A program typically includes multiple functional modules, corresponding to geometric correction, band extraction, band synthesis, and band registration steps. During runtime, the computer program can also call low-level mathematical libraries (such as BLAS and LAPACK) for matrix operations, or call image processing libraries (such as OpenCV) for feature detection and image transformation to improve execution efficiency.

[0133] Using the aforementioned computer equipment, the processor executes the computer program stored in memory, enabling the automated completion of the entire data processing workflow for a linear graded filter imaging spectrometer. This equipment embeds the aforementioned method embodiments as software within a general-purpose or dedicated computing platform, facilitating deployment at field sites, data processing centers, or in the cloud. Compared to manual processing or the use of non-dedicated equipment, this computer equipment significantly improves data processing speed and automation, while ensuring the consistency and repeatability of processing results. Since the computer program implements all the technical features of the method embodiments, this equipment also achieves the technical effects of simplified calculations, improved band data quality, and accurate registration under complex deformations, which will not be elaborated further here.

[0134] Example 4: A computer-readable storage medium

[0135] This invention provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the aforementioned data processing method for a linear graded-filter imaging spectrometer. In other words, when the storage medium is installed or connected to a computer device, and the device's processor reads and executes the computer program therein, the device acquires the ability to perform geometric correction, band extraction, band synthesis, and band registration on the original image from the linear graded-filter imaging spectrometer, ultimately outputting precisely registered spectral cube data.

[0136] Computer-readable storage media refers to media capable of storing data using physical methods such as electronic, magnetic, optical, infrared, or semiconductor methods. Specific forms include, but are not limited to: hard disk drives (HDDs), solid-state drives (SSDs), read-only memory (ROM), random access memory (RAM), flash memory, optical discs (CD-ROM, DVD), magnetic tape, magneto-optical discs, and various memory cards (such as SD cards, CF cards). In this embodiment, the storage medium is preferably a non-volatile, rewritable storage medium to facilitate the writing, updating, and repeated reading of computer programs. The storage medium can exist independently of the computer device (such as an optical disc or USB flash drive) or be embedded within the computer device (such as a built-in solid-state drive). Regardless of the form, the core function of the storage medium is to persistently store the binary code or bytecode of the computer program, ensuring that the program is not lost after power failure and is loaded by the processor when needed.

[0137] A computer program is a sequence of instructions stored on a storage medium, and its content is substantially the same as the computer program described in the method embodiments. Computer programs are typically organized in the form of files, which may include executable files (such as .exe), dynamic link libraries (.dll), script files (.py, .m), or bytecode files (.class, .jar). Computer programs on storage media can be written in various ways: for example, during factory production, they can be written to read-only memory using a programmer; or compiled by the developer and copied to media such as hard drives or USB flash drives; or downloaded from the internet and saved to storage media. When a user or system needs to perform data processing, the processor reads the instructions of the computer program from the storage medium, loads them into memory, and executes them line by line, thereby completing steps such as geometric correction, band extraction, band synthesis, and band registration.

[0138] Using the aforementioned computer-readable storage medium, users do not need to rewrite code; they only need to deploy the computer program on the storage medium to a device equipped with a processor and memory to quickly obtain the processing capabilities of linear graded filter imaging spectrometer data. This storage medium facilitates software copying, distribution, and backup, and is particularly suitable for uniformly installing the same processing program on multiple devices (such as multiple UAV ground processing workstations) to ensure consistency in the processing flow. Since the computer program implements all the technical features of the aforementioned method embodiments, when the processor executes the program, it can also achieve the technical effects of simplified calculations, improved band data quality, and accurate registration under complex deformations. Furthermore, the storage medium can also serve as a carrier for the delivery of technological achievements, facilitating patent licensing and technology transfer.

