Stripe non-uniformity correction method based on time-space domain adaptive one-dimensional filtering

Abstract

The invention discloses a stripe heterogeneity correction method based on time-space domain adaptive one-dimensional filtering, which comprises the following steps: extracting initial stripe heterogeneity by adopting one-dimensional discrete wavelet transform, and forming a time-space domain adaptive one-dimensional filter by combining time domain mean filtering and spatial domain horizontal direction one-dimensional weighted guide filtering. The method comprises the following steps: adaptively adjusting filter parameters from a spatial domain based on Median Absolute Deviation (MAD), and adaptively adjusting the filter parameters from an intensity domain based on two-dimensional discrete wavelet transform and one-dimensional edge perception weighting. Based on an iterative correction strategy, the filter has convergence capability by increasing a constraint threshold, removing an inaccurate estimated value and inhibiting error accumulation. According to the method, generalization ability, correction precision and calculation efficiency of the scene can be considered, stripe heterogeneity and scene edge distinguishing ability are achieved, and self-adaptive correction of infrared image stripe heterogeneity is achieved without depending on scene change and content.

Description

Technical Field

[0001] The invention belongs to infrared imaging non-uniformity correction technology, and in particular relates to a stripe non-uniformity correction method based on time-space domain adaptive one-dimensional filtering. Background Art

[0002] Infrared imaging systems can be divided into scanning type and staring type according to the imaging method. The scanning type uses a strip-shaped one-dimensional infrared focal plane to acquire infrared images by row or column in time-sharing, and then splices these rows or columns into an image. It is often used for long-distance satellite remote sensing imaging and acquisition of high-resolution images. Due to the discontinuity of the scanning method, the image has stripe non-uniformity. The staring type uses a flat two-dimensional infrared focal plane, which is often used for close-range imaging. The image is output through a shared amplifier and analog-to-digital converter for rows or columns. Affected by the differences in electronic components, the image has stripe non-uniformity. Stripe non-uniformity is a kind of noise superimposed on the real scene. Because it changes slowly, it is also called fixed pattern noise, which seriously degrades the image quality.

[0003] In order to restore the image quality, the correction methods are divided into two categories: calibration-based correction and scene-based correction. The calibration correction method commonly used in engineering is single-point baffle correction, which corrects the drift non-uniformity effect by periodically blocking the uniform baffle. However, the blind effect generated during the occlusion period is not suitable for real-time application scenarios. Based on the scene correction method, there are mainly statistical-based and registration-based correction methods. The statistical method relies on the full movement of the scene, and the slow-moving scene is easily regarded as a non-uniform miscorrection, resulting in negative ghosting; the registration-based correction method relies on the scene-specific global translation movement; the deep learning method has been a hot topic in recent years, but it is still far from engineering application due to limited training data, personalized differences in the non-uniformity of infrared imaging systems and high computational costs. There are also some model optimization correction methods that use image and non-uniformity prior information. By constructing a cost function to solve, the computational cost is usually high, and the use of video stream information is abandoned, which limits the correction accuracy. In summary, it is difficult for the above methods to take into account the generalization ability of the scene, correction accuracy and computational efficiency. Especially when the stripe non-uniformity coincides with the edge of the scene, it is easy for the stripes to be under-corrected, over-corrected, or blurred at the edge. Summary of the invention

[0004] The purpose of the present invention is to provide a stripe non-uniformity correction method based on spatiotemporal adaptive one-dimensional filtering, which has the ability to distinguish stripe non-uniformity and scene edges, and realizes adaptive correction of infrared image stripe non-uniformity independent of scene changes and content. Even when the stripe non-uniformity and the scene edge coincide, it can avoid the situation of stripe under-correction or stripe over-correction and edge blur.

