Deep Space Target Data Fusion Enhancement Method Based on Multi-Scale Low-Rank Structure Representation

Abstract

Deep space target data fusion enhancement method based on multi-scale low-rank structure representation, the steps include: collecting multi-source images of deep space targets and some non-image data; preprocessing the collected data and representing it in the form of multi-scale data; using low-rank sparse matrix decomposition to extract significant features from the multi-scale data; fusing the low-rank structure features at each scale; performing post-processing operations such as denoising, sharpening and enhancement on the fused image; inputting the fused feature representation into a classification or detection network to perform specific tasks. The present invention solves the problems of redundancy, noise and scale difference of multi-spectral data, improves the detection accuracy and robustness of space targets, effectively reduces redundant and noise interference, and improves the accuracy of data fusion. Through the convolutional fusion multi-source data fusion strategy, the data dimension is reduced, the complexity between different data sources is eliminated, and the efficient fusion of multi-source data with low-rank structure representation is realized.

Description

Technical Field

[0001] The present invention relates to a deep space target data fusion enhancement method based on multi-scale low-rank structure representation, belonging to the technical field of deep space exploration. Background Art

[0002] Traditional deep space target data fusion technology is a method of fusing visible light images, infrared images, and infrared spectrometer data obtained in deep space target detection. Visible light imaging technology generates images by capturing radiation in the visible light band (usually 400nm to 700nm) through optical systems such as astronomical telescopes. This technology provides high-resolution image information for deep space exploration and is widely used in fields such as astronomy, space exploration, and satellite navigation. Infrared imaging technology usually consists of an infrared sensor and an optical system, and obtains images by detecting the infrared radiation emitted by an object (the wavelength is usually between 0.75 micrometers and 1000 micrometers). These images can not only reveal the visible features of the object but also provide key information such as the temperature and material composition of the object. Infrared spectrometer technology is based on the absorption and emission characteristics of infrared radiation. By detecting the reflection, absorption, or emission spectrum of an object in the infrared band, information such as the molecular structure, chemical composition, and temperature of the substance can be obtained.

[0003] Traditional deep space target data fusion technology has certain defects in multi-scale multi-modal data processing. Traditional processing methods often directly fuse multi-modal data after separate processing, which may lead to the loss of key information and thus affect the effect of data fusion. Due to the lack of synchronization in time, space, and scale between infrared, visible light image data, and non-image infrared spectrum data, traditional fusion technology may not be able to handle this data inconsistency, resulting in information loss in the fused data. Moreover, due to the faint characteristics of space targets, the fused data obtained by traditional methods often cannot make full use of the cross-modal correlation between multi-spectral data for feature enhancement, resulting in information loss. In addition, traditional data fusion methods often involve a large number of operations such as alignment, interpolation, and registration when processing multi-modal data, thus increasing the computational and storage burden.

[0004] The low-rank feature representation method can effectively learn the correlation between different modalities during multi-modal data fusion, thus achieving more comprehensive information integration and being able to effectively reduce the redundant information of the data, making the calculation more efficient. The deep space target data fusion technology based on low-rank feature representation has stronger noise removal ability, higher fusion accuracy, robustness, and flexibility, and can achieve multi-scale fusion and feature enhancement of multi-modal deep space target data, while meeting the requirements of real-time and accuracy for deep space target detection tasks. Summary of the Invention

[0005] The object of the present invention is to provide a deep space target data fusion enhancement method based on multi-scale low-rank structure representation.

