CN112801884B - Image denoising method based on external non-local self-similarity and improved sparse representation - Google Patents
Image denoising method based on external non-local self-similarity and improved sparse representation Download PDFInfo
- Publication number
- CN112801884B CN112801884B CN202011354601.4A CN202011354601A CN112801884B CN 112801884 B CN112801884 B CN 112801884B CN 202011354601 A CN202011354601 A CN 202011354601A CN 112801884 B CN112801884 B CN 112801884B
- Authority
- CN
- China
- Prior art keywords
- image
- external
- block
- block group
- prior
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 44
- 239000000203 mixture Substances 0.000 claims abstract description 15
- 239000011159 matrix material Substances 0.000 claims description 17
- 239000013598 vector Substances 0.000 claims description 16
- 239000000126 substance Substances 0.000 claims description 9
- 238000012545 processing Methods 0.000 claims description 8
- 230000008569 process Effects 0.000 claims description 7
- 238000013459 approach Methods 0.000 claims description 4
- 230000002776 aggregation Effects 0.000 claims description 3
- 238000004220 aggregation Methods 0.000 claims description 3
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 238000000354 decomposition reaction Methods 0.000 claims description 3
- 230000006870 function Effects 0.000 claims description 3
- 238000009499 grossing Methods 0.000 claims description 3
- 230000008901 benefit Effects 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000003706 image smoothing Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000005192 partition Methods 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 238000013139 quantization Methods 0.000 description 1
- 238000000638 solvent extraction Methods 0.000 description 1
- 230000001131 transforming effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/70—Denoising; Smoothing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/40—Analysis of texture
- G06T7/41—Analysis of texture based on statistical description of texture
- G06T7/44—Analysis of texture based on statistical description of texture using image operators, e.g. filters, edge density metrics or local histograms
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/90—Determination of colour characteristics
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an image denoising method based on external non-local self-similarity and improved sparse representation, which comprises the following steps: (1) dividing an external clean image dataset into block groups; (2) dividing the block group of the external clean image into an external flat sliding block group and an external texture block group; (3) learning an external sliding block set prior; (4) learning an external texture block set prior through a Gaussian mixture model; (5) dividing the noise image into block groups; (6) guiding the block clustering of the noise images by using external non-local self-similar prior, and simultaneously calculating the flat sliding block group ratio of the noise images; (7) and optimizing the regularization parameters of the sparse representation model by using the flat sliding block group ratio, and respectively recovering the image blocks in each subspace according to the improved sparse representation model. The self-adaptability of the sparse representation model is improved by optimizing the regularization parameters of the sparse representation model, and the technical problems of low noise-to-noise ratio of a denoised image peak value and detail information loss in the existing image denoising method are solved.
Description
Technical Field
The invention relates to the technical field of image processing, in particular to an image denoising method based on external non-local self-similarity and improved sparse representation.
Background
Through decades of development, digital image processing technology has become more and more mature, and plays a great role in the fields of climate monitoring, medical diagnosis, face recognition, defect detection and the like. Digital images are inevitably interfered by noise in the processes of generation, transmission and conversion, and the operations of compressing, feature extraction and the like on the images containing the noise cause the degradation of image quality. Therefore, the problem of estimating an original clean image from an image containing noise has become one of important research contents.
The image denoising algorithm is mainly divided into a spatial domain method and a transform domain method, and the spatial domain method recovers an image by using the correlation of image pixels. Unlike the spatial domain method, the transform domain method requires transforming the noise image into another domain that facilitates distinguishing the noise signal from the original image signal, thereby removing the noise by filtering, and finally restoring the processed image into the original domain.