[0005] The technical solution to realize the present invention is: a stripe non-uniformity correction method based on spatiotemporal adaptive one-dimensional filtering, the steps are as follows:

[0006] Step 1: Use an infrared imaging system to collect images with an image resolution of m×n;

[0007] Step 2: For the first frame image z(1), the spatial mean is calculated by column, the column spatial mean vector of z(1) is obtained and normalized, and the high-frequency detail component is extracted by one-dimensional single-scale discrete wavelet transform for the normalized vector. After copying m rows, it is expanded to an m×n initial stripe non-uniformity correction image.

[0008] Step 3: Use the infrared imaging system to continuously collect images and calculate the time domain mean image of the jth frame Using the iterative correction strategy, the iteratively updated fringe non-uniformity correction image is obtained. and the iteratively updated time domain mean correction image Initialize the fringe non-uniformity correction image and residual fringe non-uniformity update image

[0009] Step 4: Correct the iteratively updated time domain mean image The Median Absolute Deviation (MAD) method is used to estimate the intensity of the overall fringe non-uniformity of the image, and the radius r of the filter is adaptively adjusted in the spatial domain.

[0010] Step 5: Correct the iteratively updated time domain mean image The improved horizontal one-dimensional edge-aware weighting matrix Γ is used to improve the image The ability to distinguish the non-uniformity of stripes and the edges of scenes. Image extraction using two-dimensional single-scale discrete wavelet transform The vertical detail component matrix V of the matrix V is taken, and the median of the absolute values ​​of each column of the matrix V is taken as the estimated value of the non-uniformity intensity of each column of the stripes. Combined with Γ, the regularization factor matrix λ′ of the filter is adaptively adjusted from the intensity domain.

[0011] Step 6: Correct the iteratively updated time domain mean image Combined with the horizontal one-dimensional weighted guided filter in the spatial domain, the adaptive filter radius r in step 4 and the regular factor matrix λ′ in step 5 are substituted to form a spatiotemporal adaptive one-dimensional filter. Subtract the output image q of the spatiotemporal adaptive one-dimensional filter to obtain the filtered image Will As the preliminary extracted residual stripe non-uniformity image.

[0012] Step 7: Initially extract the residual stripe non-uniformity image By increasing the constraint threshold TH, the inaccurate stripe non-uniformity estimation value at the edge of the scene is set to 0, and the constrained residual stripe non-uniformity image is obtained, which is denoted as right Calculate the non-zero spatial domain mean value by column to get a 1×n vector, copy m rows and expand it to an m×n residual stripe non-uniformity update image

[0013] Step 8: Calculate the corrected image of the jth frame

[0014] Compared with the prior art, the present invention has the following significant advantages:

[0015] (1) The MAD method is used for the first time to estimate the overall fringe non-uniformity intensity of the current corrected image. It is not restricted by scene content and fringe intensity and can estimate the overall fringe non-uniformity intensity of the image, and adaptively adjust the radius r of the spatiotemporal adaptive one-dimensional filter from the spatial domain.

[0016] (2) For the first time, the two-dimensional single-scale discrete wavelet transform is used to extract the median of the absolute values ​​of each column in the vertical detail component image V as the estimated value of the stripe non-uniformity intensity of each column. Combined with the one-dimensional edge-aware weighting matrix Γ in the horizontal direction, the global and local aspects are associated with each column, and the regularization factor matrix λ′ of the spatiotemporal adaptive one-dimensional filter is adaptively adjusted from the intensity domain.

[0017] (3) For the first time, the constraint threshold TH is increased to set the inaccurate stripe non-uniformity estimation value at the edge of the scene to 0, suppressing the error accumulation caused by iterative correction, making the spatiotemporal adaptive one-dimensional filter have convergence ability, and enhancing the robustness and correction accuracy of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 The present invention is a flow chart of a stripe non-uniformity correction method based on spatiotemporal adaptive one-dimensional filtering.

[0019] Figure 2 This is a model diagram of the stripe non-uniformity correction method based on spatiotemporal adaptive one-dimensional filtering in the present invention.

[0020] Figure 3 This is a subjective comparison experimental result diagram of five stripe non-uniformity correction methods and the method proposed in the present invention, where two images in the same column are a group, from left to right: non-uniformity degradation image group, SSS image group, DDBP image group, TVBP image group, WDGF image group, WGFMF image group, and Proposed image group.

