Color image restoration method based on singular value factor
By integrating color images into unified tensors and using singular value decomposition and adaptive Tubal rank adjustment technology, the problems of computational complexity and structural damage in color image recovery are solved, and efficient and stable color image recovery is achieved, maintaining the image overall structure and consistency.
Patent Information
- Application Number
- CN202411156582.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-22
- Publication Date
- 2024-11-26
AI Technical Summary
The existing image recovery method based on matrix decomposition has high computational complexity when processing color images, which may lead to color distortion and inconsistency, and destroy the overall structure of color images, resulting in a decrease in recovery quality.
Using a color image recovery method based on singular value factor, the three color channels of the color image are integrated into a unified tensor, and the image recovery model is constructed through the Frobenius norm using random tensor singular value decomposition and adaptive Tubal rank adjustment technology. , and combine the ADM method to transform multivariate problems into univariate problems, simplifying the calculation process.
The image recovery quality is improved, the calculation workload is reduced, the overall structural consistency of the color image is maintained, the recovery efficiency and practicality are improved, and the numerical stability and adaptability are good.
Smart Images

Figure CN119027351A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of image restoration, and in particular to a color image restoration method based on singular value factors. Background Art
[0002] With the development of digital image processing technology, image restoration technology plays an increasingly important role in various fields, such as medical imaging, satellite remote sensing, video surveillance, etc. Image restoration technology aims to improve image quality and enhance image usability, especially when the image is affected by degradation such as noise and blur.
[0003] Traditional image restoration methods, such as singular value thresholding methods, can achieve certain results in some cases, but they usually have limitations when dealing with high-resolution image restoration. For example, higher resolution leads to heavier computational burden and lower restoration efficiency.
[0004] In recent years, image restoration methods based on matrix decomposition have received extensive attention. However, existing image restoration methods based on matrix decomposition face some challenges when processing color images. Color images usually contain three color channels (red, green, and blue). Traditional methods need to process each channel separately, which not only increases the computational complexity, but also may cause color distortion and inconsistency problems. Therefore, many scholars have proposed to treat color images as a third-order tensor. The paper Parallel matrix factorization for low-rank tensor completion proposes TMac-dec and TMac-inc methods to deal with the problem of color image restoration, and the paper Tensor completion for estimating missing values in visual data proposes the HaLRTC method for color image restoration. Although these methods treat color images as a third-order tensor as a whole, in the process of solving the restoration problem, the tensor is expanded as a technical means according to the module, and it is necessary to expand it once along each module of the tensor. On the one hand, this technology brings about the problem of a sharp increase in the dimension after the module expansion, which leads to an increase in the amount of calculation and a decrease in the restoration efficiency; on the other hand, this technology destroys the overall structure of the color image, which leads to a decrease in the restoration quality. In order to improve these defects, the present invention proposes a color image restoration method based on singular value factors. Summary of the invention
[0005] The object of the present invention is to provide a color image restoration method based on singular value factors, which can effectively restore polluted color images while maintaining the color consistency and main information of the image.
[0006] In view of the shortcomings of the prior art, the present invention proposes a color image restoration method based on singular value factors. The method integrates the three color channels of the color image into a unified tensor, and then applies the random tensor singular value decomposition technology combined with the adaptive Tubal rank adjustment method, which effectively solves the problems encountered by the traditional method in processing color images. The present invention not only improves the quality of image restoration, but also simplifies the calculation process, and has higher efficiency and better practicality.
[0007] Technical solution: The present invention discloses a color image restoration method based on singular value factors, which mainly performs restoration processing on polluted color images to restore clear original images, including the following steps:
[0008] Step 1: Input the pollution color map to be processed;
[0009] Step 2: Use the Frobenius norm to build an image restoration model based on tensor singular value decomposition:
[0010]
[0011] Where U, S, V are singular value factors, C is the medium tensor, M is the polluted color image tensor, * represents tensor product, ‖·‖ F represents the Frobenius norm, V H represents the conjugate transpose of V, P Ω is the projection operator, Ω is the position set of the uncontaminated part of the image;
[0012] Step 3: Solve the recovery model of step 2 with the help of ADM;
[0013] Step 4: Using the solution process of step 3, an adaptive method is used to estimate the Tubal rank of the image tensor for calculation;
[0014] Step 5: Use tensor random singular value decomposition to adjust the singular value factors solved in step 3, and output the restored image after the convergence accuracy is met.
[0015] Furthermore, the image restoration model in step 2 performs Fourier transform on the third dimension of the image tensor and converts it into a restoration model:
[0016]
[0017] In the formula represents the block diagonal information matrix obtained after Fourier transform.
