Short-wave infrared image non-uniform stripe noise correction algorithm

Abstract

The invention discloses a short-wave infrared image non-uniform stripe noise correction algorithm, belongs to the technical field of short-wave infrared image processing, and designs improved adaptive fractional order total variation processing, an adaptive order is constructed by eight neighborhood gradients, and a fidelity term of fractional order gradient is added. And the generation of a gradient effect is avoided while stripe noise is removed. In order to further compare the advantages of the algorithm provided by the invention, a visual image and a series of quantitative indexes are used for evaluation. The result shows that the algorithm is excellent in visual effect, objective indexes and the like. According to the algorithm, stripe noise is quickly and effectively removed under the low calculation load, details are accurately recovered, a new way is provided for non-uniformity correction of the image, and construction of a non-uniformity correction system is further promoted.

Description

Technical Field

[0001] The invention relates to the technical field of short-wave infrared image processing, and in particular to a short-wave infrared image non-uniform stripe noise correction algorithm. Background Art

[0002] Infrared linear scanning imaging systems have the characteristics of high sensitivity and strong reliability, and are widely used in military and civilian fields. However, differences in factors such as detector array materials and manufacturing processes lead to inconsistent pixel responses during the detection process, which in turn causes deviations in the imaging signal and results in non-uniform noise in the output image in the form of stripes. This phenomenon is particularly evident in remote sensing hyperspectral and infrared detectors.

[0003] In order to obtain high-quality images, many classic stripe non-uniformity correction methods have been proposed. These methods mainly include: scene-based methods, transform domain frequency filtering methods, spatial domain statistical calculation methods, and stripe removal based on the variational framework. The scene-based method is essentially a relative radiation calibration. Different detection units should produce the same output signal value for the same input radiation value. The actual output value under the standard scene is calculated by the correlation between the uniform scene to extract the corresponding calibration coefficient. Finally, the image is adjusted by the calculated coefficient to eliminate the stripe noise. Considering that the calibrated calibration coefficient changes over time, in order to avoid the emergence of new stripe non-uniformity, it needs to be updated periodically. The transform domain method combines the spatial domain and frequency domain of the image, uses Fourier or wavelet transform to transform the image from the spatial domain to the frequency domain, and sets the corresponding filter according to the corresponding frequency of the stripe noise to filter it out. Fan et al. estimated the real scene radiation through a one-dimensional Gaussian filter with low complexity, and then established a weighted mean square error optimization function to find the optimized correction parameters, and used a variant of Landweber iteration to adaptively adjust the correction parameters. Münch B et al. developed a new stripe removal filter based on wavelet decomposition and FFT transform. The classical domain-based method is difficult to distinguish the frequency of stripe noise and normal images, which can easily cause information loss and image blur. The more classical methods of spatial domain statistical calculation include histogram matching, equalization method and moment matching. Zhang Bingxian et al. integrated local algorithms such as surface filtering and local least squares fitting into the histogram matching process. This type of method is based on the assumption that the relationship between the sub-images of the detector pixels is fixed, so there are limitations in the processing of complex stripes. The method under the basis variation framework constructs energy functionals through reasonable mathematical characterization and uses optimization algorithms to complete numerical solutions. Qian et al. constructed a constrained nonlinear optimization equation to minimize the line gradient energy of the image. By solving the optimization equation, the non-uniformity parameters of each stripe are obtained to achieve correction. Fehrenbach et al. proposed a Bayesian framework to remove stationary noise in images. Bouali M et al. introduced the single total variation model (UTV) to achieve the correction of noisy images by observing the spatial characteristics of stripe noise. The TV model is prone to the tendency of step effect in actual application and may cause excessive smoothing of the image because it more effectively retains uniform and small curvature edges. The total variation model has undergone many improvements. The generalized total variation model (TGV) model that uses both first-order and second-order regularization terms has been proposed, but it still produces a step effect. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a short-wave infrared image non-uniform stripe noise correction algorithm. The present invention introduces parameters into the regularization term by constructing an adaptive fractional order, and adds a fractional order gradient fidelity term. The wavelet transform is combined with the improved total variation method to establish an adaptive fractional order total variation image denoising model, which fully removes stripe noise while retaining detail information and greatly eliminates the influence of the global gradient effect. The content of the present invention is as follows:

[0005] The object of the present invention is to provide a short-wave infrared image non-uniform stripe noise correction algorithm, the technical point of which is that it comprises the following steps:

[0006] Step 1: Decompose the original noise image into a low-frequency information layer LA, a horizontal high-frequency information layer LH, a vertical high-frequency information layer LV and a diagonal high-frequency information layer LD through the first wavelet transform of DB1-SWT2;

[0007] Step 2: Reset the horizontal high-frequency information layer LH and the diagonal high-frequency information layer LD in step 1 to 0 and then reconstruct the image for the first time with the low-frequency information layer LA and the vertical high-frequency information layer LV by using inverse wavelet transform;

[0008] Step 3, processing the image reconstructed in step 2 using an adaptive fractional-order total variation model;

[0009] Step 4, perform DB1-SWT2 wavelet transform on the image processed in step 3 again to decompose it into low-frequency information layer LA', horizontal high-frequency information layer LH', vertical high-frequency information layer LV' and diagonal high-frequency information layer LD'. The horizontal high-frequency information layer LH and diagonal high-frequency information layer LD in step 1 are refined with the low-frequency information layer LA' and vertical high-frequency information layer LV' and reconstructed by a second wavelet inverse transform, which is the final corrected image.

[0010] In order to better solve the above technical problems, the processing formula of the fractional-order total variation model in step 3 of the short-wave infrared image non-uniform stripe noise correction algorithm of the present invention is:

[0011] Where α is the adaptive fractional order, u is the corrected image, u0 is the observed image, G is the Gaussian convolution kernel, ▽ is the gradient operator, and μ is a constant.

[0012] In order to better solve the above technical problems, the processing formula of the adaptive fractional order α in the processing formula of the fractional order total variation model in the short-wave infrared image non-uniform stripe noise correction algorithm of the present invention is:

[0013] Where k is a custom constant and F(i,j) is the gradient value of the eight-neighborhood.

[0014] Compared with the prior art, the shortwave infrared image non-uniform stripe noise correction algorithm of the present invention can achieve the following beneficial effects:

[0015] An algorithm for correcting non-uniform stripe noise in short-wave infrared images of the present invention designs an improved adaptive fractional-order total variation processing, constructs the adaptive order with eight-neighborhood gradients, and adds the fidelity term of the fractional-order gradient, thereby avoiding the generation of gradient effects while removing stripe noise. In order to further compare the advantages of the algorithm of the present invention, visual images and a series of quantitative indicators were used for evaluation. The results show that the algorithm of the present invention performs well in terms of visual effects and objective indicators. The algorithm of the present invention quickly and effectively removes stripe noise and accurately restores details under a low computational load, providing a new way for image non-uniformity correction and helping to further promote the establishment of a non-uniformity correction system. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0017] Figure 1 It is a wavelet transform block diagram of the present invention;

[0018] Figure 2 It is a two-dimensional discrete steady-state wavelet transform flow chart of the present invention;

[0019] Figure 3 For the DB1-SWT2 decomposition of the present invention;

[0020] Figure 4 is a stripe noise decomposition diagram of the present invention;

[0021] Figure 5 is the neighborhood gradient map of the present invention;

[0022] Figure 6 It is a neighborhood schematic diagram of the present invention;

[0023] Figure 7 The algorithm flow of the present invention is as follows;

[0024] Figure 8 This is a comparative test of different α values ​​for the present invention;

[0025] Fig. 9 This is a real fringe image experiment of the present invention;

[0026] Fig.10 For algorithm comparison of the present invention;

[0027] Fig.11The color diagram of the partial image before and after the processing of the present invention;

[0028] Fig.12 is the evaluation index of each algorithm of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. 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 making creative work are within the scope of protection of the present invention.

[0030] The technical solutions provided by various embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0031] To achieve the above objectives, the present invention provides an algorithm based on wavelet transform adaptive fractional order total variation function minimization, which is applied to remove non-uniform stripe noise in infrared images. The specific implementation includes:

[0032] 1. Model Architecture

[0033] 1.1 Two-dimensional discrete steady-state wavelet transform

[0034] As a global transform, Fourier transform provides global information of the signal in the frequency domain, but cannot effectively analyze non-stationary signals, and has limitations in processing non-stationary signals and capturing mutation characteristics. Wavelet transform provides a joint analysis method in the time and frequency domains, which will produce low-resolution coefficients during decomposition. The decomposition coefficients are generated by calculating the mean and gradient in the horizontal and vertical directions. Therefore, it is more effective in processing non-stationary and complex signals, and has advantages such as good time-frequency localization capability, multi-scale analysis, and flexibility.