[0021] Figure 4The figures are the quantitative comparison experimental results of five stripe non-uniformity correction methods and the method proposed in the present invention, where (a) is the PSNR graph; (b) is the SSIM graph. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] The technical solutions between the various embodiments of the present invention can be combined with each other, but it must be based on the fact that ordinary technicians in the field can implement it. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0024] The following will further introduce the specific implementation method, as well as the technical difficulties and inventive points of this invention in combination with this design example.

[0025] Combination Figure 1 and Figure 2 , a stripe non-uniformity correction method based on spatiotemporal adaptive one-dimensional filtering, comprising the following steps:

[0026] Step 1: Use an infrared imaging system to collect images with an image resolution of m×n.

[0027] Step 2: For the first frame image z(1), the spatial mean is calculated by column, the column spatial mean vector of z(1) is obtained and normalized, and the high-frequency detail component is extracted by one-dimensional single-scale discrete wavelet transform for the normalized vector. After copying m rows, it is expanded to an m×n initial stripe non-uniformity correction image. The details are as follows:

[0028] For the first frame image z(1), calculate the spatial mean by column and obtain the column spatial mean vector of the first frame image z(1) represents a vector, the subscript y represents the number of columns, Represents the spatial mean operation.

[0029] The spatial mean of the yth column of the first frame image z(1) It is expressed as:

[0030]

[0031] (x,y) represents the pixel position.

[0032] right The normalized vector is recorded as

[0033] The column spatial mean vector after normalization of the yth column It is expressed as:

[0034]

[0035] right A one-dimensional single-scale discrete wavelet transform is used to extract high-frequency detail components, and the initial stripe non-uniformity correction image is expanded to m×n after copying m rows. The details are as follows:

[0036] Pair Vector One-dimensional single-scale discrete wavelet transform is used to extract high-frequency detail component vectors

[0037]

[0038] ^ represents estimated value, DWT 1D Represents one-dimensional single-scale discrete wavelet transform operation, and sym8 represents the Symlets8 wavelet basis. Represents the high-frequency detail component vector, which mainly includes stripe non-uniformity. To maintain the high-frequency detail component vector The length of is consistent with the image column width n, DWT 1D No downsampling is done.

[0039] The high-frequency detail component vector The initial fringe non-uniformity correction image is expanded to m×n after copying m rows.

[0040]

[0041] The symbol repmat stands for the vector replication operation, and [m 1] means to replicate m rows and 1 column, converting the 1×n high-frequency detail component vector The initial fringe non-uniformity correction image expanded to m×n

[0042] Step 3: Use the infrared imaging system to continuously collect images and calculate the time domain mean image of the jth frame Using an iterative correction strategy, the fringe non-uniformity correction image is iteratively updated and time domain mean correction image Initialize the fringe non-uniformity correction image and residual fringe non-uniformity update image The details are as follows:

[0043] Use the infrared imaging system to continuously collect images and count the time domain mean image of the jth frame

[0044]

[0045] Represents the time domain mean operation.

[0046] Using the iterative correction strategy, the iteratively updated fringe non-uniformity correction image is obtained. and the iteratively updated time domain mean correction image

[0047]

[0048] represents the fringe non-uniformity corrected image of the jth frame, Represents the residual fringe non-uniformity update image of the (j-1)th frame, and initializes the fringe non-uniformity correction image And the residual fringe non-uniformity update image

[0049] Step 4: Correct the iteratively updated time domain mean image Estimate using the Median Absolute Deviation (MAD) method The overall fringe inhomogeneity intensity σ s , the radius r of the filter is adaptively adjusted from the spatial domain, as follows:

[0050] S4-1, calculate the iteratively updated time domain mean correction image Spatial mean of each column:

[0051]

[0052] in, Represents the iteratively updated time domain mean correction image The spatial mean of the yth column is represented by the vector

[0053] S4-2, Calculation Image The forward difference of the means of two adjacent columns constitutes the forward difference vector

[0054]

[0055] S4-3, calculate the MAD value to exclude Outliers in:

[0056]

[0057] med stands for median operation.