[0018] Furthermore, the solution of singular value factors U, S, V in step 3 is transformed into the solution of the transformed singular value factors in model (2) and media information Solving for Transformation Singular Value Factors and media information When using the ADM method, the multivariable problem is transformed into four single variable problems:
[0019]
[0020] Among them, equations (3), (4), and (5) are the problems of solving the singular value decomposition factors, and equation (6) is the problem of solving the medium information C.
[0021] The corresponding iterative solution is:
[0022]
[0023] The corresponding iterative solution is:
[0024]
[0025] The corresponding iterative solution is:
[0026]
[0027] In the formula, represent The Moore-Penrose inverse.
[0028] The solution of the media information C is as follows:
[0029]
[0030] Among them, Ω c represents the complement of Ω.
[0031] Furthermore, the Tubal rank of the media information C is determined by the singular value spacing, and the spacing variable gap is introduced:
[0032]
[0033] Among them, r j represents the ratio of adjacent singular values, r j =λ j / λ j+1 ,j=1,2,…,τ-1, λ j Represents the singular values of each slice after the Fourier transform of the media information C along the third dimension. The singular values are arranged in descending order. If gap ≥ 10, the rank is adjusted from τ to τ k .
[0034] Compared with the prior art, the present invention has the following beneficial effects:
[0035] 1. The present invention regards the color image as a third-order tensor, and there is no need to expand the tensor according to the model. Compared with the traditional method, the computational workload is greatly reduced, thereby improving the recovery efficiency.
[0036] 2. The present invention processes the three color channels of a color image, red, green and blue, simultaneously, thereby keeping the overall structure intact and ensuring image restoration performance.
[0037] 3. Singular value decomposition SVD, as an effective matrix decomposition technology, has been applied to image denoising, image super-resolution reconstruction and other fields. The present invention uses the singular value decomposition (T-SVD) of the tensor under the tensor product to restore the color image. This technology has good numerical stability and is conducive to solving the optimization problem. The solution form is simple.
[0038] 4. Apply the ADM method to transform the multivariable problem into four single variable problems. In actual calculations, the ADM method shows good stability and robustness, and can achieve satisfactory results even in the presence of noise or unsatisfactory initial conditions.
[0039] 5. The tensor singular value spacing is used to adaptively determine the Tubal rank of the color map without the need to give the rank in advance, which enhances the adaptability of the method and makes it easier to implement in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Figure 1 It is a comparison diagram of color image decontamination of the present invention; among them, (a) is the original image; (b) is the noise-contaminated image; (c) is the restored image by the Tmac-dec method; (d) is the restored image by the Tmac-inc method; (e) is the restored image by the HaLRTC method; (f) is the restored image by the method of the present invention.
[0041] Figure 2 This is a test picture of Example 3 of the present invention. DETAILED DESCRIPTION
[0042] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.
[0043] Example 1
[0044] The technical solution provided in this embodiment is a color image restoration method based on singular value factors, which mainly performs decontamination processing on a polluted image to restore a clear original image, including the following steps:
[0045] Set P Ω(M) is the image information of the uncontaminated part, P Ω is the projection operator, and Ω is the position set of the uncontaminated part of the image.
[0046] In order to restore the complete image information M, the present invention proposes the following image restoration model based on tensor singular value decomposition:
[0047]
[0048] Apply Fourier transform along the third dimension of the tensor to convert it into a restored model:
[0049]
[0050] Formula (2) is a multivariable problem, which can be transformed into the following single variable problem:
[0051]
[0052] Among them, equations (3), (4), and (5) are the problems of solving the singular value decomposition factors, and equation (6) is the problem of solving the medium information C.
[0053] The solution of formula (3) is:
[0054]
[0055] The solution of formula (4) is:
[0056]
[0057] The solution of formula (5) is:
[0058]
[0059] The solution of the media information C is as follows:
[0060]
[0061] The Tubal rank of the medium information C is adjusted by the singular value gap. If gap ≥ 10, adjust the rank from τ to τ k .
[0062] When the calculation process reaches a given accuracy, the restored image C of M is output.
[0063] In the above solution process, The orthogonality of the tensor The correction is performed by using the singular value decomposition or random singular value decomposition of the algorithm. During the correction, the orthogonal columns are determined according to the adaptive Tubal rank. The convergence accuracy in the experiment is taken as 0.00001. When , the calculation is terminated and the recovery result is output. The convergence accuracy can also be selected according to the actual scene requirements to select an ideal value.