The non-local self-similar prior denoising method belongs to a spatial domain method, and the core idea is to recover an image according to the fact that many similar image blocks exist in the image. For example, selecting any image block, many similar image blocks can be found at other positions of the image. Therefore, when an image block is contaminated by noise, the image block can be restored by using an image block similar to the image block. The non-local self-similar prior method has higher flexibility, and can extract similar image blocks from a noise image or an external image, wherein the similar image blocks are called as external non-local self-similar prior. As can be seen from the principle of the non-local self-similarity method, this method only provides sufficient image texture information for restoring the image. In recent years, sparse representation models have become a focus of research because combining non-local self-similarity methods with sparse representation models can efficiently remove noise in images. The core idea of the sparse representation model is to extract orthogonal vectors capable of representing key texture detail information of an image from non-local self-similar prior, and the image can be recovered by linearly combining the orthogonal vectors. In the sparse representation process, the image signal has sparsity, and the noise signal does not have sparsity, so that the noise is filtered.
In 2015, Xu Jun et al first proposes a Patch Group Prior Denoising (PGPD) model, the PGPD model learns non-local self-similar Prior information of an external clean image by using a gaussian mixture model, finds external non-local self-similar Prior information most matched with a noise image block by a probability matching method, and finally performs sparse representation on the external non-local self-similar Prior to restore the image. The PGPD model has the advantage of finding texture and detail information suitable for recovering an image, for example, in the case of a high noise level, the texture and detail information of the image are seriously damaged by noise, it is difficult to extract accurate texture information from a noisy image itself to recover the image, and an external non-local self-similar prior is derived from an external clean image without considering the interference of the noise on the external non-local self-similar prior.
However, the PGPD model cannot solve the problem that the sparse representation model lacks adaptability, and for an image with a low noise level, the sparse representation model can generate the problem of incomplete denoising, so that noise residue exists in the image; for an image with a higher noise level, the sparse representation model can destroy the texture and detail information of the image while removing noise, and the utility of external non-local self-similarity prior cannot be effectively exerted. The sparse representation model has the defect that the regularization parameters of the model cannot be automatically adjusted. Therefore, the regularization parameters of the sparse representation model are optimized, so that the regularization parameters can effectively remove noise and simultaneously retain texture and detail information of the image as much as possible.
Disclosure of Invention
The invention aims to provide an image denoising method based on external non-local self-similarity and improved sparse representation, so as to solve the problems in the background technology.
The invention aims to provide an image denoising method based on external non-local self-similarity and improved sparse representation aiming at the problems in the prior art, and the method increases the self-adaptability of a sparse representation model by optimizing the regularization parameter of the sparse representation model so as to solve the technical problems of low peak signal-to-noise ratio of a denoised image and detail information loss in the conventional image denoising method. To achieve the above object, the present invention is characterized by comprising the steps of:
(1) dividing an external clean image dataset into block groups;
(2) dividing the block group of the external clean image into an external flat sliding block group and an external texture block group;
(3) learning an external sliding block set prior;
(4) learning an external texture block set prior through a Gaussian mixture model;
(5) dividing the noise image into block groups;
(6) guiding the noise image block clustering by using the external flat sliding block group and the external texture block group, and simultaneously calculating the flat sliding block group ratio of the noise image;
(7) and optimizing the regularization parameters of the sparse representation model by using the flat block group ratio, and respectively recovering the image blocks in each subspace according to the improved sparse representation model and the external non-local self-similar prior.
In the step (1), the external clean image data set is divided into block groups according to a non-local self-similarity method. A reference image block (with the size of p multiplied by p) in the external clean image is selected, M-1 image blocks most similar to the reference image block can be found in the range of W multiplied by W of a search window by utilizing Euclidean distance, and then the M similar image blocks are combined into a block group. Suppose that
Represents a block group, then
Representing a block vector consisting of three color channels in an image block, the block group with the dc component removed is represented as:
wherein the content of the first and second substances,
representing the dc component of the block set X. After the direct current component of the block group is removed, the similarity between the block groups can be obviously increased, so that the number of Gaussian components required by the block group fitting is reduced.