[0006] To achieve the above object, the present invention provides a deep space target data fusion enhancement method based on multi-scale low-rank structure representation, including the following steps:

[0007] Step 1. Collect multi-spectral data of the same space target under different time series conditions, including visible light images, infrared images, and infrared spectral data;

[0008] Step 2. Preprocess the visible light and infrared image data collected in Step 1 by means of mean filtering denoising and adaptive histogram equalization (AHE), and preprocess the infrared spectral data by means of Gaussian smoothing denoising and polynomial baseline correction;

[0009] Step 3. Use the Gaussian pyramid method to decompose and represent the multi-source data obtained in Step 2 in the form of a multi-scale matrix, and use the alternating least squares optimization algorithm to solve the low-rank sparse matrix decomposition (LRaSMD) to extract shared features from the multi-time series multi-spectral data;

[0010] Step 4. Use the method of cross-modal attention mechanism to fuse the multi-scale low-rank structure features to obtain complete fused feature data;

[0011] Step 5. Perform post-processing of denoising and image sharpening on the fused feature data obtained in Step 4 to obtain the final multi-source fused feature data.

[0012] Further, the steps of preprocessing the collected visible light and infrared image data by means of mean filtering denoising and adaptive histogram equalization, and preprocessing the infrared spectral data by means of Gaussian smoothing denoising and polynomial baseline correction in Step 2 are as follows:

[0013] Step 2.1-1. Read in the visible light and infrared image data, and perform denoising operations on the images using a denoising method based on mean filtering. Mean Filtering is a common image denoising method that smooths the image and removes noise by replacing the value of each pixel with the average value of its neighboring pixels. Let represent the pixel value at position in the image, represent the filter neighborhood centered on in the image. Select a filter of size 5×5 and perform image boundary processing by copying the edges. The output of the mean filter is represented by the following formula:

[0014] ;

[0015] Assign the calculated mean value to the current pixel to smooth the image.

[0016] Step 2.1-2. Apply the adaptive histogram equalization method to enhance the visible light and infrared images. Adaptive Histogram Equalization (AHE) is an image enhancement method that aims to improve local details by adjusting the contrast of the image. Divide the entire image into windows of size 8×8. For each local window , calculate its histogram normalized cumulative distribution function , and the calculation formula is as follows:

[0017] ;

[0018] Among them, is the normalized histogram frequency of the local window at the pixel value , is the number of pixels in the window, and L is the number of gray levels of the image data;

[0019] For each original pixel in the image, update the value of this pixel to according to the value of the histogram normalized cumulative distribution function of the local window where it is located. The calculation formula is:

[0020] ;

[0021] Step 2.2-1. Apply the Gaussian smoothing method to denoise the infrared spectral data. Gaussian Smoothing is a common image denoising method that smooths the image and removes noise by applying a Gaussian filter. The Gaussian filter is based on the Gaussian function (normal distribution function), and its weight gradually decreases as the distance from the center point increases. Therefore, it can effectively smooth the details and noise in the image while retaining the main structure and edge information of the image. For each pixel point , use the Gaussian filter to perform weighted averaging on its surrounding neighborhood to obtain the smoothed pixel value , and the calculation formula is as follows:

[0022] ;

[0023] Step 2.2-2. Preprocess the infrared spectral data using the method of polynomial baseline correction. Polynomial baseline correction is a technique used to remove baseline drift or background noise in signals and is widely applied in spectral data processing. This method represents the baseline part of the signal by fitting a polynomial curve (usually a low-order polynomial, such as a first- or second-order polynomial) to the signal, and then removes this baseline part from the original signal to correct the true information of the signal. Assume the original data is an array containing wavelength values and corresponding absorption intensities :

[0024] ;

[0025] where is the wavelength and is the corresponding absorption intensity;

[0026] Subsequently, identify the baseline drift region in the spectral data, select a polynomial function of a certain order , and fit the baseline drift in the spectrum using the least squares method. The resulting polynomial form of the baseline drift is:

[0027] ;

[0028] Through the fitted polynomial of the baseline drift , the baseline drift can be removed from the original spectrum to obtain the spectral data after removing the baseline drift , and the calculation formula is:

[0029] ;

[0030] Furthermore, in Step 3, the Gaussian pyramid method is used to decompose and represent the multi-source data obtained in Step 2 in the form of a multi-scale matrix, and the alternating least squares optimization algorithm is used to solve the low-rank sparse matrix decomposition (LRSMD). The steps for extracting shared features from multi-temporal multi-spectral data are as follows:

[0031] Step 3.1. To extract the main components of the multi-source data, use the Gaussian pyramid analysis method to perform multi-scale decomposition on the data. The Gaussian pyramid analysis method is a multi-scale image processing technique that generates a series of images with different resolutions by continuously applying Gaussian filtering to the image and gradually downsampling;

[0032] Represent the preprocessed multi-source data of deep space targets as a matrix , as the first layer of the pyramid, i.e., the bottom layer of the Gaussian pyramid;

[0033] First, apply Gaussian filtering to the data of the first layer. The calculation formula of the filter is:

[0034] ;

[0035] Among them, is the standard deviation, which determines the width of the filter and controls the degree of smoothing;

[0036] Perform downsampling on the Gaussian filtering result, that is, obtain the data of the second layer ;

[0037] Subsequently, continue to apply Gaussian filtering and downsampling to the data of the second layer, that is, obtain the data of the third layer , thus obtaining multi-source data information at three different scales. Each layer of the Gaussian pyramid represents the characteristics of the data at different scales;

[0038] Step 3.2-1. To extract the shared features of multi-source and multi-scale data, the alternating least squares optimization algorithm is used to perform low-rank and sparse matrix decomposition (LRaSMD) on the data. This method aims to simultaneously recover the low-rank part and the sparse part from the given matrix to remove noise and extract the main structural information;

[0039] Let be the data at each scale. The purpose of low-rank and sparse matrix decomposition is to decompose into the sum of two matrices and . Among them, is the low-rank matrix, representing the main structure or trend of the data, is the sparse matrix, representing the outliers, noise or rare features in the data. This method uses the alternating least squares optimization method to optimize the low-rank and sparse matrix decomposition, that is, fixing one matrix in each step and minimizing the other matrix, so as to gradually approach the optimal solution;

[0040] At the initial stage, it is necessary to initialize the matrices and . Initialize as the low-rank approximation matrix of the singular value decomposition of matrix M, The matrix is initialized as a zero matrix. The optimization process of low-rank and sparse matrix decomposition is to minimize the decomposition error of the low-rank matrix and the sparse matrix. The error formula is:

[0041] ;

[0042] Among them, Norm is denoted by, rank represents the rank of the matrix, card represents the number of matrix components, and r represents the low-rank background matrix. The maximum value of the rank, and k represents the sparse matrix. The sparsity of, which reflects the sparse components in the image and is usually defined as of norm;

[0043] Step 3.2-2. Alternately optimize the low-rank matrix and the sparse matrix to gradually reduce the decomposition error of the low-rank matrix and the sparse matrix through the process of alternating optimization;

[0044] First, fix the sparse matrix and optimize the low-rank matrix . The calculation formula for the optimization process is:

[0045] ;

[0046] where t is the dimension of the data matrix;

[0047] Subsequently, fix the low-rank matrix and optimize the sparse matrix . The calculation formula for the optimization process is:

[0048] ;

[0049] As the number of iterations increases, the decomposition error will monotonically decrease; when ( is the selected fault tolerance coefficient), it is judged that the optimization process converges, and the low-rank matrix corresponding to the background information and the sparse matrix corresponding to the abnormal target information are obtained;

[0050] Furthermore, the steps of fusing multi-scale low-rank structural features by using the cross-modal attention mechanism in Step 4 to obtain the complete fused feature data are as follows:

[0051] By adaptively learning the relationships between different modalities and different scales through the cross-modal attention mechanism, the model can automatically focus on important features and suppress redundant or irrelevant information, thereby effectively fusing multi-source information. Specifically, for the image feature and the non-image feature , define query, key, and value respectively, and the calculation formula is:

[0052] ;

[0053] where , , is the learned weight matrix, which is used to map queries, keys, and values respectively;

[0054] By calculating the dot product of the query and the key , the attention weights are obtained, and the calculation formula is:

[0055] ;

[0056] where, is the dimension of the key, which is used to scale the dot product to prevent gradient explosion;

