Image noise reduction method and image noise reduction system based on graph smoothing RPCA
By introducing graph smoothing technology and non-convex punishment in RPCA, combined with ADMM algorithm, the problem of insufficient noise reduction effect of existing RPCA in non-strict low-rank data processing is solved, and efficient images under K non-strict low-rank situations are achieved Noise reduction.
Patent Information
- Application Number
- CN202410997120.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-24
- Publication Date
- 2024-11-15
AI Technical Summary
The existing robust principal component analysis method (RPCA) has poor noise reduction effect when processing non-strict low-rank data and cannot effectively recover potential features in the image.
Using a graph smoothing-based RPCA method, combined with non-convex punishment and graph learning technology, the data matrix is processed through the ADMM algorithm, and graph Laplace operator is introduced to smooth the graph signal, and a new RPCA model is constructed to improve noise reduction performance.
In the case of K non-strictly low rank, the new method significantly improves the recovery performance of image noise reduction, ensures effective extraction of low-rank components and sparse components of the data matrix, and enhances the robustness of the algorithm.
Smart Images

Figure CN118967496A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of image processing, and in particular relates to an image denoising method and an image denoising system based on graph smoothing RPCA. Background Art
[0002] With the advent of the big data era and the continuous progress of scientific research, data is becoming more and more important to us. In recent years, due to the continuous breakthroughs in data acquisition and storage technology, the data we need often presents extremely high dimensions. This extremely high-dimensional data will not only greatly increase the computing and storage load, but also cause the "dimensionality disaster". The so-called "dimensionality disaster" is to ensure that the classification system always maintains good performance. The number of samples will increase geometrically with the increase of dimensions, but this high-dimensional complexity often contains many potential features that are not easy to detect. For these high-dimensional data, how do we mine these potential features and how to reduce their dimensions while ensuring that the original information is not lost as much as possible? This is data dimensionality reduction. Data dimensionality reduction can help people obtain the available information in the original high-dimensional data more quickly and efficiently, can effectively deal with the dimensionality disaster, and improve the efficiency of the classification system. In fact, data dimensionality reduction is an important method of feature extraction, which can effectively extract potential features in high-dimensional data and reduce the computing and storage costs of the system.
[0003] Principal component analysis (PCA) is a commonly used tool for data analysis and data dimensionality reduction. It can be understood as the projection of high-dimensional data into a low-dimensional space. However, when the data is noisy, PCA becomes less effective. Based on this problem, researchers proposed the robust principal component analysis (RPCA).
[0004] Robust principal component analysis (RPCA) is very effective in processing data with large noise pollution. It has been widely used in face recognition, collaborative filtering, video surveillance, graphical model learning and linear system recognition. The existing algorithm combines the generalized non-convex penalty induced by low-rank sparsity and proposes a new RPCA algorithm model based on generalized non-convex regularization. The optimization problem of this algorithm is:
[0005]
[0006] in, and g 2(.) are the generalized non-convex penalties induced by low rank and sparsity, respectively. A more practical situation is considered here, that is, there may be an input noise, that is, M = K + S + N, N is the input noise, and σ is the noise tolerance. This method proposes an RPCA algorithm that uses a generalized non-convex penalty under low rank sparsity induction, and proves through experiments that when K is strictly low rank, accurate recovery can only be achieved through non-convex regularization, but when K is not strictly low rank, the denoising effect of this method is poor.
[0007] Therefore, based on the above technical problems, it is necessary to design a new image denoising method and image denoising system based on graph smoothing RPCA. Summary of the invention
[0008] The purpose of the present invention is to provide an image denoising method and an image denoising system based on graph smoothing RPCA.
[0009] In order to solve the above technical problems, the present invention provides an image denoising method based on graph smoothing RPCA, comprising:
[0010] The RPCA model based on graph smoothing is obtained according to the non-convex penalty and graph learning method;
[0011] The RPCA model based on graph smoothing is processed by ADMM algorithm.