[0058] S4-4, removed The vector after the outlier is recorded as

[0059]

[0060] S4-5, estimate the iteratively updated time domain mean correction image The overall fringe inhomogeneity intensity σ s :

[0061]

[0062] std stands for standard deviation operation.

[0063] S4-6, according to the confidence interval of the normal distribution, the radius r of the filter is adaptively adjusted from the spatial domain:

[0064]

[0065] in, Represents rounding up.

[0066] Step 5: Correct the iteratively updated time domain mean image The improved horizontal one-dimensional edge-aware weighting matrix Γ is used to improve the image The ability to distinguish the non-uniformity of stripes and the edges of scenes. Image extraction using two-dimensional single-scale discrete wavelet transform The vertical detail component matrix V is taken, and the median of the absolute values ​​of each column of the matrix V is taken as the estimated value of the non-uniformity intensity of each column of the stripes. Combined with Γ, the regular factor matrix λ′ of the filter is adaptively adjusted from the intensity domain, as follows:

[0067]

[0068] Where L is The dynamic range of the sensor is ε, the sensitivity coefficient is ε, and the proportional coefficient is K.

[0069] Represents the time domain mean correction image updated in iteration The variance of a 1×3 local window centered at pixel p′. represents the variance of a 1×3 local window centered on pixel p. Pixel p represents All pixels of . Γ(p′) represents the one-dimensional edge-aware weighting in the horizontal direction at p′. In order to associate the local and the global, Γ(p′) is used The variance of all pixels in their respective 1×3 local window is normalized.

[0070] The present invention proposes for the first time to add a proportional coefficient K to the sensitivity coefficient ε formula to adjust the resolution sensitivity to stripe non-uniformity and scene edges, and usually K is set to 0.1; the two-dimensional edge perception weighting is changed to one-dimensional edge perception weighting, so that it focuses on the resolution ability of vertical stripe non-uniformity and scene edges, further improving the resolution sensitivity.

[0071] Image Extraction Using Two-Dimensional Single-Scale Discrete Wavelet Transform The vertical detail component matrix V of:

[0072]

[0073] DWT 2D represents a two-dimensional single-scale discrete wavelet transform, haar represents the Haar wavelet basis, and V represents the detail component matrix in the vertical direction. To keep the matrix V consistent with the image The resolution of DWT is consistent 2D No downsampling is done.

[0074] Take the column vectors of matrix V The median of the absolute values ​​is used as an estimate of the inhomogeneity intensity of each row of fringes:

[0075]

[0076] Represents the y-th column vector of the matrix V. σ y Represents the estimated standard deviation of the non-uniformity intensity of the y-th column fringe. y represents the regularization factor of the yth column, and the vector is expressed as Vector After copying m rows, it expands to an m×n regular factor matrix λ:

[0077]

[0078] Combined with Γ, λ is improved to the regular factor matrix λ′ of the adaptively adjusted dimensional filter from the intensity domain, associating the global and local with each column:

[0079]

[0080] symbol Represents Hadamard.

[0081] Step 6: Correct the iteratively updated time domain mean image Combined with the horizontal one-dimensional weighted guided filter in the spatial domain, the adaptive filter radius r in step 4 and the adaptive regular factor matrix λ′ in step 5 are substituted to form a spatiotemporal adaptive one-dimensional filter. Subtract the output image q of the spatiotemporal adaptive one-dimensional filter to obtain the filtered image Will The residual stripe non-uniformity image extracted initially is as follows:

[0082] The output image q of the spatiotemporal adaptive one-dimensional filter is as follows:

[0083]

[0084] Where w represents a horizontal window of 1×(2r+1), represents the variance within the window w centered on pixel p′, a(p′) is the linear multiplication coefficient of the filter transformation of pixel p′, and b(p′) is the linear addition coefficient of the filter transformation of pixel p′. is the spatial mean of a(p′) in the window w centered at pixel p′, It is the spatial mean of b(p′) in the window w centered at pixel p′. It means that the pixel p′ is the center and the window w is The spatial mean of . q(p′) is the input image The filter output at pixel p′ is passed through w to traverse all pixels of the window image to obtain the output image q.