[0035]

[0036] Where WT(a,τ) is the coefficient of wavelet transform, f(t) is the original signal, Ψ * is the complex conjugate of the wavelet function Ψ, a is the scale parameter, and τ is the translation parameter.

[0037] The decomposition process of wavelet transform can mine the gradient characteristics of infrared image noise without losing any original information in the decomposition process, retaining the integrity of the image signal to the greatest extent, and can directly restore the complete image from the coefficients, avoiding the introduction of computational noise. Figure 1As shown in the figure, the image is first decomposed by row, and the mean and difference of two adjacent pixels in the horizontal direction are calculated, where the mean represents the low-frequency information in the image, and the difference represents the high-frequency information in the image. Then the upper part of the low-frequency information and high-frequency information is decomposed in the column direction to solve the low-frequency information. The difference of adjacent pixels in the column direction is calculated for the lower part of the low-frequency information and high-frequency information to obtain the high-frequency information, and finally four wavelet subbands are generated. The LA subband retains the image structure and detail information that is most similar to the original image data. The LV subband is very similar to the horizontal gradient of the image. The stripe noise contained in the infrared image is further highlighted, which helps to reflect the noise-related features. The LH subband is highly complementary to the LV subband. It contains detail information that is very similar to the LV subband, but the noise information is not reflected in this subband. Therefore, the LV and LH subbands can refer to each other and provide noise and detail information to each other, so as to suppress noise while maintaining image details. The LD subband contains the second-order gradient information of the image, reflects the sharpest image details, and helps the auxiliary network to complete the reconstruction and restoration of high-frequency image texture details.

[0038] like Figure 2 The present invention uses the SWT2 algorithm to perform two-dimensional discrete steady-state wavelet transform, aiming to achieve more sophisticated signal analysis. Compared with the traditional discrete wavelet transform (DWT), the SWT2 algorithm maintains the original sampling rate of the signal during the transformation process, without the need for downsampling. This feature ensures that the image after SWT2 transformation is consistent with the original image in size, so that multi-scale analysis can be performed more accurately.

[0039] The DB series wavelet basis functions are compactly supported orthogonal wavelets, which help reduce redundant information when decomposing signals or images, and have good orthogonality and compact support. At the same time, experimental verification shows that the slight error of db1 wavelet reconstruction is much smaller than other wavelets in magnitude. Therefore, the present invention selects db1 wavelet as the wavelet basis function of the algorithm.

[0040] Stripe noise has significant directionality. The two-dimensional stationary discrete wavelet transform (DB1Two-Dimensional Discrete StationaryWaveletTransform) based on DB1 wavelet shows strong high-frequency information extraction capabilities in the horizontal, vertical, and diagonal directions, and can maintain the pattern structure of the low-frequency sub-band, which helps to completely separate the part containing noise components during the non-uniformity correction process. Figure 3The DB1-SWT2 decomposition results of the short-wave infrared image collected by the Resource-1 02D (ZY1E) satellite are shown. It can be seen that the vertical stripe noise shows obvious response in the low-frequency information layer LA and the vertical high-frequency information layer LV, while the horizontal high-frequency information layer LH and the diagonal high-frequency information layer LD mainly preserve some details in the scene. Therefore, it is feasible to separate and reprocess each component. Figure 4 DB1-SWT2 decomposition was performed on multiple groups of infrared images containing different ground object information. The results showed that vertical stripe noise exists in low-frequency components and vertical high-frequency components with extremely high universality, providing strong support for subsequent algorithm decomposition processing.

[0041] 1.2 Adaptive TV Model

[0042] Rudin, Osher and Fatemi first proposed a nonlinear total variation image denoising model in 1992 and derived a constrained minimization algorithm, which is a time-dependent nonlinear partial differential equation in which the constraints are determined by noise information. The basis of the TV model is that the total variation of a noisy image is much larger than the total variation of an image without noise, and image denoising can be achieved by solving the minimum overall variation of the image.