[0064] Example 2
[0065] This embodiment is attached Figure 1 For the test image, the image resolution is 640×426×3. Figure 1 The noise addition process causes a 30% pixel loss, which is equivalent to a 70% sampling rate. After that, the third dimension of the sampled image tensor is Fourier transformed and converted into a recovery model:
[0066]
[0067] Then follow the following formula Update:
[0068]
[0069] in The orthogonality correction is done by the tensor The singular value decomposition or random singular value decomposition is performed. Here, the test uses random singular value decomposition, and the steps are as follows:
[0070] (1) Generate a 426×(r+p) Gaussian random matrix Ψ, where r is Tubal rank, p is the oversampling parameter, and in this embodiment, p=5;
[0071] (2)
[0072] (3) [Q, R] = qr(Y, 0), where qr(·) represents the qr decomposition of Y;
[0073] (4)
[0074] (5) [U, S, V] = svd(B), where svd(·) represents the svd decomposition of B;
[0075] (6) The convergence accuracy is taken as 0.00001, that is, when When the calculation is terminated, the output is restored. The results are shown in the attached Figure 1 It can be seen that the recovery effect of the method of the present invention is relatively good.
[0076] Example 3
[0077] This embodiment is attached Figure 2For the test picture, the picture resolution is 640×427×3. The embodiment considers the restoration effect of the present invention at different sampling rates. The sampling rates are 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, and 90%, respectively. The third dimension of the sampled image tensor at different sampling rates is Fourier transformed and converted into a restoration model:
[0078]
[0079] Then follow the following formulas: Update:
[0080]
[0081] in The orthogonality correction is done by the tensor The singular value decomposition and random singular value decomposition are respectively performed, and the representative results of the present invention (random) in Table 1 and Table 2 are the results under the random singular value decomposition. Table 1 statistics the signal-to-noise ratio of the image restoration of the method of the present invention at different sampling rates. It can be seen that the signal-to-noise ratio of the restoration method of the present invention is relatively high. Referring to Table 1, it is shown that the image restoration quality is relatively good.
[0082] Table 1: Comparison of restored signal-to-noise ratios of the present invention and other methods at different sampling rates
[0083]
[0084] Note: The present invention (random) means that random singular value decomposition is used in orthogonality correction.
[0085] Table 2 summarizes the image recovery time of the method of the present invention at different sampling rates. It can be seen that when the sampling rate is very low, the time taken to perform correction using the conventional singular value decomposition is slightly longer. However, as long as the sampling rate is not very low, regardless of whether the conventional singular value decomposition or the random singular value decomposition is used, the image recovery operation time of the present invention is relatively short, indicating that the image recovery efficiency is high.
[0086] Table 2: Comparison of recovery time between the present invention and other methods at different sampling rates
[0087]
[0088] Note: The present invention (random) means that random singular value decomposition is used in orthogonality correction.
[0089] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A color image restoration method based on singular value factors, characterized in that: The steps include: Step 1: Input the pollution color map to be processed; Step 2: Use the Frobenius norm to build an image restoration model based on tensor singular value decomposition: Where U, S, V are singular value factors, C is the medium tensor, M is the contaminated color image tensor, * represents tensor product, ‖·‖ F represents the Frobenius norm, V H represents the conjugate transpose of V, P Ω is the projection operator, Ω is the position set of the uncontaminated part of the image; Step 3: Solve the recovery model of step 2 with the help of ADM; Step 4: Using the solution process of step 3, an adaptive method is used to estimate the Tubal rank of the image tensor for calculation; Step 5: Use tensor random singular value decomposition to correct and characterize the singular value factors solved in step 3, and output the restored image after the convergence accuracy is met.
2. The color image restoration method based on singular value factors according to claim 1, characterized in that: The image restoration model in step 2 performs Fourier transform on the third dimension of the image tensor and converts it into a restoration model: In the formula represents the block diagonal information matrix obtained after Fourier transform.
3. The color image restoration method based on singular value factors according to claim 1, characterized in that: The solution of singular value factors U, S, V in step 3 is transformed into the solution of transformation singular value factors and media information 4. The color image restoration method based on singular value factors according to claim 2, characterized in that: The recovery model (2) solves the transformation singular value factor and media information When , the multivariable problem is transformed into four single variable problems: Among them, equations (3), (4), and (5) are the singular value decomposition factors The solution problem is: The problem of solving.
5. The color image restoration method based on singular value factors according to claim 4, characterized in that: The solution of the singular value decomposition factor is as follows: in, represent The Moore-Penrose inverse.
6. The color image restoration method based on singular value factors according to claim 4, characterized in that: The media information The solution results are as follows: Among them, Ω c represents the complement of Ω.
7. The color image restoration method based on singular value factors according to claim 1, characterized in that: In the implementation of the adaptive method in step 4, the Tubal rank of the medium information C is determined by the singular value spacing, and the spacing is defined as follows: Among them, r j represents the ratio of adjacent singular values, r j =λ j / λ j+1 ,j=1,2,…,τ-1, λ j Represents the singular value of each slice after the Fourier transform of the media information C along the third dimension. The singular values are arranged in descending order. If gap ≥ 10, the Tubal rank of the media information C is adjusted from τ to τ k .
8. The color image restoration method based on singular value factors according to claim 1, characterized in that: The convergence accuracy in step 5 can be taken as 10 -2 ~10 -5 .