In the step (2), the purpose of classifying the external clean image block group is to divide the block group into an external flat slider group and an external texture block group. The sliding block set does not belong to key image information a priori, and a sliding block set space can be constructed quickly; the texture block group stores key information such as texture and detail of the image, and an external texture prior space needs to be constructed by a learning method. Calculate each block group
Variance of, if block
If the maximum value of the variance of the middle image block is less than or equal to the threshold value xi, the image block is combined into a smooth block group, and the smooth block group is marked as a smooth block group
If a block group
If the maximum value of the variance of the middle image block is greater than the threshold value xi, classifying the block group as a texture block group and marking the texture block group as the texture block group
In step (3), it is assumed that a total of N is extracted from the external clean image dataset1A flat slider group, the outer flat slider group space is expressed as
Because the flat sliding block group does not belong to key image information a priori, a smooth priori subspace can be directly constructed, and P is removed firstlySHas a direct current component of
Then solve for PSThe covariance matrix of
Finally obtaining external smooth prior subspace N (0, Sigma)k) And N (-) represents a Gaussian component. Because of PSDC removal is performed so that the Gaussian components N (0, Sigma)k) Is a zero vector.
In step (4) above, it is assumed that a total of N is extracted from the external clean image dataset2Individual texture block groups, representing texture block group space as
Since texture block sets belong to key image information a priori, a Gaussian mixture model is used to fit the block set PxThe non-local self-similar prior, gaussian mixture model has the advantage of being able to efficiently partition the external prior subspace, i.e. each gaussian component represents one prior subspace. The technical scheme is that the parameters of the Gaussian mixture model are solved by using an EM algorithm for iteration J times, if the difference value between the current iteration result and the last iteration result is smaller than an iteration threshold theta, the iteration process is ended in advance, and finally K external texture prior subspaces are obtained
Each subspace N (0, Σ)k) Image information, Σ, describing certain textureskA covariance matrix representing the gaussian component.
In the step (5), the noise image is divided into blocks to increase the matching degree between the noise image blocks and the external blocks, and then the image is restored by selecting the external blocks which are most matched with each other a priori. One reference image block (with the size of p multiplied by p) in the noise image is selected, M-1 image blocks which are most similar to the reference image block are found in the W multiplied by W range of a search window, and the M similar image blocks are combined into a block group. Suppose that
Representing a group of blocks in a noisy image, in which
Representing a column vector formed by combining the three color channels of the image block. And performing direct current component removal processing on the block group to obtain:
wherein the content of the first and second substances,
group of presentation blocks
The purpose of removing the dc component from the noise image block group is to increase the prior matching degree with the external block group.
In the step (6), the clustering aims to classify similar block groups in the noise image into the same subspace, and the flat sliding block group ratio is used for optimizing the regularization coefficient of the sparse representation model. Compared with the whole image, the subspace only stores certain specific types of image information, and the denoising treatment of each subspace one by one can obviously reduce the difficulty of the denoising problem. Fitting the outer prior subspace
The noise image block group is guided to be clustered as a reference subspace, and the speed of clustering the noise image block group can be remarkably increased because an external prior subspace is prepared in advance. In groups of blocks
For example, the handle block group
Projection into an external prior subspace
In the method, the probability estimation is reused for each block group in the noise image
Finding the best matching outer prior subspace:
wherein the content of the first and second substances,
group of presentation blocks
The probability of matching the kth external prior subspace, K1, 2, 3. Assume that a total of L blocks are extracted from the noisy image, with a subspace N (0, ∑ of the outer smoothing prior)k) The most matched block group number is L0Then the flat-slider group ratio of the noisy image is σ ═ L0The noise image smoothing block set ratio indirectly reflects the noise level of the noise image. The higher the noise level of the image, the smaller the value of the smooth block group ratio; the lower the noise level of the image, the greater the value of the smooth block group ratio.