[0043] The first step is to build a model: Let Ω be a bounded open subset in Rn, the function u is locally integrable, and the total variation is defined as:

[0044]

[0045] Through Lagrange number multiplication, the Lagrange multiplier λ is introduced, and the total variation model defines the minimum form of the energy functional as:

[0046]

[0047] The gradient ▽u is a vector whose components are the rates of change of u in all directions, u is the corrected image, u0 is the observed image, and Ω is the domain.

[0048] The first term is the regularization term (i.e. the total variation defined), which is equivalent to the sum of the changes, describing the smoothness of the image, and its function is to remove noise in the minimization process; the second term is the fidelity term, which ensures that the degradation process is met without changing the original image structure, and its function is to make the denoised image as close to the original image as possible. λ is used to balance the regularization term and the fitting term. is a first-order difference operation.

[0049] The energy functional specifically refers to the integration of the squared differential norm. If the true solution of the functional is known, then this solution must be the minimum value of the energy functional. The idea of ​​the Euler equation is to transform the function differentiation problem into a single variable differentiation. Based on the principle of variational calculus, the Euler-Lagrange equation is derived:

[0050]

[0051] The solution to the extreme value of the energy functional is obtained through the Euler-Lagrange equation corresponding to the functional, which can quickly and effectively handle linear equation problems. For complex nonlinear equations, it is impossible to use them directly without processing. The time variable is introduced for numerical calculation, and the static problem is transitioned to the dynamic problem. When the equation evolution is completed, a steady-state solution can be obtained, which is the gradient descent method. The gradient descent method will be converted into the following equation:

[0052]

[0053] is the gradient operator, u i，j Represents the gray value of the image u at the pixel point, represents the nth iteration value, dt is the time step, the partial derivative of the image at the pixel point (i, j) is replaced by the difference, and the derivatives in each direction are obtained:

[0054]

[0055] Where (u x ) ij represents the first-order difference of u along the x direction at the pixel point (i, j), (u y ) ij is the first-order difference along the y direction, (u xx ) ij and(u yy ) ij are the second-order differences along the x and y directions, respectively, (u xy ) ij is the mixed second-order difference along the x-direction and the y-direction.

[0056] The divergence calculation formula is:

[0057]

[0058] As the number of iterations increases, the corrected image can be finally obtained, and the equation is as follows:

[0059]

[0060] represents the nth iteration value, Δt is the time step, and ▽ is the gradient operator.

[0061] The regularization term in the TV model can effectively remove noise, but it is easy to over-smooth, produce blocky shadows, and have a staircase effect. Bai et al. proposed a fractional-order anisotropic diffusion equation that can eliminate the staircase effect, but it cannot be adaptively adjusted according to image features. The present invention introduces parameters into the regularization term through an adaptive fractional order constructed by eight-neighborhood gradients, adds a fractional-order gradient fidelity term, and establishes an improved adaptive fractional-order total variation image denoising model, which greatly eliminates the influence of the global gradient effect while removing stripe noise.

[0062]

[0063] The model is shown in formula (10), the first term F α (u) is the adaptive fractional-order regularization term constructed, which is used to remove noise according to the gradient information of each region neighborhood in the process of minimizing the energy functional. The third term is the fractional-order gradient term, with a constant value of μ, which makes the gradient change of the denoised image close to This further smoothes the denoised image and improves the staircase effect. The model is transformed into:

[0064]

[0065] Where λ is the Lagrange multiplier, u is the corrected image, u0 is the observed image, G is the Gaussian convolution kernel, ▽ is the gradient operator, and μ is a constant.

[0066] The algorithm of the present invention takes into account the structural complexity of the local area of ​​the image, and by setting the adaptive fractional order, the differential order can be selected according to the characteristics of different image areas. Figure 5 From the neighborhood gradient map, we can see that the eight-neighborhood gradient values ​​pay more attention to image edge and texture detail information, and the edge and texture areas in the image are more obvious.

[0067] like Figure 6 As shown, the pixel eight-neighborhood analysis feature information is selected, and the eight-neighborhood gradient average calculation formula (12) is defined for any pixel point P(i, j). The fractional order is adaptively determined by the gradient information of the image pixel. The experimental verification shows that the value formula is feasible and achieves good results. The adaptive fractional order is constructed as shown in formula (13), where k is a custom constant.

[0068]

[0069]

[0070] The model is regarded as a problem of minimizing energy functional. The solution is to obtain the Euler-Lagrange equation based on the variational principle, and use the gradient descent idea to obtain the corresponding gradient descent flow diffusion formula. Then, the discretized iterative formula is obtained by numerical solution. The final iterative result is the required corrected image.