[0085] By inputting the image Subtract the output image q of the spatiotemporal adaptive one-dimensional filter to obtain the filtered image

[0086]

[0087] Will As the preliminary extracted residual stripe non-uniformity image.

[0088] Step 7: Initially extract the residual stripe non-uniformity image By increasing the constraint threshold TH, the inaccurate stripe non-uniformity estimation value at the edge of the scene is set to 0, and the constrained residual stripe non-uniformity image is obtained, which is denoted as right Calculate the non-zero spatial domain mean value by column to get a 1×n vector, copy m rows and expand it to an m×n residual stripe non-uniformity update image The details are as follows:

[0089] The residual stripe non-uniformity image extracted initially Due to the edge-preserving effect of the one-dimensional weighted guided filter, the edge of the scene is close to 0, which is an inaccurate stripe non-uniformity estimate. By increasing the constraint threshold TH, the inaccurate stripe non-uniformity estimate at the edge of the scene is set to 0, and the constrained residual stripe non-uniformity image is obtained, which is recorded as The error accumulation caused by iteration can be suppressed, so that the time-space domain adaptive one-dimensional filter has the ability to converge.

[0090]

[0091] For 8-bit or 16-bit images, TH=0.1 is usually set.

[0092] right Calculate the non-zero empty domain mean column by column to get a 1×n vector Represents the non-zero spatial domain mean.

[0093] It is expressed as:

[0094]

[0095] Representing images The non-zero spatial domain mean of the yth column, ||||0 represents the L0 norm, Representing images The number of non-zero elements in column y.

[0096] After copying m rows, the residual stripe non-uniformity image is expanded to m×n.

[0097]

[0098] The symbol repmat represents the vector replication operation, and [m 1] represents the replication of m rows and 1 column. Expanded to m×n

[0099] Step 8: Calculate the corrected image of the jth frame

[0100]

[0101] Example 1

[0102] In the embodiment, it is verified by Matlab R2020b on a computer with a 2.10 GHz processor and 16 GB memory. In order to better evaluate the effectiveness of the stripe non-uniformity correction method based on spatiotemporal adaptive one-dimensional filtering proposed in this paper, this section selects spectral shaping statistics (SSS) (Liu T, Sui X, Wang Y, et al. Strong non-uniformity correction algorithmbased on spectral shaping statistics and LMS[J].Optics Express, 2023, 31(19):30693-30709.), dual-domain filtering BP neural network (DDBP) (Liu Y, Qiu B, Tian Y, et al. Scene-based dual domain non-uniformity correction algorithm for stripe and optics-caused fixed pattern noise removal[J].Optics Express, 2024, 32(10):16591-16610.), and total variation BP neural network (TVBP) (Vera E, Meza P, Torres S. Total variation approach for adaptive nonuniformity correction in focal-plane arrays[J].Optics letters, 2011, 36(2): 172-174.), multi-scale wavelet decomposition and guided filtering (WDGF) (Cao Y, He Z, Yang J, et al. A multi-scale non-uniformity correction methodbased on wavelet decomposition and guided filtering for uncooled long wave infrared camera[J]. Signal Processing: Image Communication, 2018, 60: 13-21.), weighted guided filtering and multi-domain fusion (WGFMF) (Hong Y, Rao P, Zhou Y, et al. A weighted guided filter-based multi-domain fusion destriping method[J].IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2024.) A total of five stripe non-uniformity correction methods were compared. On 500 frames of the FLIR advanced driver assistance system thermal infrared public data set (16-bit data format, image resolution 512×640), Gaussian distribution stripe non-uniformity with a mean of 0 and a standard deviation of 10 was added, and the stripe non-uniformity intensity of each column was the same, simulating non-uniformity degraded images, and a comparative experiment was designed. The correction results of the 100th and 500th frames are shown in. Figure 3 The peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM) curve are used to quantify the evaluation curve. Figure 4 As shown. Figure 3 and Figure 4 As a result, the stripe non-uniformity correction method based on spatiotemporal adaptive one-dimensional filtering proposed by the present invention can converge within 100 frames, effectively correcting the stripe non-uniformity without generating ghost images. The PSNR of the last frame increased from 57.08dB to 78.04dB, and the SSIM increased from 0.537 to 0.997, both of which were higher than the other five stripe non-uniformity correction methods, and the single-frame processing speed can reach within 80ms, which can effectively show that the stripe non-uniformity correction method proposed by the present invention has excellent performance.