In the step (7), the key of the sparse representation model is to construct a dictionary and sparse coding coefficients, and the noise image can be restored by obtaining a dictionary matrix and a sparse coding coefficient matrix. The dictionary is composed of a plurality of feature vectors that hold key information in the image. The role of the sparse coding coefficients is to pick the appropriate feature vectors to recover the image, so that a signal can be represented by a few feature vectors. Hypothesis and block set
The best matching external prior subspace is N (0, Sigma)k) For covariance matrix ∑kPerforming singular value decomposition to obtain:
wherein D iskIs an external dictionary holding feature vectors, SkIs a diagonal matrix holding eigenvalues, diagonal matrix SkThe size of the middle characteristic value characterizes the external dictionary DkThe degree of importance of the medium feature vector. In groups of blocks
Image block of
For example, according to the sparse representation model principle, estimating sparse coding coefficients:
wherein D iskIs formed by an external prior subspace sigmakDecomposed external dictionary, SkIs a reaction of with DkThe corresponding eigenvalue matrix, exp (-) represents an exponential function, δ is an artificially set threshold, and ε is a positive number that approaches zero. According to the regularization term on the right side of the equation (5), the regularization parameter λ is adjusted by using the property that the exponential function is monotonically increased, when the noise level of the image is higher, the number of the smooth block groups in the image is smaller, the value of the smooth block group ratio σ is smaller, and the value of exp (δ - σ) is larger, so that the value of λ is increased, and the texture and detail information of the image is prevented from being damaged by the sparse representation model. Conversely, when the noise level of the image is lower, the larger the number of the slider groups in the image, the larger the value of the slider group ratio σ, the smaller the value of exp (δ - σ), and therefore the value of λ can be reduced, thereby avoiding the problem of incomplete denoising caused by the sparse representation model.
The formula (5) expresses the method for optimizing the sparse representation model by using the sliding block group ratio, and simultaneously describes the process of estimating the coding coefficient by using the improved sparse representation model, and the invention adopts a soft threshold method to solve the sparse coding coefficient in the formula (5):
wherein the content of the first and second substances,
λ is the regularization parameter and ε is a positive number that approaches zero.
Finally, an external dictionary D is utilizedkAnd sparse coding coefficient alphan,mFor yn,mSparse representation is carried out to obtain:
yn,m＝Dkαn,m+μy (7)
wherein, muyAs a block group YnThe whole image can be recovered by carrying out aggregation processing on the denoised image blocks.
Compared with the prior art, the invention has the beneficial effects that:
the method is divided into two stages, wherein the first stage is a preparation stage, namely learning external non-local self-similar prior; the second stage is a denoising stage, i.e., removing noise using an external non-local self-similar prior. The first stage and the second stage are independent from each other, so that the denoising efficiency can be accelerated.
The regularization parameters of the sparse representation model are optimized by utilizing the smooth block group ratio of the noise image, so that the adaptability of the sparse representation model is improved, and the sparse representation model is suitable for recovering images with different noise levels. Compared with the prior art, the method can efficiently remove noise, simultaneously can keep the details and texture information of the image as much as possible, and can remarkably improve the peak signal-to-noise ratio of the image.
Drawings
FIG. 1 is a flow chart of an image denoising method based on an external non-local self-similar prior and an improved sparse representation.
Detailed Description
The present invention will be further described with reference to the following examples, which are intended to illustrate only some, but not all, of the embodiments of the present invention. Based on the embodiments of the present invention, other embodiments used by those skilled in the art without any creative effort belong to the protection scope of the present invention.
Example 1:
as shown in fig. 1, the method is a flow of an image denoising method based on external non-local self-similarity and improved sparse representation, an algorithm needs to be iteratively processed when a denoising process is completed, and in order to improve denoising efficiency, clustering processing is performed on a noise image block group only when the number of iterations is odd. In addition, the smooth block group ratio of the noisy image only needs to be calculated at the first iteration.