[0057] The image features and the non-image features are fused through feature concatenation to obtain the fused features , and the calculation formula is:

[0058] ;

[0059] Furthermore, the step of performing denoising and image sharpening data post-processing on the fused feature data obtained in Step 4 in Step 5 to obtain the final multi-source fused feature data is as follows:

[0060] The low-rank recovery algorithm is used to perform denoising on the fused feature information. The data matrix is decomposed into the sum of a low-rank matrix and a sparse matrix through the low-rank recovery algorithm. The low-rank matrix represents the main patterns and structures of the data, while the sparse matrix contains noise or outliers, thereby achieving the denoising effect. Specifically, first, the fused feature matrix F obtained in Step 4 is decomposed by SVD to obtain:

[0061] ;

[0062] where, and are singular vectors, is the singular value matrix;

[0063] Subsequently, the singular value matrix is thresholded, that is, the singular values less than a certain threshold are set to zero, to obtain:

[0064] ;

[0065] where, is the threshold parameter, which controls the compression degree of the singular values;

[0066] Finally, the low-rank matrix is obtained through matrix reconstruction:

[0067] ;

[0068] Through the above denoising operation, a low-rank matrix that removes noise and retains the main structure of the original matrix is obtained. ;

[0069] The Laplace filtering method is used to sharpen and enhance the feature data. Specifically, in the application process, first, the original feature data is Laplace filtered to obtain the second derivative (i.e., the rate of change) of the data. Then, by weighted superposition of the result after Laplace filtering and the original data (usually setting the weighting coefficient as a positive value), the edge part in the image or signal is enhanced, thereby improving the detail contrast and structural information, achieving the effect of feature enhancement.

[0070] Let represent the matrix data before processing. The calculation formula of the Laplace filtering method is:

[0071] ;

[0072] where is a parameter that controls the sharpening intensity.

[0073] After the above post-processing operations, the final multi-scale fusion enhanced image data of deep space targets is obtained, which can meet the requirements of subsequent tasks such as deep space target detection.

[0074] Beneficial effects: It solves the problems of redundancy, noise, and scale differences in multi-spectral data, can make full use of multi-source deep space target images and spectral data, improves the data fusion effect and the quality of the fused image, has strong robustness and stability, effectively reduces redundancy and noise interference, improves the accuracy of data fusion, and at the same time can reduce the complexity of calculation and storage, realizes the efficient fusion of multi-source data with low-rank structure representation, and is applicable to the detection and recognition tasks of deep space dim target objects. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0076] Figure 1 is the flow chart of the present invention;

[0077] Figure 2 is the flow chart of multi-scale decomposition of multi-source data using the Gaussian pyramid method;

[0078] Figure 3 is the flow chart of low-rank sparse matrix decomposition using the alternating least squares optimization algorithm;

[0079] Figure 4 It is a schematic diagram of the method for fusing multi-source low-rank structure features by using a cross-modal attention mechanism. Specific implementation manners

[0080] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0081] As Figure 1 shown, the deep space target data fusion enhancement method based on multi-scale low-rank structure representation includes the following steps:

[0082] Step 1. Collect multi-spectral data of the same space target under different time series conditions, including visible light images, infrared images, and infrared spectral data;

[0083] Step 2. Preprocess the visible light and infrared image data collected in Step 1 by means of mean filtering denoising and adaptive histogram equalization (AHE), and preprocess the infrared spectral data by means of Gaussian smoothing denoising and polynomial baseline correction;

[0084] Step 3. Use the Gaussian pyramid method to decompose and represent the multi-source data obtained in Step 2 in the form of a multi-scale matrix, and use the alternating least squares optimization algorithm to solve the low-rank sparse matrix decomposition (LRaSMD) to extract shared features from the multi-time series multi-spectral data;

[0085] Step 4. Use the method of cross-modal attention mechanism to fuse the multi-scale low-rank structure features to obtain complete fused feature data;

[0086] Step 5. Perform post-processing of denoising and image sharpening on the fused feature data obtained in Step 4 to obtain the final multi-source fused feature data.