[0012] Further, the method of obtaining the graph smoothing-based RPCA model according to the non-convex penalty and graph learning method includes:
[0013] RPCA model based on graph smoothing:
[0014]
[0015] Where K is the low-rank component of the data matrix; S is the sparse component; L is the graph Laplacian operator, which is obtained by the graph learning method; τ is the noise tolerance, τ>0; λ is the weight factor, λ>0; γ is a constant, γ>0; and g 2 (.) are the low-rank and sparsity-induced generalized non-convex penalties, respectively.
[0016] Furthermore, the method of obtaining the graph Laplacian operator by the graph learning method includes:
[0017] L = DW;
[0018] Among them, D is the degree matrix; W is the weighted adjacency matrix.
[0019] Furthermore, the graph Laplacian operator L is used as a differential operator, and the smoothness of the graph signal relative to the underlying graph is measured using the quadratic form of the graph Laplacian matrix:
[0020]
[0021] Among them, w ij represents the weight of the edge between two adjacent vertices i and j, and f(i) and f(j) are the signal values associated with these two vertices;
[0022] The smaller the quadratic shape, the smoother the signal on the graph, and the graph smoothing operator is:
[0023]
[0024] The smaller the quadratic shape, the smoothing operator tr(X T The smaller LX) is, the smoother the graph signal is. Then the graph smoothing operator constructed based on the low-rank component of the data matrix is tr(K T LK).
[0025] Furthermore, the graph smoothing-based RPCA model is converted into an unconstrained form:
[0026]
[0027] Among them, β is a penalty parameter related to the noise parameter τ, β>0, when β infinitely approaches 0, it means that there is no input noise.
[0028] Furthermore, the optimization problem in unconstrained form is solved by the alternating direction multiplier method;
[0029] Through two auxiliary variables K = Z and S = V, the unconstrained form is transformed into:
[0030]
[0031] The corresponding augmented Lagrangian function is:
[0032]
[0033] Among them, W and U are dual variables; ρ 1 >0 and ρ 2 >0 is the penalty parameter; in k+1 iterations, the alternating direction multiplication method is used to update the original variables and the dual variables.
[0034] Furthermore, the method for updating the original variable and the dual variable using the alternating direction multiplier method includes:
[0035]
[0036]
[0037]
[0038]
[0039] W k+1 =W k +ρ 1 (W k+1 -Z k+1 )
[0040] U k+1 =U k +ρ 2 (S k+1 -V k+1 ). Further, the solutions of K and S are:
[0041] K k+1 =((β+ρ 1 )I+2γL) -1 (β(MS k )+ρ 1 Z k+1 -W k )
[0042]
[0043] Furthermore, when calculating K, the Moore-Penrose generalized inverse method is used.
[0044] On the other hand, the present invention also provides an image denoising system using the above-mentioned image denoising method based on graph smoothing RPCA, comprising:
[0045] A model building module, configured to obtain a graph smoothing-based RPCA model according to a non-convex penalty and graph learning method;
[0046] The processing module is configured to process the RPCA model based on graph smoothing through an ADMM algorithm.
[0047] The beneficial effects of the present invention are that the present invention obtains a graph-smoothing-based RPCA model according to a non-convex penalty and graph learning method; processes the graph-smoothing-based RPCA model through an ADMM algorithm; realizes that on the basis of the existing RPCA model, the non-convex penalty is combined with the graph learning technology to improve the model, and proposes a new graph-smoothing-based RPCA model, which uses generalized non-convex penalties and graph smoothing regularization to process data, and proposes a new ADMM algorithm for the model, which can effectively solve the formulas in the model. The original algorithm uses non-convex penalties to enable the algorithm to obtain better results when K (the low-rank component of the data matrix) is strictly low-rank, but when K is not strictly low-rank, the advantage of non-convex penalties is not very obvious. On this basis, the graph learning technology is introduced to make up for this shortcoming, so that the algorithm can also have good performance when K is not strictly low-rank.
[0048] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description and the drawings.