The image denoising method based on the external non-local self-similarity and the improved sparse representation comprises the following steps:
(1) partitioning an external clean image dataset into block groups
Kodak PhotoCD dataset (http:// r0k. us/graphics/Kodak /) was used as the external clean image dataset, which contained 24 high quality natural images, each with dimensions of 500 × 500. If the value of the side length p of an image block is 6, the size of the image block is 6 × 6. In this embodiment, the side length W of the search window is 31, and the number M of image blocks in the block group is 10. Using Euclidean distance to combine M similar image blocks into a block group, and removing DC component of the block group to obtain
(2) Grouping blocks of an external clean image into external flat slider groups and external texture block groups
In the present embodiment, the variance threshold ξ ═ 0.001, is calculated for each block group
Variance of, if block
If the maximum value of the variance of the middle image block is less than or equal to the threshold value xi, the image block is combined into a smooth block group, and the smooth block group is marked as a smooth block group
If a block group
If the maximum value of the variance of the middle image block is greater than the threshold value xi, classifying the block group as a texture block group and marking the texture block group as the texture block group
(3) Learning external smooth block set priors
Because the flat sliding block group does not belong to key image information a priori, the flat sliding block group can be used for correcting the key image informationConstructing a smooth prior subspace to obtain an external smooth prior subspace N (0, sigma)k) And N (-) represents a Gaussian component. Because the outer flat slider group P is assembledSThe DC component removing process is performed so that the Gaussian components N (0, Sigma)k) Is a zero vector.
(4) Learning external texture block set priors through a Gaussian mixture model
In this embodiment, the number K of gaussian components in the gaussian mixture model is 32, the number of iterations for solving the gaussian mixture model by the EM algorithm is J100, and the iteration threshold θ of the EM algorithm is 1 × 10-10. Fitting an external texture block set P with a Gaussian mixture modelxThe non-local self-similar prior is used for solving the parameters of the Gaussian mixture model by using an EM algorithm to finally obtain K external texture prior subspaces
Each subspace describes some texture-specific image information, ΣkA covariance matrix representing the gaussian component.
(5) Dividing noisy images into blocks
A noise image is selected from the CC15 dataset, the size of the noise image being 512 x 512. Setting the side length p of image block to 6, selecting one reference image block (with size p × p) in the noise image, finding M-1 image blocks most similar to the reference image block in the search window W × W, combining M similar image blocks into one block group, and removing DC component of the block group
(6) External non-local self-similar prior guiding noise image block clustering
As can be seen from fig. 1, the whole algorithm needs to be iterated to complete denoising, and the number of iterations of the whole algorithm is set to 6 in this embodiment. In order to accelerate the denoising efficiency, the noise image is clustered only when the iteration number is odd. Handle block group
Projection into an external prior subspace
In the method, the probability estimation is reused for each block group in the noise image
The best matching outer prior subspace is found. Assuming that a total of L blocks are extracted from the noisy image, as can be seen from FIG. 1, in the first iteration of the algorithm, the statistics and outer smoothing a priori subspaces N (0, ∑ are used for the first timek) Number of most matched blocks L0Thereby calculating the flat slider group ratio sigma-L of the noisy image0/L。
(7) Optimizing sparse representation models
For covariance matrix ∑kSingular value decomposition is carried out to obtain an external dictionary DkAnd eigenvalue matrix SkIn groups of blocks
Image block of
For example, estimating sparse coding coefficients yields:
where the regularization coefficient λ of the sparse representation model is 0.0005, ∈ is a positive number close to zero, the regularization parameter threshold δ is 0.5, and σ is a flat-slider group ratio.
Solving sparse coding coefficients by a soft threshold method to obtain:
wherein the content of the first and second substances,
λ is regularThe quantization parameter, ε, is a positive number that approaches zero.