[0071] 2 Algorithm Flow

[0072] Through the study of non-uniformity correction technology and the analysis of wavelet transform, the present invention separates the components containing non-uniform noise, removes the non-uniform noise and protects the other two detail components from being blurred. It uses an improved adaptive fractional-order total variation model to perform non-uniformity correction on the wavelet components containing only noise, and solves the gradient effect with fractional-order differential constraints. The specific algorithm steps and algorithm flow are as follows: Figure 7 shown.

[0073] Step 1: Decompose the original noise image into the following: Figure 3 The low frequency information layer LA, the horizontal high frequency information layer LH, the vertical high frequency information layer LV and the diagonal high frequency information layer LD are shown;

[0074] Step 2: Reset the horizontal high-frequency information layer LH and the diagonal high-frequency information layer LD in step 1 to 0 and then reconstruct the image for the first time with the low-frequency information layer LA and the vertical high-frequency information layer LV by using inverse wavelet transform;

[0075] Step 3, processing the image reconstructed in step 2 using a fractional-order total variation model;

[0076] Step 4, perform DB1-SWT2 wavelet transform on the image processed in step 3 again to decompose it into low-frequency information layer LA', horizontal high-frequency information layer LH', vertical high-frequency information layer LV' and diagonal high-frequency information layer LD'. The horizontal high-frequency information layer LH and diagonal high-frequency information layer LD in step 1 are reconstructed with the low-frequency information layer LA' and vertical high-frequency information layer LV' through a second inverse wavelet transform to obtain the final corrected image.

[0077] 3 Technical Effects

[0078] 3.1 Adaptive fractional order experiment

[0079] The present invention uses different fractional order α values ​​to conduct comparative experimental processing on infrared images. Specifically, the α values ​​are set to 1.2, 1.6 and 2 respectively. The experimental results are shown in Figure 2. Figure 8As shown. It can be observed from the figure that when the α value is set to 1.2, as shown in group (a), the removal of some stripe noise in the overall image is not good; when α is set to 1.6, as shown in group (b), the removal of the detail stripe noise in the red frame is not thorough; in addition, when α is set to 2, it can be clearly seen that in the enlarged image of the red frame in group (c), especially the edge details, there is a gradient blur phenomenon, and the detail information is seriously lost. The adaptive fractional order value formula set by the algorithm of the present invention is applied in image processing. This innovative formula adaptively sets the fractional order according to the regional neighborhood characteristics, and retains the details according to the regional characteristics while removing the stripes, as shown in the enlarged image of the red frame in group (d), which further confirms the applicability of the algorithm and shows the importance of the adaptive α value of the algorithm of the present invention.

[0080] 3.2 Real Image Data

[0081] The robustness of the algorithm is verified by experimentally processing multiple sets of real stripe noise images of the short-wave infrared of Resource One 02D (ZY1E) satellite, involving a variety of complex scenes. The experimental results are as follows Fig. 9 The component graph obtained by the image DB1-SWT2 transformation is brought into the algorithm model of the present invention for processing, and then each component is reconstructed. Fig. 9 From the comparison pictures marked with red frames in groups (b) and (d), it can be seen that the algorithm processing results have obviously removed the non-uniform stripe noise, and from the 1st and 4th pictures in group (d), it can be seen that the visual effect of the model has been enhanced, and the image details and edge structure information are well preserved.

[0082] In addition, the present invention also conducts comparative experiments on real stripe noise images using non-local mean filtering (NLM), adaptive neighborhood bandpass filtering (ANBF), total variation method (TV) and the algorithm of the present invention. The experimental results are shown in Figure 2. Fig.10 As shown in the figure, after comprehensive evaluation of comparative experiments, NLM has a certain effect in removing stripe noise, but there is an obvious problem of detail information loss, and its overall performance is not satisfactory. ANBF shows good stripe removal ability in specific areas, but in areas with a large depth of field (as shown in the red box diagram of group c), it not only introduces new noise, but also produces artifacts, which significantly reduces its practical application value. The TV algorithm maintains details by strengthening the edges of the image, but the gradient effect (as shown in the box diagram of group d) and excessive blurring effect it produces make the overall visual effect of the image appear unsatisfactory. The algorithm of the present invention solves the problem of detail loss by layered processing, and takes into account both noise removal effect and information retention by establishing an adaptive fractional-order TV model. Therefore, the algorithm proposed in the present invention shows significant advantages in removing stripe noise and preserving detail information, and has achieved better experimental results and visual effects than other algorithms.