[0049] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0051] Figure 1 It is a flow chart of an image denoising method based on graph smoothing RPCA of the present invention;
[0052] Figure 2 is a simple undirected graph of the present invention;
[0053] Figure 3 is the undirected graph and graph signal x of the present invention;
[0054] Figure 4 is the original image in the example of the present invention;
[0055] Figure 5 This is a schematic diagram of the algorithm recovery performance in the case where K is not strictly low rank in the example of the present invention;
[0056] Figure 6 This is a schematic diagram showing the effect of the value of K on the recovery performance of the algorithm when K is not strictly low-rank in the example of the present invention;
[0057] Figure 7 It is a schematic diagram of the recovery performance of the algorithm when K is not strictly low rank in the example of the present invention when there is input noise;
[0058] Figure 8 Schematic diagram of the effect of the value of K on the recovery performance of the algorithm when there is input noise in the example of the present invention;
[0059] Fig. 9 It is a schematic diagram of the application test of the original algorithm and the new algorithm in the example of the present invention in dynamic and static background separation. DETAILED DESCRIPTION
[0060] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are 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.
[0061] like Figures 1 to 9 In at least one embodiment, a method for image denoising based on graph smoothing RPCA is provided, including: obtaining a graph smoothing-based RPCA model according to a non-convex penalty and graph learning method; processing the graph smoothing-based RPCA model through an ADMM algorithm; improving the model by combining non-convex penalty with graph learning technology on the basis of the existing RPCA model, proposing a new graph smoothing-based RPCA model, which processes data using generalized non-convex penalty and graph smoothing regularization, and proposing a new ADMM algorithm for the model, which can effectively solve the formulas in the model. The original algorithm uses non-convex penalty to enable the algorithm to obtain better results when K (the low-rank component of the data matrix) is strictly low-rank, but when K is not strictly low-rank, the advantage of non-convex penalty is not very obvious. On this basis, graph learning technology is introduced to make up for this shortcoming, so that the algorithm can also have good performance when K is not strictly low-rank.
[0062] For many sparse and low-rank recovery problems, neighbor operators play an important role in developing efficient first-order algorithms; for a lower semi-continuous penalty function g(.), its neighbor operator can be defined as;
[0063]
[0064] Among them, η>0 is the penalty operator;
[0065] Hard threshold penalty (Hard), whose penalty function is l 0 norm, i.e. g(x) = |x| 0 , which can also be written as:
[0066]
[0067] Then its neighboring operator is:
[0068]
[0069] Soft threshold penalty (Soft), the penalty is l 1 norm, g(x) = |x|, then its neighboring operator is:
[0070] P g,η (v)=sign(v)·max{|v|-1 / η, 0} (4);
[0071] Lq norm penalty (Lq), whose penalty function is:
[0072] g(x)=|x| q , 0<q<1 (5);
[0073] Then its neighboring operator is:
[0074]
[0075] Where β = [2(1-q) / η] 1 / (2-q) ,τ=β+qβ q-1 / η,y * ∈(β,|v|) is the equation h(y)=qy q-1 +ηy-η|v|=0. Since h(y) is convex and differentiable, y * It can be calculated by an iterative algorithm.
[0076] The low-rank regularization of a matrix is usually achieved by sparse regularization of the matrix singular values. For a matrix M with rank r, its singular value decomposition is M = U∑V T , this method defines a penalty To achieve the low rank improvement of M;
[0077]
[0078] Among them, g(.) is the several threshold penalties mentioned above;
[0079] Penalty for definition This can be transformed into a minimization problem:
[0080]
[0081] The solution is:
[0082]
[0083] Among them, P g,η (.) refers to the several neighboring operators introduced earlier. This generalized singular value threshold penalty is an extension of the scalar threshold penalty defined in the previous section.
[0084] The original algorithm combines the generalized non-convex penalty induced by low-rank sparsity and proposes a new RPCA algorithm model based on generalized non-convex regularization. The optimization problem of this algorithm is:
[0085]
[0086] in, and g 2 (.) are low-rank and sparsity-induced generalized non-convex penalties, respectively. Here we consider a more practical situation, that is, there may be an input noise, that is, M = K + S + N, N is the input noise, and σ is the noise tolerance.
[0087] The principle of graph signal processing is described based on the architecture of graph models, which are structural graphs composed of points and lines to describe the characteristics between data. Graph models are essentially structural graphs, which are often used to represent the connections between individuals in human society or nature. Each node in the graph model can be represented by an individual around it, and the edges between nodes in the model can represent the relationship between individuals. The edges between nodes are assigned weights, thus forming a weighted graph. The weight is the data value of the physical distance, flow, etc. between two points. The emergence of graph models provides a good tool for studying complex systems, but the graphs used to construct graph models are different from geometric figures in the general sense.