[0087] As a preferred implementation manner, the specific steps of the said Step 2 are:

[0088] Step 2.1-1. Read in the visible light and infrared image data, and perform denoising operations on the images by using a denoising method based on mean filtering. Mean Filtering is a common image denoising method that smooths the image and removes noise by replacing the value of each pixel with the average value of its neighboring pixels. Let represent the position in the image The pixel value, represents the filter neighborhood centered on in the image. A 5×5 filter is selected, and the image boundary is processed by copying the edges. The output of the mean filter is expressed by the following formula:

[0089] ;

[0090] Assign the calculated mean value to the current pixel to smooth the image;

[0091] Step 2.1-2. Use the adaptive histogram equalization method to enhance the visible light and infrared images. Adaptive Histogram Equalization (AHE) is an image enhancement method that aims to improve local details by adjusting the contrast of the image. The entire image is divided into 8×8 windows. For each local window , calculate its histogram normalized cumulative distribution function , and the calculation formula is as follows:

[0092] ;

[0093] where, is the normalized histogram frequency of the local window at the pixel value , is the number of pixels in the window, and L is the number of gray levels of the image data;

[0094] For each original pixel in the image, update the value of this pixel to according to the value of the histogram normalized cumulative distribution function of the local window where it is located. The calculation formula is:

[0095] ;

[0096] Step 2.2-1. Use the Gaussian smoothing method to denoise the infrared spectral data. Gaussian Smoothing is a common image denoising method that smooths the image and removes noise by applying a Gaussian filter. The Gaussian filter is based on the Gaussian function (normal distribution function), and its weight gradually decreases as the distance from the center point increases. Therefore, it can effectively smooth the details and noise in the image while retaining the main structure and edge information of the image. For each pixel point , use the Gaussian filter to perform weighted averaging on its surrounding neighborhood to obtain the smoothed pixel value , and the calculation formula is as follows:

[0097] ;

[0098] Step 2.2-2. Preprocess the infrared spectral data using the method of polynomial baseline correction. Polynomial Baseline Correction is a technique for removing baseline drift or background noise in signals and is widely used in spectral data processing. This method represents the baseline part of the signal by fitting a polynomial curve (usually a low-order polynomial, such as a first- or second-order polynomial) to the signal, and then removes this baseline part from the original signal to correct the true information of the signal. Assume the original data is an array containing wavelength values and corresponding absorption intensities :

[0099] ;

[0100] where is the wavelength and is the corresponding absorption intensity;

[0101] Subsequently, identify the baseline drift region in the spectral data, select a polynomial function of a certain order, and fit the baseline drift in the spectrum by the least squares method. The polynomial form of the obtained baseline drift is:

[0102] ;

[0103] Through the fitted baseline drift polynomial , the baseline drift can be removed from the original spectrum to obtain the spectral data after removing the baseline drift. The calculation formula is:

[0104] ;

[0105] As a preferred implementation manner, the specific steps of Step 3 are:

[0106] Step 3.1. In order to extract the main components of the multi-source data, use the Gaussian pyramid analysis method to perform multi-scale decomposition on the data. The Gaussian pyramid analysis method is a multi-scale image processing technique that generates a series of images with different resolutions by continuously applying Gaussian filtering to the image and gradually downsampling, as shown in Figure 2 ;

[0107] Represent the preprocessed multi-source data of deep space targets as a matrix , as the first layer of the pyramid, i.e., the bottom layer of the Gaussian pyramid;

[0108] First, apply Gaussian filtering to the data of the first layer. The calculation formula is:

[0109] ;

[0110] Among them, is the standard deviation, which determines the width of the filter and controls the degree of smoothing;

[0111] Perform downsampling on the Gaussian filtering result, that is, obtain the data of the second layer ;

[0112] Subsequently, continue to apply Gaussian filtering and downsampling to the data of the second layer, that is, obtain the data of the third layer , thus obtaining multi-source data information at three different scales. Each layer of the Gaussian pyramid represents the characteristics of the data at different scales;