[0083] 3. Evaluation and analysis

[0084] 3.1 Subjective evaluation

[0085] Subjective evaluation mainly involves visual observation and comparison of processed images, considering two aspects:

[0086] (1) Observe the effect of removing noise in the overall image area;

[0087] (2) Observe the preservation effect of high-frequency areas such as image texture details and edges.

[0088] In order to more fully and intuitively demonstrate the effect of the correction technology of the algorithm of the present invention on improving image quality, by comparing the color images before and after correction, the effectiveness of noise removal and detail retention is intuitively revealed.

[0089] from Fig.11 It can be seen from the direct comparison that after careful processing by the algorithm flow proposed by the present invention, the vertical stripe noise in the original image is removed, and the removal effect is significant. At the same time, the details and high-frequency information of each part of the image are fully preserved, showing extremely high clarity and rich details. The high-frequency and low-frequency junction areas of the image also achieve a smooth transition effect, further improving the overall perception and quality of the image.

[0090] 3.2 Objective evaluation

[0091] The present invention adopts information entropy, structural similarity (SSIM), mean square error (MSE), peak signal-to-noise ratio (PSNR) and the average value of pixel difference (D) as objective evaluation criteria, which can effectively eliminate the interference of human factors and evaluate the image processing results more comprehensively.

[0092] (1) Information entropy

[0093] For images, the wider the grayscale distribution, the richer the details, and the greater the information entropy. Information entropy is an important indicator to measure the richness of image information. It can be used to measure whether the image processing results retain edge and texture detail information well. The larger the value of information entropy, the richer the image features retained by the processing results, and the better the denoising effect.

[0094] The calculation formula is as follows, where I(i, j) is the frequency of the feature binary (i, j) appearing in the image, P ij Represents the probability of the occurrence of the feature binary group (i, j).

[0095]

[0096]

[0097] (2)MSE

[0098] The root mean square error is an important evaluation index for image processing. It measures the average of the squares of the difference between the true value and the predicted value, and intuitively reflects the closeness between the predicted image and the original image. The smaller the MSE value, the smaller the deviation between the predicted value and the true value, which reflects the higher the similarity between the images. The calculation formula is shown in the following formula (9):

[0099]

[0100] (3)PSNR

[0101] Peak signal-to-noise ratio (PSNR) is a parameter used to evaluate image quality. It assesses the degree of distortion that an image may suffer during processing and can reflect the impact of noise on the image. The larger the peak signal-to-noise ratio of an image, the better the image quality. Its expression is as follows: Max I Indicates the maximum pixel value.

[0102]

[0103] (4)SSIM

[0104] The Structural Similarity Index (SSIM) is an indicator used to measure the degree of structural change between two images. Taking into account structural information, SSIM not only considers the difference between pixel values, but also comprehensively considers the structural information that is crucial to image quality.

[0105] Therefore, SSIM can evaluate image quality more comprehensively. The higher the SSIM value, the higher the structural similarity between the two images, which reflects the better image quality. I , δ I , is image I and image The mean and standard deviation of is the covariance between the two graphs, and c1 and c2 are constant terms used to maintain stability.

[0106]

[0107] (5)D

[0108] The pixel difference average D measures the average absolute difference between the corresponding pixel values ​​of the original image and the processed image, reflecting the changes that may occur in the image during the processing. The smaller the D value, the smaller the difference between the processed image and the original image, thus retaining more original information. The calculation formula is as follows:

[0109]

[0110] Depend on Fig.12It can be seen that the ANBF algorithm causes great damage to the structure and edge information of the image during the stripe removal process, while the TV method and the algorithm proposed in the present invention retain the important structural features of the image, with a higher SSIM value and a lower pixel difference mean D. The algorithm of the present invention has obvious advantages in PSNR value compared with the other three algorithms, and the noise removal effect is significant. At the same time, the information entropy value also shows that the image information is better preserved.