[0139] In summary, the data processing method, system, device, and medium for linear graded-filter imaging spectrometers provided in this invention can function independently or work together to effectively address many shortcomings in existing data processing workflows for linear graded-filter imaging spectrometers. Specific beneficial effects are as follows:

[0140] This invention avoids the cumbersome process of "extracting bands first and then performing geometric correction on each frame of the original image and simultaneously generating a band index file for locating irregularly arranged band pixels. Then, based on this index file, it extracts the same band data frame by frame from multiple corrected images. Geometric correction only needs to be performed once on the original frame image, and band extraction can directly locate irregularly arranged band pixels using the index file, eliminating the need for row-by-row geometric correction for each single band. This significantly reduces computational complexity and improves data processing efficiency.

[0141] In the geometric correction step, the collinear equation indirect method combined with the nearest neighbor resampling method can accurately eliminate geometric deformation caused by platform attitude changes and terrain undulations, while preventing spectral aliasing between different bands, ensuring the geometric accuracy and spectral fidelity of the corrected image, and providing a high-quality data foundation for subsequent band extraction and synthesis.

[0142] In the band synthesis step, an initial fusion is performed by averaging adjacent bands of multiple single-band images. This effectively fills in data gaps caused by terrain undulations or sampling frequency mismatches in the single-band images, ensuring complete pixel coverage for each single-band image. Based on this, the Sobel operator is used to detect texture features. Using the single-band image corresponding to the center wavelength as a reference band, the texture differences between other single-band images and the reference band are calculated. Fusion weights are determined based on these texture differences, with larger differences resulting in lower weights and smaller differences in weight. Finally, a weighted fusion is performed. This method effectively suppresses the "black line" missing problem caused by irregular pixel resampling or sampling frequency mismatches in single-band images, while improving the signal-to-noise ratio and mitigating visual artifacts such as jagged edges and misalignments that may appear in the synthesized band images. This ensures that the quality of the synthesized band images meets the requirements for subsequent quantitative analysis.

[0143] During the weight calculation process, the average texture difference within a pixel block is calculated. This average value is then compared with multi-level thresholds (e.g., 0.1, 0.2, 0.5), and corresponding fusion weights (e.g., 1.0, 0.8, 0.4, 0.1) are assigned to all pixels within the block. The smaller the texture difference, the higher the weight. Using a block-based approach avoids abrupt weight changes caused by noise affecting single-pixel texture differences, resulting in a more stable weight distribution. The fused image maintains a reasonable weight transition in both texture-rich and uniform regions, avoiding abrupt fusion boundaries and further improving the quality of the synthesized bands.

[0144] In the band registration step, by constructing an image pyramid, feature point pairs are extracted at small-scale layers using the Radiometric Invariant Feature Transform (RIFT) algorithm, and the similarity transformation matrix is ​​solved, which can eliminate global displacement bias between bands. The RIFT algorithm is robust to nonlinear radiometric differences and can stably extract corresponding feature points between images of different bands, effectively eliminating overall deformation caused by changes in platform attitude, flight altitude, etc., and obtaining coarse registration parameters.

[0145] Building upon coarse registration, further fine-tuning is performed locally: feature points are extracted and corresponding point pairs are obtained through template matching. A Delaunay triangulation is constructed based on these corresponding point pairs. For each triangular region, the local affine transformation matrix is ​​solved, and piecewise linear transformation is used to eliminate local displacement deviations between bands. This method effectively handles local distortions caused by factors such as the height of ground features and terrain undulations, ensuring that the registered bands remain aligned in detailed regions, thus compensating for the shortcomings of simple global registration.

[0146] In the multi-scale stepwise registration process, the registration parameters are passed stepwise according to the inter-layer scaling factor of the image pyramid to achieve stepwise compensation for displacement deviation. This transfer method avoids the problems of large computational load and easy getting trapped in local optima caused by directly solving complex deformations on large-scale images, and achieves stable registration from coarse to fine. Finally, it obtains sub-pixel level accurate registration at the original image scale, which meets the needs of quantitative applications of hyperspectral data.