[0113] Step 3.2-1. In order to extract the shared features of multi-source and multi-scale data, the alternating least squares optimization algorithm is used to perform low-rank sparse matrix decomposition (LRaSMD) on the data. This method aims to simultaneously recover the low-rank part and the sparse part from a given matrix to remove noise and extract the main structural information;

[0114] Let be the data at each scale. The purpose of low-rank sparse matrix decomposition is to decompose into the sum of two matrices and . Among them, is the low-rank matrix, representing the main structure or trend of the data, is the sparse matrix, representing the outliers, noise or rare features in the data. This method uses the alternating least squares optimization method to optimize the low-rank sparse matrix decomposition, that is, fixing one matrix in each step and minimizing the other matrix, so as to gradually approach the optimal solution, as shown in Figure 3 ;

[0115] At the initial stage, it is necessary to initialize the matrices and . Initialize as the low-rank approximation matrix of the singular value decomposition of matrix M, The matrix is initialized as a zero matrix. The optimization process of low-rank sparse matrix decomposition is to minimize the decomposition error of the low-rank matrix and the sparse matrix. The error formula is:

[0116] ;

[0117] Among them, Norm is denoted as, rank represents the rank of a matrix, card represents the number of matrix components, and r represents the low-rank background matrix The maximum value of the rank, and k represents the sparse matrix The sparsity of, which reflects the sparse components in the image and is usually defined as of norm;

[0118] Step 3.2-2. Alternately optimize the low-rank matrix and the sparse matrix to gradually reduce the decomposition error of the low-rank matrix and the sparse matrix through the process of alternating optimization;

[0119] First, fix the sparse matrix and optimize the low-rank matrix . The calculation formula for the optimization process is:

[0120] ;

[0121] where t is the dimension of the data matrix;

[0122] Subsequently, fix the low-rank matrix and optimize the sparse matrix . The calculation formula for the optimization process is:

[0123] ;

[0124] As the number of iterations increases, the decomposition error will monotonically decrease. When ( is the selected fault tolerance coefficient), it is determined that the optimization process converges, and the low-rank matrix corresponding to the background information and the sparse matrix corresponding to the abnormal target information are obtained.

[0125] As a preferred implementation manner, the specific steps of Step 4 are as follows:

[0126] By adaptively learning the relationships between different modalities and different scales through the cross-modal attention mechanism, the model can automatically focus on important features, suppress redundant or irrelevant information, and thus effectively fuse multi-source information together, as Figure 4 shown. Specifically, define queries, keys, and values for the image feature and the non-image feature respectively. The calculation formula is:

[0127] ;

[0128] where , , is the learned weight matrix, which is used to map queries, keys, and values respectively;

[0129] By calculating the dot product of the query and the key the attention weights are obtained, and the calculation formula is:

[0130] ;

[0131] where, is the dimension of the key, which is used to scale the dot product to prevent gradient explosion;

[0132] The image features and non-image features are fused through feature concatenation to obtain the fused feature , and the calculation formula is:

[0133] ;

[0134] As a preferred implementation manner, the specific steps of the said Step 5 are:

[0135] The low-rank recovery algorithm is adopted to denoise the fused feature information. By the low-rank recovery algorithm, the data matrix is decomposed into the sum of a low-rank matrix and a sparse matrix. The low-rank matrix represents the main patterns and structures of the data, while the sparse matrix contains noise or outliers, so as to achieve the denoising effect. Specifically, first, the fused feature matrix F obtained in Step 4 is subjected to SVD decomposition to obtain:

[0136] ;

[0137] where, and are singular vectors, is the singular value matrix;

[0138] Subsequently, the singular value matrix is thresholded, that is, the singular values less than a certain threshold are set to zero, and we get:

[0139] ;

[0140] where, is the threshold parameter, which controls the compression degree of the singular values;

[0141] Finally, the low-rank matrix is obtained through matrix reconstruction:

[0142] ;

[0143] Through the above denoising operation, a low-rank matrix .