Finally, an external dictionary D is utilizedkAnd sparse coding coefficient alphan,mFor yn,mSparse representation is carried out to obtain:
yn,m＝Dkαn,m+μy
wherein, muyAs a block group YnThe whole image can be recovered by carrying out aggregation processing on the denoised blocks.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents and improvements made within the spirit and principle of the present invention are intended to be included within the scope of the present invention.
Claims (1)
1. The image denoising method based on external non-local self-similarity and improved sparse representation is characterized by comprising the following steps of:
(1) dividing an external clean image dataset into block groups;
(2) dividing the block group of the external clean image into an external flat sliding block group and an external texture block group;
(3) learning an external sliding block set prior;
(4) learning an external texture block set prior through a Gaussian mixture model;
(5) dividing the noise image into block groups;
(6) guiding the block clustering of the noise images by using external non-local self-similar prior, and simultaneously calculating the flat sliding block group ratio of the noise images;
(7) utilizing a flat sliding block group ratio to optimize regularization parameters of the sparse representation model, and respectively recovering image blocks in each subspace according to the improved sparse representation model;
the step (1) of dividing the external clean image data set into block groups comprises dividing the external clean image data set into block groups according to a non-local self-similarity method, selecting a reference image block in the external clean image, wherein the size of the reference image block is p multiplied by p, and M-1 image blocks which are most similar to the reference image block can be found in a W multiplied by W range of a search window by using Euclidean distance,further, the M similar image blocks are combined into a block group; suppose that
Represents a block group, then
Representing a block vector consisting of three color channels in an image block; the block group from which the dc component is removed is represented as:
wherein the content of the first and second substances,
a direct current component representing the block set X;
the step (2) of dividing the block groups of the external clean image into an external flat block group and an external texture block group includes calculating each block group
Variance of, if block
If the maximum value of the variance of the middle image block is less than or equal to the threshold value xi, the image block is combined into a smooth block group, and the smooth block group is marked as a smooth block group
If a block group
If the maximum value of the variance of the middle image block is greater than the threshold value xi, classifying the block group as a texture block group and marking the texture block group as the texture block group
Learning an external smooth block set prior in step (3), including assuming that N is extracted altogether from the external clean image dataset1A flat slider group, the outer flat slider group space is expressed as
Removing PSHas a direct current component of
Then solve for PSThe covariance matrix of
Finally obtaining external smooth prior subspace N (0, Sigma)k) N (·) denotes a Gaussian component; because of PSDC removal is performed so that the Gaussian components N (0, Sigma)k) The mean value of (a) is a zero vector;
learning an external texture block set prior through a Gaussian mixture model in the step (4), wherein N is assumed to be extracted from an external clean image dataset2Individual texture block groups, representing texture block group space as
Because the texture block group is prior to the key image information, a Gaussian mixture model is used for fitting the block group set PxThe parameters of the Gaussian mixture model are solved by using an EM algorithm to iterate J times, if the difference value between the iteration result of the time and the iteration result of the last time is less than an iteration threshold theta, the iteration process is ended in advance, and finally K external texture prior subspaces are obtained
Each subspace N (0, Σ)k) Image information, Σ, describing certain textureskA covariance matrix representing the gaussian component;
the step (5) of dividing the noise image into block groups comprises selecting a reference image block in the noise image, wherein the size of the reference image block is p multiplied by p, finding M-1 image blocks which are most similar to the reference image block in a W multiplied by W range of a search window, and combining the M similar image blocks into one block group; suppose that