[0103] In summary, the present invention can take into account the generalization ability of scenes, correction accuracy and computational efficiency, has the ability to distinguish stripe non-uniformity and scene edges, and realizes adaptive correction of infrared image stripe non-uniformity independent of scene changes and content. Even when stripe non-uniformity and scene edges coincide, it can avoid the situation of stripe under-correction or stripe over-correction and edge blur.

Claims

1. A stripe non-uniformity correction method based on spatiotemporal adaptive one-dimensional filtering, characterized in that: Here are the steps: Step 1: Use an infrared imaging system to collect images with an image resolution of m×n; Step 2: For the first frame image z(1), the spatial mean is calculated by column, the column spatial mean vector of z(1) is obtained and normalized, and the high-frequency detail component is extracted by one-dimensional single-scale discrete wavelet transform for the normalized vector. After copying m rows, it is expanded to an m×n initial stripe non-uniformity correction image. Step 3: Use the infrared imaging system to continuously collect images and calculate the time domain mean image of the jth frame Using the iterative correction strategy, the iteratively updated fringe non-uniformity correction image is obtained. and the iteratively updated time domain mean correction image Initialize the fringe non-uniformity correction image and residual fringe non-uniformity update image Step 4: Correct the iteratively updated time domain mean image The MAD method is used to estimate the intensity of the overall fringe non-uniformity of the image, and the radius r of the filter is adaptively adjusted in the spatial domain; Step 5: Correct the iteratively updated time domain mean image The improved horizontal one-dimensional edge-aware weighting matrix Γ is used to improve the image The ability to resolve the non-uniformity of stripes and the edges of scenes; extracting images using two-dimensional single-scale discrete wavelet transform The vertical detail component matrix V is taken, and the median of the absolute values ​​of each column of the matrix V is taken as the estimated value of the non-uniformity intensity of each column of the stripes. Combined with Γ, the regularization factor matrix λ′ of the filter is adaptively adjusted from the intensity domain; Step 6: Correct the iteratively updated time domain mean image Combined with the horizontal one-dimensional weighted guided filter in the spatial domain, the adaptive filter radius r in step 4 and the regular factor matrix λ′ in step 5 are substituted to form a spatiotemporal adaptive one-dimensional filter. Subtract the output image q of the spatiotemporal adaptive one-dimensional filter to obtain the filtered image Will As a preliminary extracted residual fringe non-uniformity image; Step 7: Initially extract the residual stripe non-uniformity image By increasing the constraint threshold TH, the inaccurate stripe non-uniformity estimation value at the edge of the scene is set to 0, and the constrained residual stripe non-uniformity image is obtained, which is denoted as right Calculate the non-zero spatial domain mean value by column to get a 1×n vector, copy m rows and expand it to an m×n residual stripe non-uniformity update image Step 8: Calculate the corrected image of the jth frame 2. The stripe non-uniformity correction method based on spatiotemporal adaptive one-dimensional filtering according to claim 1, characterized in that: In step 2, for the first frame image z(1), the spatial mean is calculated by column, the column spatial mean vector of z(1) is obtained and normalized, as follows: For the first frame image z(1), calculate the spatial mean by column and obtain the column spatial mean vector of the first frame image z(1) represents a vector, the subscript y represents the number of columns, Represents the spatial mean operation; The spatial mean of the yth column of the first frame image z(1) It is expressed as: (x,y) represents the pixel position; right The normalized vector is recorded as The column spatial mean vector after normalization of the yth column It is expressed as:

3. The stripe non-uniformity correction method based on spatiotemporal adaptive one-dimensional filtering according to claim 2, characterized in that: In step 2, A one-dimensional single-scale discrete wavelet transform is used to extract high-frequency detail components, and the initial stripe non-uniformity correction image is expanded to m×n after copying m rows. The details are as follows: Pair Vector One-dimensional single-scale discrete wavelet transform is used to extract high-frequency detail component vectors Represents estimated value, DWT 1D represents one-dimensional single-scale discrete wavelet transform operation, sym8 represents Symlets8 wavelet basis; Represents the high-frequency detail component vector; to maintain the high-frequency detail component vector The length of is consistent with the image column width n, DWT 1D No downsampling; The high-frequency detail component vector The initial fringe non-uniformity correction image is expanded to m×n after copying m rows. The symbol repmat stands for the vector replication operation, and [m 1] stands for replicating m rows and 1 column, converting the 1×n high-frequency detail component vector The initial fringe non-uniformity correction image expanded to m×n 4. The stripe non-uniformity correction method based on spatiotemporal adaptive one-dimensional filtering according to claim 3 is characterized in that: In step 3, the infrared imaging system is used to continuously collect images and the time domain mean image of the jth frame is counted. Using the iterative correction strategy, the iteratively updated fringe non-uniformity correction image is obtained. and the iteratively updated time domain mean correction image Initialize the fringe non-uniformity correction image and residual fringe non-uniformity update image The details are as follows: Use the infrared imaging system to continuously collect images and count the time domain mean image of the jth frame represents the time domain mean operation; Using the iterative correction strategy, the iteratively updated fringe non-uniformity correction image is obtained. and the iteratively updated time domain mean correction image represents the fringe non-uniformity corrected image of the jth frame, Represents the residual fringe non-uniformity update image of the (j-1)th frame, and initializes the fringe non-uniformity correction image And the residual fringe non-uniformity update image 5. The stripe non-uniformity correction method based on spatiotemporal adaptive one-dimensional filtering according to claim 4, characterized in that: In step 4, the iteratively updated time domain mean correction image The MAD method is used to estimate the intensity of the overall fringe non-uniformity of the image, and the radius r of the filter is adaptively adjusted from the spatial domain, as follows: S4-1, calculate the iteratively updated time domain mean correction image Spatial mean of each column: in, Represents the iteratively updated time domain mean correction image The spatial mean of the yth column is represented by the vector S4-2, Calculation Image The forward difference of the means of two adjacent columns constitutes the forward difference vector S4-3, calculate the MAD value to exclude Outliers in: med stands for median operation; S4-4, removed The vector after the outlier is recorded as S4-5, estimate the iteratively updated time domain mean correction image The overall fringe inhomogeneity intensity σ s : std stands for standard deviation operation; S4-6, according to the confidence interval of the normal distribution, the radius r of the filter is adaptively adjusted from the spatial domain: in, Represents rounding up.

6. The stripe non-uniformity correction method based on spatiotemporal adaptive one-dimensional filtering according to claim 5, characterized in that: In step 5, the iteratively updated time domain mean correction image An improved one-dimensional edge-aware weighting matrix Γ in the horizontal direction is used, as follows: Where L is The dynamic range, ε represents the sensitivity coefficient, and K represents the proportionality coefficient; Represents the time domain mean correction image updated in iteration The variance of the 1×3 local window centered at pixel p′; represents the variance of a 1×3 local window centered on pixel p. Pixel p represents All pixels of ; Γ(p′) represents the one-dimensional edge-aware weighting in the horizontal direction at p′; In order to associate the local and the global, Γ(p′) is used The variance of all pixels in their respective 1×3 local window is normalized.