[0111] Table 1 Evaluation indicators

[0112]

[0113] According to the evaluation index data in Table 1, the algorithm proposed in the present invention shows significant advantages in image processing. In terms of information entropy, the algorithm of the present invention is higher than the other three algorithms, which fully proves the richness of the algorithm in retaining image features. At the same time, in terms of MSE (mean square error) and PSNR (peak signal-to-noise ratio), the key indicators for measuring image quality, the performance of the algorithm of the present invention is far better than other algorithms, which strongly illustrates the significant effect of the algorithm in removing stripe noise. In addition, through the high score of the algorithm of the present invention in the SSIM (structural similarity metric) index, it can be seen that the algorithm of the present invention shows excellent performance in retaining the original image structure information, and the lower pixel difference mean also confirms the ability to protect the image detail information in the process of removing stripe noise. This further proves that the algorithm can effectively remove stripe noise and restore the image detail information on the basis of maintaining the edge contour structure of the image. In summary, the algorithm of the present invention shows significant effects in processing image stripe non-uniformity correction, not only has a wide range of applications, but also can achieve efficient image processing with a low computational load. The algorithm proposed in the present invention can effectively remove stripe noise and accurately restore the image detail information while ensuring image quality.

[0114] 4. Summary

[0115] Aiming at the problem of stripe non-uniformity in push-broom imaging system, and considering that the traditional non-uniformity correction method directly processes the image, it will greatly damage the information of the image and cause the loss of high-frequency information. The present invention proposes an innovative non-uniformity correction algorithm based on minimization of adaptive fractional-order total variation function of wavelet transform. Firstly, the characteristics of wavelet transform and stripe noise are studied, and the two-dimensional discrete steady-state wavelet transform is deeply studied through theoretical analysis. The control experiment of wavelet basis function and decomposition function is set to screen out the appropriate basis function and decomposition function. The image is processed by layering using wavelet transform. The study shows that stripe noise has significant directional characteristics. The vertical stripe noise exists in the low-frequency information layer and the vertical high-frequency information layer after decomposition. Therefore, the targeted processing of these two components can well retain the contour information of the horizontal high-frequency and diagonal high-frequency information layers. The components with stripe non-uniformity are processed by inverse wavelet transform, and the generated image contains all the noise. The present invention designs an improved adaptive fractional-order total variation processing, constructs the adaptive order with eight-neighborhood gradient, and adds the fidelity term of fractional-order gradient, so as to avoid the generation of gradient effect while removing stripe noise. In order to further compare the advantages of the algorithm of the present invention, visual images and a series of quantitative indicators were used for evaluation. The results show that the algorithm of the present invention performs well in terms of visual effects and objective indicators. The algorithm of the present invention can quickly and effectively remove stripe noise and accurately restore details under a low computational load, providing a new way for image non-uniformity correction and helping to further promote the establishment of a non-uniformity correction system.

[0116] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A shortwave infrared image non-uniform stripe noise correction algorithm, characterized in that: The following steps are involved: Step 1: Decompose the original noise image into a low-frequency information layer LA, a horizontal high-frequency information layer LH, a vertical high-frequency information layer LV and a diagonal high-frequency information layer LD through the first wavelet transform of DB1-SWT2; Step 2: Reset the horizontal high-frequency information layer LH and the diagonal high-frequency information layer LD in step 1 to 0 and then reconstruct the image for the first time with the low-frequency information layer LA and the vertical high-frequency information layer LV by using inverse wavelet transform; Step 3, processing the image reconstructed in step 2 using a fractional-order total variation model; Step 4, perform DB1-SWT2 wavelet transform on the image processed in step 3 again to decompose it into low-frequency information layer LA', horizontal high-frequency information layer LH', vertical high-frequency information layer LV' and diagonal high-frequency information layer LD'. The horizontal high-frequency information layer LH and diagonal high-frequency information layer LD in step 1 are refined with the low-frequency information layer LA' and vertical high-frequency information layer LV' and reconstructed by a second wavelet inverse transform, which is the final corrected image.

2. The shortwave infrared image non-uniform stripe noise correction algorithm according to claim 1, characterized in that: The processing formula of the fractional-order total variation model in step 3 is: Adapt to the fractional order, u is the corrected image, u0 is the observed image, and G is the defined Gaussian convolution kernel.

3. The shortwave infrared image non-uniform stripe noise correction algorithm according to claim 2, characterized in that: The processing formula of the adaptive fractional order α in the processing formula of the fractional order total variation model is: is a defined constant, and F(i,j) is the gradient value of the eight-neighborhood.

Images