[0144] The Laplace filtering method is used to sharpen and enhance the feature data. Specifically, in the application process, first, the original feature data is subjected to Laplace filtering to obtain the second derivative (i.e., the rate of change) of the data. Then, by weighted superposition of the result after Laplace filtering and the original data (usually setting the weighting coefficient as a positive value), the edge part in the image or signal is enhanced, thereby improving the detail contrast and structural information, achieving the effect of feature enhancement;

[0145] Let represent the matrix data before processing. The calculation formula of the Laplace filtering method is:

[0146] ;

[0147] where is a parameter to control the sharpening intensity.

[0148] After the above post-processing operations, the final multi-scale fusion enhanced image data of deep space targets is obtained, which can meet the requirements of subsequent tasks such as deep space target detection.

[0149] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and its equivalent technologies, the present invention also intends to include these changes and modifications.

Claims

1. A deep space target data fusion enhancement method based on multi-scale low-rank structure representation, characterized in that It includes the following steps: Step 1. Collect multi-spectral data of the same space target under different temporal conditions, including visible light images, infrared images, and infrared spectral data; Step 2. Preprocess the visible light and infrared image data collected in Step 1 by means of mean filtering denoising and adaptive histogram equalization, and preprocess the infrared spectral data by means of Gaussian smoothing denoising and polynomial baseline correction; Step 3. Use the Gaussian pyramid method to decompose and represent the multi-source data obtained in Step 2 in the form of a multi-scale matrix, and use the alternating least squares optimization algorithm to solve the low-rank sparse matrix decomposition to extract shared features from the multi-temporal multi-spectral data; The specific steps include: Step 3.1-1. To extract the main components of the multi-source data, use the Gaussian pyramid analysis method to perform multi-scale decomposition on the data; Represent the preprocessed multi-source deep space target data as a matrix , which serves as the first layer of the pyramid, i.e., the bottom layer of the Gaussian pyramid; First, apply Gaussian filtering to the first-layer data. The filter calculation formula is: ； Among them, is the standard deviation, which determines the width of the filter and controls the degree of smoothing; Downsample the Gaussian filtering result to obtain the second-layer data ; Subsequently, Gaussian filtering and downsampling are continued to be applied to the second-layer data, and thus the third-layer data is obtained. , thereby obtaining multi-source data information at three different scales. Each layer of the Gaussian pyramid represents the characteristics of the data at different scales. Step 3.2-1. To extract the shared features of multi-source and multi-scale data, the data is decomposed into a low-rank sparse matrix using the alternating least squares optimization algorithm. Let be the data at each scale. The purpose of the low-rank sparse matrix decomposition is to decompose into the sum of two matrices, and . Among them, is the low-rank matrix, representing the main structure or trend of the data, and is the sparse matrix, representing the outliers, noise or rare features in the data. The alternating least squares optimization method is used to optimize the low-rank sparse matrix decomposition, that is, in each step, one matrix is fixed and the other matrix is minimized, so as to gradually approach the optimal solution; At the beginning, the matrix and need to be initialized. is initialized as the low-rank approximation matrix of the singular value decomposition of matrix M. The matrix is initialized as a zero matrix. The optimization process of low-rank sparse matrix decomposition is to minimize the decomposition error between the low-rank matrix and the sparse matrix. The error formula is: ； Among them, represents the norm, rank represents the matrix rank, card represents the number of matrix components, and r represents the low-rank background matrix The maximum value of the rank, and k represents the sparse matrix The sparsity, which reflects the sparse components in the image, is defined as of norm; Step 3.2-2. Alternately optimize the low-rank matrix and the sparse matrix : First, fix the sparse matrix , and optimize the low-rank matrix . The calculation formula for the optimization process is as follows: ； where t is the dimension of the data matrix; Subsequently, the low-rank matrix is fixed , and the sparse matrix is optimized , and the calculation formula for the optimization process is as follows: ； As the number of iterations increases, the decomposition error will monotonically decrease. When is reached, it is determined that the optimization process has converged, and a low-rank matrix corresponding to the background information and a sparse matrix corresponding to the abnormal target information are obtained. is the selected fault tolerance coefficient; Step 4. Use the cross-modal attention mechanism method to fuse the multi-scale low-rank structural features to obtain complete fused feature data; Step 5. Perform post-processing of denoising and image sharpening on the fused feature data obtained in Step 4 to obtain the final multi-source fused feature data.