[0088] The graph can be defined as G = (V, ε, W), where the vertex set V = {v 1 , ..., v N} means G has N vertices, Represents the edge set, and each edge β∈ε in the graph can be represented as W represents the weighted adjacency matrix of the graph. Figure 2 As shown, this is a simple weighted undirected graph with five nodes and four edges, that is, the vertex set of the graph is V = {v 1 , v 2 , v 3 , v 4 , v 5}, the edge set of the graph is ε={β 12 , β13 , β 14 , β 15}. The weighted adjacency matrix W of the undirected graph is:
[0089]
[0090] The weighted adjacency matrix of a graph generally satisfies the following conditions:
[0091]
[0092] in represents the set of non-negative real numbers, w ij Represents the connected vertex v i and v j The edge ij The weight of the unweighted graph is 1. However, the unweighted graph is a special case of the weighted graph, and the edge weight of each edge is 1. The adjacency matrix W of the undirected graph is a symmetric matrix, but the adjacency matrix of the directed graph is not necessarily symmetric. The diagonal matrix of the graph can be recorded as D G =diag{d i} i=1,2,3...N , d i Here, d represents the degree of the node, so the diagonal matrix of the graph can also be called the degree matrix, i =∑ j W ij That is, the sum of the connection weights of node i, which is actually the sum of each row of the weighted adjacency matrix. The data on the graph is the signal value attached to each node on the graph. The data on the graph can be characterized as a finite set of samples, which is the graph signal. Figure 3 As shown, It is a graph signal. The i-th component x i With the i-th vertex v on the graph i One-to-one correspondence. Therefore, the graph signal can be considered as a function that maps the vertices of the graph to real values, that is, The graph Laplacian matrix is also a tool for describing the topological structure of a graph, which is defined as: L = DW (11); D is the degree matrix, W is the weighted adjacency matrix, So the Laplace matrix must satisfy L≥0 at the same time, i≠j when L ij ≤0, L1=0, where 0 and 1 represent all-0 column vectors and all-1 column vectors respectively. The normalized Laplacian matrix L * Defined as:
[0093]
[0094] if Figure 2 The graph in is an unweighted graph, then:
[0095]
[0096]
[0097]
[0098] Graph signals can be used to represent data information in different network architectures. For example, in a sensor network, the information value measured by each node can be converted into a graph signal. In a traffic network, the car flow at each intersection can also be used as a graph signal, and the node is each intersection. In a social network, each user is a node, and the information value related to him is also a graph signal.
[0099] The graph Laplacian operator L is used as a differential operator. The smoothness of the graph signal relative to the underlying graph is measured using the quadratic form of the graph Laplacian matrix:
[0100]
[0101] Among them, w ij represents the weight of the edge between two adjacent vertices i and j, and f(i) and f(j) are the signal values associated with these two vertices; considering that the weight is non-negative, the above formula shows that if strongly connected vertices (the edges between them have large weights) have similar values, then the graph signal is considered smooth.
[0102] The smaller the quadratic form in the above formula, the smoother the signal on the graph, and the graph smoothing operator is:
[0103]
[0104] The smaller the quadratic shape, the smoothing operator tr(X T The smaller LX) is, the smoother the graph signal is. Then the graph smoothing operator constructed based on the low-rank component of the data matrix is tr(K T LK).
[0105] Aiming at the RPCA algorithm based on generalized non-convex regularization, the concept of graph smoothing regularization in graph signal processing is introduced, and a RPCA algorithm based on graph smoothing is proposed. The optimization problem of the algorithm can be described by an RPCA model based on graph smoothing; the RPCA model based on graph smoothing obtained according to the non-convex penalty and graph learning method includes: RPCA model based on graph smoothing:
[0106]
[0107] Where K is the low-rank component of the data matrix; S is the sparse component; L is the graph Laplacian operator, which is obtained by the graph learning method and is known a priori; τ is the noise tolerance, τ>0; λ is the weight factor, λ>0; γ is a constant, γ>0; and g2 (.) are the low-rank and sparsity-induced generalized non-convex penalties, respectively.