[0147] The above method can be implemented through a corresponding system, which includes a geometric correction module, a band extraction module, a band synthesis module, and a band registration module. Each module executes the corresponding steps in the above method, achieving the same technical effects as the method embodiment, namely, simplified calculation, improved band data quality, and accurate registration under complex deformations. Furthermore, storing the above method as a computer program in a computer-readable storage medium and executing it with a computer device can also achieve the aforementioned beneficial effects, facilitating the deployment and application of this data processing method.

[0148] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should be covered within the scope of protection of the present invention.

Claims

1. A data processing method for a linear graded-filter imaging spectrometer, characterized in that, Includes the following steps: Geometric correction steps: Perform geometric correction on each frame of the acquired original image to generate a corrected image, and record the band row number of each pixel during the resampling process to generate a band index file for locating irregularly arranged band pixels in the corrected image. Band extraction step: Based on the corrected image and the band index file, extract the same band data frame by frame from multiple consecutive corrected images, and stitch them together to form multiple single-band images; Band synthesis step: The multiple single-band images are weighted and fused based on texture differences to obtain a synthesized band image; and Band registration steps: Based on the image pyramid, the synthetic band image is registered in a multi-scale stepwise manner, combining global coarse registration and local fine-tuning, to achieve accurate registration between bands.

2. The data processing method according to claim 1, characterized in that, The geometric correction step is implemented using the collinear equation indirect method, and the nearest neighbor method is used during resampling to prevent spectral aliasing between different bands.

3. The data processing method according to claim 1, characterized in that, The band synthesis step specifically includes: Initial fusion sub-step: Perform initial fusion of the multiple single-band images by averaging adjacent bands to fill in the missing data areas in the single-band images; Texture detection sub-step: Detect texture features of each single-band image using the Sobel operator; Weight calculation sub-step: Using the single-band image corresponding to the center wavelength as the reference band, calculate the texture difference between other single-band images and the reference band, and determine the fusion weight of each single-band image based on the texture difference; and Weighted fusion sub-step: Weighted fusion of each single-band image is performed according to the fusion weight to obtain a composite band image.

4. The data processing method according to claim 3, characterized in that, The weight calculation sub-step is executed on a pixel block basis. By calculating the average texture difference within each pixel block and comparing the average value with a multi-level threshold, a corresponding fusion weight is assigned to all pixels within the block. The smaller the texture difference, the higher the corresponding fusion weight.

5. The data processing method according to claim 1, characterized in that, The global coarse registration extracts feature point pairs using the radiation-invariant eigenvalue transformation algorithm, solves the similarity transformation matrix, eliminates global displacement deviations between bands, and obtains coarse registration parameters.

6. The data processing method according to claim 1, characterized in that, The local fine-tuning includes: extracting feature points and obtaining corresponding point pairs through template matching; constructing a Delaunay triangulation based on the corresponding point pairs; solving the local affine transformation matrix for each triangular region; and performing piecewise linear transformation to eliminate local displacement deviations between bands.

7. The data processing method according to claim 1, characterized in that, In the multi-scale step-by-step registration, the registration parameters are passed step-by-step according to the inter-layer scaling factor of the image pyramid to compensate for displacement deviations step by step, and finally achieve band registration with sub-pixel level accuracy.

8. A data processing system for a linear graded-filter imaging spectrometer, characterized in that, include: The geometric correction module is used to perform geometric correction on each frame of the acquired original image, generate the corrected image, and record the band row number of each pixel during the resampling process, generating a band index file for locating irregularly arranged band pixels in the corrected image. The band extraction module is used to extract the same band data frame by frame from multiple consecutive corrected images based on the corrected images and the band index file, and stitch them together to form multiple single-band images. The band synthesis module is used to perform weighted fusion based on texture differences on the multiple single-band images to obtain a synthesized band image; The band registration module is used to perform multi-scale step-by-step registration of the synthetic band image based on the image pyramid, combining global coarse registration with local fine-tuning, to achieve accurate registration between bands.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the data processing method of the linear gradient filter imaging spectrometer as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the data processing method of the linear gradient filter imaging spectrometer as described in any one of claims 1 to 7.

Images