2. The deep space target data fusion enhancement method based on multi-scale low-rank structure characterization according to claim 1, wherein The specific steps of Step 2 are: Step 2.1-1. Read in the visible light and infrared image data, and perform denoising operations on the images using a denoising method based on median filtering. Let represent the pixel value at position in the image. denote the filter neighborhood centered on in the image. Select a 5×5 filter and perform image boundary processing by copying the edges. The output of the median filter is expressed by the following formula: ； Assign the calculated mean value to the current pixel to smooth the image; Step 2.1-2. Use the adaptive histogram equalization method to enhance the visible light and infrared images; divide the entire image into windows of 8×8 size, and for each local window , calculate its histogram normalized cumulative distribution function , and the calculation formula is as follows: ； Among them, is a local window is the normalized histogram frequency at the pixel value is the number of pixels in the window, and L is the number of gray levels of the image data; For each original pixel in the image , update the value of the pixel to according to the value of the normalized cumulative distribution function of the histogram of the local window where it is located. The calculation formula is: ； Step 2.2-1. Denoise the infrared spectrum data using the Gaussian smoothing method. For each pixel point , use a Gaussian filter to perform weighted averaging on its surrounding neighborhood to obtain the smoothed pixel value . The calculation formula is as follows: ； Step 2.2-2. Preprocess the infrared spectral data using the method of polynomial baseline correction. Assume that the original data is an array containing wavelength values and the corresponding absorption intensities : ； Among them, is the wavelength, is the corresponding absorption intensity; Subsequently, the baseline drift region in the spectral data is identified, and a polynomial function is selected, and the baseline drift in the spectrum is fitted by the least squares method. The resulting polynomial form of the baseline drift is: ； Through the fitted baseline drift polynomial , remove the baseline drift from the original spectrum to obtain the spectral data after removing the baseline drift , and the calculation formula is: .

3. The deep space target data fusion enhancement method based on multi-scale low-rank structure characterization according to claim 1, characterized in that The specific steps of Step 4 are: Step 4.1-1: Define query, key, and value for image features and non-image features respectively, and the calculation formula is as follows: ； Among them, is the learned weight matrix, which is used to map queries, keys, and values respectively; Step 4.1-2: Query by calculation and key to obtain the attention weight. The calculation formula is as follows: ； Among them, is the dimension of the key; Step 4.1-3. Fuse the image features and non-image features to obtain fused features . The calculation formula is as follows: 。 4. The deep space target data fusion enhancement method based on multi-scale low-rank structure characterization according to claim 1, characterized in that The specific steps of Step 5 are: Step 5.

1. Use the low-rank recovery algorithm to perform denoising operations on the fused feature information; First, perform SVD decomposition on the fused feature matrix F obtained in Step 4 to get: ； Among them, and are singular vectors, is a singular value matrix; Threshold the singular value matrix, that is, set the singular values less than a certain threshold to zero to get: ； Among them, is a threshold parameter that controls the compression degree of singular values; Finally, a low-rank matrix is obtained through matrix reconstruction : ； Through the above denoising operation, a low-rank matrix that removes noise and retains the main structure of the original matrix is obtained. ; Step 5.2-1. Sharpen and enhance the feature data using the Laplace filtering method. Let represent the matrix data before processing. The calculation formula of the Laplace filtering method is: ； Among them, is a parameter for controlling the sharpening intensity; After the above post-processing operations, the final multi-scale fusion enhanced image data of deep space targets is obtained.

Images