[0108] The graph smoothing-based RPCA model is converted into an unconstrained form, namely, the unconstrained form of formula (16):
[0109]
[0110] Among them, β is a penalty parameter related to the noise parameter τ, β>0, when β infinitely approaches 0, it means that there is no input noise.
[0111] The optimization problem in the unconstrained form is solved by the alternating direction multiplier method, that is, the optimization problem in equation (17) is solved by the alternating direction multiplier method (ADMM);
[0112] Through two auxiliary variables K = Z and S = V, the unconstrained form is transformed into, and the optimization problem in equation (17) can be restated as:
[0113]
[0114] The corresponding augmented Lagrangian function is:
[0115]
[0116] Among them, W and U are dual variables; ρ 1 >0 and ρ 2 >0 is the penalty parameter; in k+1 iterations, the alternating direction multiplication method is used to update the original variables and the dual variables.
[0117] The method for updating the original variable and the dual variable by the alternating direction multiplier method includes:
[0118]
[0119]
[0120]
[0121]
[0122] W k+1 =W k +ρ 1 (K k+1 -Z k+1 ) (twenty four);
[0123] U k+1 =U k +ρ 2 (S k+1 -V k+1) (25).
[0124] Here, the solution to the optimization problem of Z can be regarded as a form of formula (8), and its solution is formula (9), while the optimization problem of V can be solved by formula (1). The solutions of K and S are:
[0125] K k+1 =((β+ρ 1 )I+2γL) -1 (β(MS k )+ρ 1 Z k+1 -W k ) (26);
[0126]
[0127] When calculating K, the concept of Moore-Penrose generalized inverse is used to solve this problem.
[0128] Specifically, the new algorithm proposed in this embodiment is verified through three experiments. Figure 4 The 256×256 image shown in the figure, S is sparse noise, and M = K + S. In the experiment, we considered the case where K is not strictly low-rank, that is, using the original image, S has different sparsities: (1) K is not strictly low-rank, S is not strictly sparse; (2) K is not strictly low-rank, S is strictly sparse.
[0129] For the case where S is strictly sparse, this experiment uses salt and pepper noise, and the experiment sets 25% of the pixels in the image to be destroyed by salt and pepper noise; for the case where S is not strictly sparse, the experiment uses symmetric α-stable (SαS) noise. SαS distribution is widely used in impulse noise modeling, and the characteristic function of zero-position SαS distribution can be expressed as:
[0130]
[0131] Among them, ε>0 is a proportional parameter; 0<α≤2 is a characteristic exponential parameter. The smaller the α value, the stronger the noise impulse. When α=1 and α=2, the SαS distribution becomes a Cauchy distribution and a Gaussian distribution. In the experiment of this embodiment, SαS noise with α=1 and ε=0.5 was used.
[0132] The restoration performance of our proposed algorithm in the above two cases is shown. The restoration performance is evaluated by relative suppression ratio (RelErr) and peak signal-to-noise ratio (PSNR). Figure 5 In the above figure, “Corrupted” represents the original image; “Soft” represents the use of soft threshold penalty for the first two items of the objective function, i.e., l 1Penalty; "Hard" means using hard threshold penalty for the first two items of the objective function, i.e. l 0 Penalty; "Lq" means q 1 ,q 2 The recovery performance of the algorithm when taking different values, where q 1 Express The constraint, q 2 Indicates g 2 The constraint is , and 0<q 1 ,q 2 <1, that is, in the non-convex case, the recovery performance of the algorithm. 1 ,q 2 <1, we first run q 1 =q 2 =1, that is, the soft threshold penalty is used to obtain an initialized K and S. The input values of the hard threshold penalty and Lq penalty are the initialized K and S.
[0133] In the experiment, the value range of parameter λ is [10 -3 , 10 0 ], with a step size of 10 0.1 ; The range of parameter γ is [10 -4 , 10 -1 ], with a step size of 10 0.1 . The recovery performance of the algorithm in the case where K is not strictly low-rank. Figure 5 (a) S is not strictly sparse, and (b) S is strictly sparse.
[0134] Figure 6 In the case where K is not strictly low-rank, q 1 and q 2 The value of affects the recovery performance of the algorithm. The horizontal axis represents q 2 , the vertical axis represents q 1 (a) S is not strictly sparse; (b) S is strictly sparse.
[0135] Figure 6 It is shown that when K is not strictly low-rank, q 1 and q 2 The value of affects the recovery performance of the algorithm. From the figure we can see that when K is not strictly low-rank and S is not strictly sparse, q 1 =0.6,q 2 = 0.9, the algorithm has the best recovery performance; when K is not strictly low-rank and S is strictly sparse, q 1 =0.8,q 2 =0.9 algorithm has the best recovery performance.
[0136] Table 1 shows the comparison of the results of the algorithm proposed in this patent and the comparative algorithm when K is not strictly low-rank, S is not strictly sparse, and S is strictly sparse. It can be seen from the table that in most cases, the algorithm proposed in this patent has better results.
[0137] Table 1 Comparison of the results of the original algorithm and the proposed algorithm when S is not strictly sparse and S is strictly sparse
[0138]
[0139] In the second experiment, possible input noise in real life was taken into consideration, so possible input noise was artificially added to the image. The input noise was modeled as zero-mean Gaussian noise with a variance of 4. In both cases, the recovery performance of the algorithm is as follows: Figure 7 As shown, q 1 and q 2 The value of Figure 8 As shown, when K is not strictly low-rank and S is not strictly sparse, q 1 =0.9,q 2 = 0.9, the algorithm has the best recovery performance; when K is not strictly low-rank and S is strictly sparse, q 1 =0.9,q 2 =0.5 algorithm has the best recovery performance. Figure 7 The recovery performance of the algorithm is shown when K is not strictly low-rank in the presence of input noise, (c) S is not strictly sparse, and (d) S is strictly sparse. Figure 8 It shows that when there is input noise, q 1 and q 2 The value of affects the recovery performance of the algorithm. The horizontal axis represents q 2 , the vertical axis represents q 1 , (c) S is not strictly sparse, (d) S is strictly sparse.
[0140] In the third experiment, the present embodiment tests the new algorithm in the application of dynamic and static background separation, using a 160×130 escalator video with 200 frames. Without using any prior information about the sparse error, a graph Laplacian is constructed between the video frames. Fig. 9 The results show that the original algorithm and the algorithm proposed in this embodiment are tested in the application of dynamic and static background separation. Compared with the original algorithm, the algorithm proposed in this embodiment has a smaller error and a larger peak signal-to-noise ratio, and its effect is better than the original algorithm.
[0141] This embodiment proposes a new algorithm, which combines non-convex penalty with graph learning technology to improve the model on the basis of the existing RPCA model, and proposes a new RPCA model based on graph smoothing, which uses generalized non-convex penalty and graph smoothing regularization to process data. This embodiment proposes a new ADMM algorithm for the model, which can effectively solve the formulas in the model. The results show that although the original algorithm uses non-convex penalty to obtain better results when K is strictly low-rank, the advantage of non-convex penalty is not very obvious when K is not strictly low-rank. On this basis, this embodiment introduces graph learning technology to make up for this shortcoming, so that the algorithm can still have good performance when K is not strictly low-rank.
[0142] In at least one other embodiment, an image denoising system adopting the above-mentioned image denoising method based on graph smoothing RPCA is also provided, including: a model building module, which is configured to obtain a graph smoothing-based RPCA model according to a non-convex penalty and graph learning method; a processing module, which is configured to process the graph smoothing-based RPCA model through an ADMM algorithm.
[0143] In summary, the present invention obtains a graph-smoothing-based RPCA model according to a non-convex penalty and graph learning method; processes the graph-smoothing-based RPCA model through an ADMM algorithm; realizes that on the basis of the existing RPCA model, the non-convex penalty is combined with the graph learning technology to improve the model, and proposes a new graph-smoothing-based RPCA model, which uses generalized non-convex penalties and graph smoothing regularization to process data, and proposes a new ADMM algorithm for the model, which can effectively solve the formulas in the model. The original algorithm uses non-convex penalties to enable the algorithm to obtain better results when K (the low-rank component of the data matrix) is strictly low-rank, but when K is not strictly low-rank, the advantage of non-convex penalties is not very obvious. On this basis, the graph learning technology is introduced to make up for this shortcoming, so that the algorithm can also have good performance when K is not strictly low-rank.
[0144] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.
[0145] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0146] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.
[0147] With the above-mentioned ideal embodiments of the present invention as inspiration, through the above-mentioned description content, relevant personnel can make various changes and modifications without departing from the scope of the technical idea of the present invention.
Claims
1. An image denoising method based on graph smoothing RPCA, characterized in that: include: The RPCA model based on graph smoothing is obtained according to the non-convex penalty and graph learning method; The RPCA model based on graph smoothing is processed by ADMM algorithm.
2. The image denoising method based on graph smoothing RPCA as claimed in claim 1, characterized in that: The method of obtaining the graph smoothing-based RPCA model according to the non-convex penalty and graph learning method includes: RPCA model based on graph smoothing: Where K is the low-rank component of the data matrix; S is the sparse component; L is the graph Laplacian operator, which is obtained by the graph learning method; τ is the noise tolerance, τ>0; λ is the weight factor, λ>0; γ is a constant, γ>0; and g2(.) are low-rank and sparsity-induced generalized non-convex penalties, respectively.
3. The image denoising method based on graph smoothing RPCA as claimed in claim 2, characterized in that: The method for obtaining the graph Laplacian operator by the graph learning method includes: L = DW; Among them, D is the degree matrix; W is the weighted adjacency matrix.
4. The image denoising method based on graph smoothing RPCA as claimed in claim 3, characterized in that: The graph Laplacian operator L is used as a differential operator. The smoothness of the graph signal relative to the underlying graph is measured using the quadratic form of the graph Laplacian matrix: Among them, w ij represents the weight of the edge between two adjacent vertices i and j, and f(i) and f(j) are the signal values associated with these two vertices; The smaller the quadratic shape, the smoother the signal on the graph, and the graph smoothing operator is: The smaller the quadratic shape, the smoothing operator tr(X T The smaller LX) is, the smoother the graph signal is. Then the graph smoothing operator constructed based on the low-rank component of the data matrix is tr(K T LK).
5. The image denoising method based on graph smoothing RPCA as claimed in claim 4, characterized in that: Convert the graph smoothing based RPCA model to an unconstrained form: Among them, β is a penalty parameter related to the noise parameter τ, β>0, when β infinitely approaches 0, it means that there is no input noise.
6. The image denoising method based on graph smoothing RPCA as claimed in claim 5, characterized in that: Solve the optimization problem in unconstrained form by the alternating direction multiplier method; Through two auxiliary variables K = Z and S = V, the unconstrained form is transformed into: stK=Z,S=T,L=DM.; The corresponding augmented Lagrangian function is: Among them, W and U are dual variables; ρ1>0 and ρ2>0 are penalty parameters; the alternating direction multiplication method is used to update the original variables and the dual variables in k+1 iterations.
7. The image denoising method based on graph smoothing RPCA as claimed in claim 6, characterized in that: The method for updating the original variable and the dual variable by the alternating direction multiplier method includes: IN k+1 =In k +ρ1(K k+1 -WITH k+1 ) U k+1 =U k +ρ2(S k+1 -V k+1 )。 8. The image denoising method based on graph smoothing RPCA as claimed in claim 7, characterized in that: The solutions for K and S are: K k+1 =((β+ρ1)I+2γL) -1 (β(MS k )+ρ1Z k-1 -W k ) 9. The image denoising method based on graph smoothing RPCA as claimed in claim 8, characterized in that: When calculating K, the Moore-Penrose generalized inverse method is used.
10. An image denoising system using the image denoising method based on graph smoothing RPCA as claimed in claim 1, characterized in that: include: A model building module, configured to obtain a graph smoothing-based RPCA model according to a non-convex penalty and graph learning method; The processing module is configured to process the RPCA model based on graph smoothing through an ADMM algorithm.
Citation Information
Patent Citations
Metallographic image denoising method and system
CN112435175A
Method for Recovering Low-Rank Matrices and Subspaces from Data in High-Dimensional Matrices
US20130191425A1