Representing a group of blocks in a noisy image, in which
Representing a column vector formed by combining three color channels of an image block; and performing direct current component removal processing on the block group to obtain:
wherein the content of the first and second substances,
group of presentation blocks
A direct current component of (a);
the step (6) of guiding the block group clustering of the noise image by using external non-local self-similar prior and simultaneously calculating the smooth block group ratio of the noise image comprises the step of guiding the external prior subspace
As a reference subspace to guide the clustering of noisy image patches
Projection into an external prior subspace
In the method, the probability estimation is reused for each block group in the noise image
Finding the best matching outer prior subspace:
wherein the content of the first and second substances,
group of presentation blocks
Probability of matching with kth outer prior subspace, K ═ 1,2,3, …, K + 1; assume that a total of L blocks are extracted from the noisy image, with a subspace N (0, ∑ of the outer smoothing prior)k) The most matched block group number is L0Then the flat-slider group ratio of the noisy image is σ ═ L0/L；
The step (7) optimizes the regularization parameters of the sparse representation model by utilizing the flat sliding block group ratio, and respectively recovers the image blocks in each subspace according to the improved sparse representation model, wherein the image blocks comprise hypothesis and block groups
The best matching external prior subspace is N (0, Sigma)k) For covariance matrix ∑kPerforming singular value decomposition to obtain:
wherein D iskIs an external dictionary holding feature vectors, SkIs a diagonal matrix holding eigenvalues, diagonal matrix SkMiddle featureThe magnitude of the value characterizes the external dictionary DkThe degree of importance of the medium feature vector;
according to the sparse representation model principle, estimating sparse coding coefficients to obtain:
wherein D iskIs formed by an external prior subspace sigmakDecomposed external dictionary, SkIs a reaction of with DkA corresponding eigenvalue matrix, exp (·) represents an exponential function, δ is an artificially set threshold, and ε is a positive number approaching zero;
using an external dictionary DkAnd sparse coding coefficient alphan,mFor yn,mSparse representation is carried out to obtain:
yn,m＝Dkαn,m+μy (7)
wherein, muyAs a block group YnThe whole image can be recovered by carrying out aggregation processing on the denoised blocks by using the direct current component;
sparse coding coefficient:
wherein the content of the first and second substances,
λ is the regularization parameter and ε is a positive number that approaches zero.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011354601.4A CN112801884B (en) | 2020-11-26 | 2020-11-26 | Image denoising method based on external non-local self-similarity and improved sparse representation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011354601.4A CN112801884B (en) | 2020-11-26 | 2020-11-26 | Image denoising method based on external non-local self-similarity and improved sparse representation |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112801884A CN112801884A (en) | 2021-05-14 |
| CN112801884B true CN112801884B (en) | 2022-04-05 |
Family
ID=75806303
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011354601.4A Active CN112801884B (en) | 2020-11-26 | 2020-11-26 | Image denoising method based on external non-local self-similarity and improved sparse representation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112801884B (en) |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105279740A (en) * | 2015-11-17 | 2016-01-27 | 华东交通大学 | Image denoising method based on sparse regularization |
| US9317903B1 (en) * | 2013-08-27 | 2016-04-19 | Pixelworks, Inc. | Self-similarity based single frame super-resolution |
| CN106169180A (en) * | 2016-07-13 | 2016-11-30 | 桂林电子科技大学 | A kind of non-local sparse based on group represents additive noise minimizing technology |
| CN107004255A (en) * | 2014-09-23 | 2017-08-01 | 英特尔公司 | The non-local mean image denoising retained with details of mixing is driven using self-similarity |
| CN107038730A (en) * | 2017-03-30 | 2017-08-11 | 湖北工业大学 | The rarefaction representation image rebuilding method being grouped based on Gauss mesostructure block |
| CN107067367A (en) * | 2016-09-08 | 2017-08-18 | 南京工程学院 | A kind of Image Super-resolution Reconstruction processing method |