7. The stripe non-uniformity correction method based on spatiotemporal adaptive one-dimensional filtering according to claim 6, characterized in that: In step 5, the image is extracted using a two-dimensional single-scale discrete wavelet transform The vertical detail component matrix V is taken, and the median of the absolute values ​​of each column of the matrix V is taken as the estimated value of the non-uniformity intensity of each column of the stripes. Combined with Γ, the regular factor matrix λ′ of the filter is adaptively adjusted from the intensity domain, as follows: Image extraction using two-dimensional single-scale discrete wavelet transform The vertical detail component matrix V of: DWT 2D represents a two-dimensional single-scale discrete wavelet transform, haar represents the Haar wavelet basis, and V represents the detail component matrix in the vertical direction; in order to keep the matrix V and the image The resolution of DWT is consistent 2D No downsampling; Take the column vectors of matrix V The median of the absolute values ​​is used as an estimate of the inhomogeneity intensity of each row of fringes: represents the y-th column vector of the matrix V; σ y represents the estimated standard deviation of the non-uniformity intensity of the y-th column stripes; y represents the regularization factor of the yth column, and the vector is expressed as Vector After copying m rows, it expands to an m×n regular factor matrix λ: Combined with Γ, λ is improved to the regular factor matrix λ′ of the adaptively adjusted dimensional filter from the intensity domain, associating the global and local with each column: symbol Represents Hadamard.

8. The stripe non-uniformity correction method based on spatiotemporal adaptive one-dimensional filtering according to claim 7, characterized in that: Step 6: Correct the iteratively updated time domain mean image Combined with the horizontal one-dimensional weighted guided filter in the spatial domain, the adaptive filter radius r in step 4 and the adaptive regular factor matrix λ′ in step 5 are substituted to form a spatiotemporal adaptive one-dimensional filter. Subtract the output image q of the spatiotemporal adaptive one-dimensional filter to obtain the filtered image Will The residual stripe non-uniformity image extracted initially is as follows: The output image q of the spatiotemporal adaptive one-dimensional filter is as follows: Where w represents a horizontal window of 1×(2r+1), represents the variance within the window w centered on pixel p′, a(p′) is the linear multiplication coefficient of the filter transformation of pixel p′, and b(p′) is the linear addition coefficient of the filter transformation of pixel p′; is the spatial mean of a(p′) in the window w centered at pixel p′, is the spatial mean of b(p′) in the window w centered at pixel p′; It means that the pixel p′ is the center and the window w is The spatial mean of ; q(p′) is the input image The filter output at pixel p′ is passed through w to traverse all pixels of the window image to obtain the output image q; By inputting the image Subtract the output image q of the spatiotemporal adaptive one-dimensional filter to obtain the filtered image Will As the preliminary extracted residual stripe non-uniformity image.

9. The stripe non-uniformity correction method based on spatiotemporal adaptive one-dimensional filtering according to claim 8, characterized in that: In step 7, the details are as follows: The residual stripe non-uniformity image extracted initially By increasing the constraint threshold TH, the inaccurate stripe non-uniformity estimation value at the edge of the scene is set to 0, and the constrained residual stripe non-uniformity image is obtained, which is denoted as right Calculate the non-zero empty domain mean column by column to get a 1×n vector Represents the non-zero spatial domain mean; It is expressed as: Representing images The non-zero spatial domain mean of the yth column, || ||0 represents the L0 norm, Representing images The number of non-zero elements in column y; After copying m rows, the residual stripe non-uniformity image is expanded to m×n. The symbol repmat represents the vector replication operation, and [m 1] represents the replication of m rows and 1 column. Expanded to m×n 10. The stripe non-uniformity correction method based on spatiotemporal adaptive one-dimensional filtering according to claim 9, characterized in that: In step 8, the corrected image of the jth frame is calculated

Images