| CN107358589A (en) * | 2017-07-17 | 2017-11-17 | 桂林电子科技大学 | A kind of combination histogram of gradients and the denoising method of low-rank constraint |
| CN109636734A (en) * | 2018-10-29 | 2019-04-16 | 昆明理工大学 | A kind of sparse regular terms of weighting based on group rarefaction representation constrains image de-noising method |
-
2020
- 2020-11-26 CN CN202011354601.4A patent/CN112801884B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9317903B1 (en) * | 2013-08-27 | 2016-04-19 | Pixelworks, Inc. | Self-similarity based single frame super-resolution |
| CN107004255A (en) * | 2014-09-23 | 2017-08-01 | 英特尔公司 | The non-local mean image denoising retained with details of mixing is driven using self-similarity |
| CN105279740A (en) * | 2015-11-17 | 2016-01-27 | 华东交通大学 | Image denoising method based on sparse regularization |
| CN106169180A (en) * | 2016-07-13 | 2016-11-30 | 桂林电子科技大学 | A kind of non-local sparse based on group represents additive noise minimizing technology |
| CN107067367A (en) * | 2016-09-08 | 2017-08-18 | 南京工程学院 | A kind of Image Super-resolution Reconstruction processing method |
| CN107038730A (en) * | 2017-03-30 | 2017-08-11 | 湖北工业大学 | The rarefaction representation image rebuilding method being grouped based on Gauss mesostructure block |
| CN107358589A (en) * | 2017-07-17 | 2017-11-17 | 桂林电子科技大学 | A kind of combination histogram of gradients and the denoising method of low-rank constraint |
| CN109636734A (en) * | 2018-10-29 | 2019-04-16 | 昆明理工大学 | A kind of sparse regular terms of weighting based on group rarefaction representation constrains image de-noising method |
Non-Patent Citations (1)
| Title |
|---|
| Image denoising using group sparsity residual and external nonlocal self-similarity prior;Zhiyuan zha et al;《2017 IEEE international conference on image processing》;20180222;第2956-2960页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112801884A (en) | 2021-05-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107358589B (en) | A kind of denoising method of combination histogram of gradients and low-rank constraint | |
| CN109671029B (en) | Image denoising method based on gamma norm minimization | |
| CN111369457B (en) | Remote sensing image denoising method for sparse discrimination tensor robustness PCA | |
| CN111127354B (en) | Single-image rain removing method based on multi-scale dictionary learning | |
| CN110349112B (en) | Two-stage image denoising method based on self-adaptive singular value threshold | |
| CN112967210B (en) | Unmanned aerial vehicle image denoising method based on full convolution twin network | |
| CN110675422B (en) | Video foreground and background separation method based on generalized non-convex robust principal component analysis | |
| CN107292855B (en) | Image denoising method combining self-adaptive non-local sample and low rank | |
| CN113379618A (en) | Optical remote sensing image cloud removing method based on residual dense connection and feature fusion | |
| CN107451961B (en) | Method for recovering sharp image under multiple fuzzy noise images | |
| Onur | Improved image denoising using wavelet edge detection based on Otsu’s thresholding | |
| CN111126169A (en) | Face recognition method and system based on orthogonalization graph regular nonnegative matrix decomposition | |
| CN112801884B (en) | Image denoising method based on external non-local self-similarity and improved sparse representation | |
| Hu et al. | Patch-based low-rank minimization for image denoising | |
| CN110136164B (en) | Method for removing dynamic background based on online transmission transformation and low-rank sparse matrix decomposition | |
| Nejati et al. | Low-rank regularized collaborative filtering for image denoising | |
| CN112927169B (en) | Remote sensing image denoising method based on wavelet transformation and improved weighted kernel norm minimization | |
| CN114495037A (en) | Video prediction method and system based on key points and Kalman filtering | |
| Choi et al. | Fast super-resolution algorithm using ELBP classifier | |
| Wang et al. | A review of image denoising methods | |
| CN116109519B (en) | Image denoising method | |
| CN105139353B (en) | A kind of blind separating method for replacing aliased image | |
| CN117611484B (en) | Image denoising method and system based on denoising self-decoding network | |
| Liu et al. | The validity analysis of the non-local mean filter and a derived novel denoising method | |
| CN111709344B (en) | EPLL image illumination removal recognition processing method based on Gaussian mixture model |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |