CN110223229A - Based on the image magnification method of self-similarity, system and storage medium between scale - Google Patents
Based on the image magnification method of self-similarity, system and storage medium between scale Download PDFInfo
- Publication number
- CN110223229A CN110223229A CN201910433432.4A CN201910433432A CN110223229A CN 110223229 A CN110223229 A CN 110223229A CN 201910433432 A CN201910433432 A CN 201910433432A CN 110223229 A CN110223229 A CN 110223229A
- Authority
- CN
- China
- Prior art keywords
- image
- adaptivity
- input picture
- self
- low
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformation in the plane of the image
- G06T3/40—Scaling the whole image or part thereof
- G06T3/4023—Decimation- or insertion-based scaling, e.g. pixel or line decimation
Abstract
The invention discloses a kind of based on the image magnification method of self-similarity, system and storage medium between scale, the invention is obtained from training data set using input picture, the probability density function of the self-training data acquisition system is obtained based on gauss hybrid models, adaptivity prior image is obtained according to the high-resolution low-frequency image of the probability density function and the input picture, the output image of amplification is generated according to the low resolution fidelity image of the adaptivity prior image and the adaptivity prior image.The present invention is used in the reconstruction process of image amplification without external image training dataset using adaptivity as a kind of intrinsic property, and algorithm is novel, efficient, the better quality of image amplification, and the scope of application is wider.
Description
Technical field
The present invention relates to image processing techniques, it is especially a kind of based on the image magnification method of self-similarity between scale, be
System and storage medium.
Background technique
Image amplification be also known as image super-resolution, it refers to the low-resolution image that will input, calculate mathematics with
Under the support of the subjects theories such as machine learning, its corresponding high-definition picture is reconstructed in such a way that software calculates.It rebuilds
High-definition picture is consistent with the low-resolution image content of input, and size is proportional to become larger, and comprising low in reconstruction image
The high-frequency information that image in different resolution had not occurred.Because it has widely in fields such as video monitoring, medical diagnosis, remotely sensed images
Using being constantly subjected to the attention of researcher.
Image magnification method can be divided by the training data source used at present are as follows: utilize the image of extraneous training dataset
Amplification method and the image magnification method for using self-similarity.It is needed using the image magnification method of extraneous training dataset a large amount of
Extraneous high-definition picture constructs training dataset, and the scale and content of training dataset influence very the quality of enlarged drawing
Greatly.
Image self-similarity refers to (such as figure of 5 × 5,8 × 8 equidimensions when piece image is investigated from image sheet
Photo), any one small image sheet can be found from the lower scale version of the associated pictures of other angles or diagram picture
With its very similar or even identical content；It only need to be by means of input picture sheet using the image magnification method of self-similarity
Body construction training set carries out image amplification.On the basis of image self-similarity, Glasner etc. propose one it is unified super
Resolution algorithm frame, basic thought are the different views for the similar diagram photo in same scale being considered as Same Scene, are used
Similar diagram photo between different scale constructs training sample piece pair, then by the super-resolution of multiple image and based on learning method
Two kinds of thoughts of super-resolution integrated, using image self-similarity to input low-resolution image amplify.But
Edge is fuzzy or the image comprising noise in, the image sheet of arbitrary extracting is with it in corresponding lower scale in the input image
The nearest-neighbor found in version has larger difference, causes reconstructed results inaccurate；It is tight to will appear mistake for details position simultaneously
The high-frequency information of weight, to reduce the quality of enlarged drawing.
Summary of the invention
In order to solve the above technical problems, the purpose of the embodiment of the present invention is: providing a kind of based on self-similarity between scale
Image magnification method, system and storage medium；The inventive embodiments are without external image training dataset, image amplification
Better quality, the scope of application is wider.
The technical solution that first aspect of the embodiment of the present invention is taken is, a kind of image based on self-similarity between scale is put
Big method, comprising:
Training data set is obtained from using input picture；
The probability density function of the self-training data acquisition system is obtained based on gauss hybrid models；
Adaptivity priori is obtained according to the high-resolution low-frequency image of the probability density function and the input picture
Image；
It is generated according to the low resolution fidelity image of the adaptivity prior image and the adaptivity prior image
The output image of amplification.
Preferably, described to be directed to input picture, it is obtained from training data set, is specifically included:
Institute is obtained for s times to up-sampling after s times of down-sampling of input picture bicubic of result, then to the result bicubic
The low-frequency image of input picture is stated, wherein s is positive number；
All image sheets of the low-frequency image of the input picture and the input picture and corresponding connection are extracted, is obtained from
Training data set.
Preferably, the expression formula of the probability density function is as follows:
Wherein, yjWith y 'jIndicate the jth extracted from the low-frequency image of the input picture and the input picture respectively
A image sheet, K are the number for introducing Gaussian function, πk、μk、∑kWeight coefficient, the mean vector of respectively k-th Gaussian function
And covariance matrix,Indicate k-th of Gaussian function.
Preferably, the high-resolution low-frequency image according to the probability density function and the input picture is obtained from
Adaptability prior image, specifically includes:
Input picture bicubic is obtained into the high-resolution low-frequency image of the input picture for s times to up-sampling；
Existed according to all image sheets of the high-resolution low-frequency image of the input picture and high-resolution low-frequency image piece
Corresponding probability density in the probability density function obtains adaptivity prior image.
Preferably, described to be protected according to the low resolution of the adaptivity prior image and the adaptivity prior image
True image generates the output image of amplification, specifically includes:
Adaptivity prior image is done into downsampled and gaussian filtering process, obtains low point of adaptivity prior image
Resolution image；
According to the input picture and the low-resolution image of adaptivity prior image, adaptivity prior image is obtained
Low resolution fidelity image；
According to the low resolution fidelity image, adaptivity prior image and ratio system of the adaptivity prior image
Number, generates the output image of amplification.
Preferably, the calculation formula of the output image of the amplification are as follows:
Wherein,For fidelity term, Y indicates that input picture matrix, X indicate that adaptivity prior image, A indicate
Degrade matrix, and λ is proportionality coefficient,For adaptivity prior image expression formula form, N is in X
The image sheet quantity of extraction, Pi、AndRespectively indicate the probability of i-th of image sheet in X, the mean vector of Gaussian function and
Covariance matrix.
Preferably, the amplification method amplifies for color image, comprising:
Color image is transformed into yuv space from rgb space；
The channel U and V of the color image is amplified to target size according to bicubic sampling method；
The channel Y of the color image is amplified to the target size according to above-mentioned image magnification method；
The channel Y, U and V of the color image target size is merged, and is converted into the output of rgb space image.
The technical solution that second aspect of the embodiment of the present invention is taken is, a kind of image based on self-similarity between scale is put
Big system, comprising:
Self-training data acquisition system obtains module, for being obtained from training data set using input picture；
Probability density function computing module, for obtaining the probability of the self-training data acquisition system based on gauss hybrid models
Density function；
Adaptivity prior image determining module, for the high score according to the probability density function and the input picture
Resolution low-frequency image obtains adaptivity prior image；
Image generation module is exported, for according to the adaptivity prior image and the adaptivity prior image
Low resolution fidelity image generates the output image of amplification.
The technical solution that the third aspect of the embodiment of the present invention is taken is, a kind of image based on self-similarity between scale is put
Big system, comprising:
At least one processor；
At least one processor, for storing at least one program；
When at least one described program is executed by least one described processor, so that at least one described processor is realized
Image magnification method based on self-similarity priori described in above-mentioned first aspect.
The technical solution that fourth aspect of the embodiment of the present invention is taken is a kind of storage medium, wherein being stored with processor can
The instruction of execution, the executable instruction of the processor is when executed by the processor for executing base described in above-mentioned first aspect
In the image magnification method of self-similarity priori.
A kind of image magnification method based on self-similarity priori provided in an embodiment of the present invention, is obtained from input picture
Training data set is gathered without external image training, avoids the dependence to extraneous training dataset；It is mixed based on Gauss
Molding type obtains the probability density function of the self-training data acquisition system, avoids the neighbouring field search for being easy to introduce error,
Computer resource is saved；It is obtained according to the high-resolution low-frequency image of the probability density function of acquisition and input picture adaptive
Property prior image, and combine low resolution fidelity image generate amplification output image, using adaptivity as the intrinsic of image
Property, for constructing adaptivity prior image, algorithm is novel, efficient, while by high-resolution low-frequency image in view of output
In image, keep output picture structure more complete；The image magnification method based on self-similarity priori in the embodiment of the present invention, makes
The calculation method for exporting image is efficient, and high-quality, the scope of application is wider.
Detailed description of the invention
Fig. 1 is the flow diagram of the image magnification method based on self-similarity priori of one embodiment；
Fig. 2 is that the data of the image magnification method based on self-similarity priori of one embodiment generate schematic diagram；
Fig. 3 is the color image amplification process signal of the image magnification method based on self-similarity priori of one embodiment
Figure；
Fig. 4 is the image for amplifying 3 times based on Glasner method of one embodiment；
Fig. 5 is that the image magnification method based on self-similarity priori of one embodiment amplifies 3 times of image；
Fig. 6 is 3 times of amplification of the ideal image of one embodiment；
Fig. 7 is the structural schematic diagram of the image enhancement system based on self-similarity priori of one embodiment；
Fig. 8 is a kind of data processing equipment of image enhancement system based on self-similarity priori of one embodiment.
Specific embodiment
The present invention is further explained and is illustrated with specific embodiment with reference to the accompanying drawings of the specification.For of the invention real
The step number in example is applied, is arranged only for the purposes of illustrating explanation, any restriction is not done to the sequence between step, is implemented
The execution sequence of each step in example can be adaptively adjusted according to the understanding of those skilled in the art.
The flow diagram of the image magnification method based on self-similarity priori of one embodiment in Fig. 1 is please referred to, such as
Shown in Fig. 1, including step S1 to S4.
S1 is obtained from training data set using input picture.
In the present embodiment, the self-training data acquisition system is all figures of the low-frequency image of input picture and input picture
The data acquisition system for being correspondingly connected with relationship composition of photo；It is obtained from training data set with input picture, is schemed without the external world
As training set, the dependence to extraneous training dataset is avoided.
Preferably, described to be directed to input picture, it is obtained from training data set, referring to Fig. 2, specifically including:
Institute is obtained for s times to up-sampling after s times of down-sampling of input picture bicubic of result, then to the result bicubic
The low-frequency image of input picture is stated, wherein s is positive number；Extract the institute of the low-frequency image of the input picture and the input picture
There are image sheet and corresponding connection, is obtained from training data set.
Wherein, Y is input picture, and Y ' is after s times of down-sampling of input picture bicubic of result, again to the result double three
It is secondary to up-sampling s times obtain the low-frequency image of the input picture, yjWith y 'jIt indicates respectively from j-th of figure of the middle extraction of Y and Y '
Photo.All image sheets in Y and Y ' are extracted, and establish corresponding connection, being coupled indicates yjWith y 'jImage sheet data all become
At column vector, then two column vectors are joined end to end and become a column vector, obtains self-training data acquisition systemM table
Show that the sum of abstract image piece, T indicate to operate the transposition of vector.
S2 obtains the probability density function of the self-training data acquisition system based on gauss hybrid models.
In the present embodiment, gauss hybrid models are introduced to approach the probability density function of training set, to be obtained from
The probability density function of training data set avoids the neighbouring field search for being easy to introduce error, has saved computer resource.
Referring to Fig. 2, the high-resolution low-frequency image of X ' expression input picture, X ' is that input picture is adopted upwards through bicubic
S times of sample obtains, and X indicates adaptivity prior image, xi,x′iIt indicates respectively from i-th of image sheet of the middle extraction of X and X '.Root
According to the Self-similar Feature of image it is found that if X (X ') is clear enough, the image sheet x of arbitrary extracting in X (X ')i(x′i) in correspondence
Y (Y ') in can find similar or even identical image sheet yj(y′j), serial number i (j) indicates that the image sheet is in corresponding diagram
I-th (j) a image sheet extracted as in；Therefore,In self-training setProbability density function in answer
With the presence of highest probability, gauss hybrid models are introduced to approach the probability density function of training set.
Preferably, the expression formula of the probability density function is as follows:
Wherein, yjWith y 'jIndicate the jth extracted from the low-frequency image of the input picture and the input picture respectively
A image sheet, K are the number for introducing Gaussian function, πk、μk、ΣkWeight coefficient, the mean vector of respectively k-th Gaussian function
And covariance matrix,Indicate that k-th of Gaussian function, z indicate stochastic variable.According to the definition of gauss hybrid models
Known to:Expression are as follows:
For parameterSeek, using expectation-maximization algorithm (Expectation-
Maximization, EM).
It, can since each of training data sample is all spliced by corresponding image sheet in image Y and Y '
With by the mean vector μ in gauss hybrid modelskAnd ΣkTwo class parameters are rewritten are as follows:
It is first to obtain adaptivity according to the high-resolution low-frequency image of the probability density function and the input picture by S3
Test image.
In the present embodiment, it is obtained from according to the high-resolution low-frequency image of the probability density function of acquisition and input picture
Adaptability prior image, using adaptivity as the intrinsic property of image, for constructing adaptivity prior image, algorithm is new
It is clever, efficient.
Preferably, the high-resolution low-frequency image according to the probability density function and the input picture is obtained from
Adaptability prior image, specifically includes:
Input picture bicubic is obtained into the high-resolution low-frequency image of the input picture for s times to up-sampling；
Existed according to all image sheets of the high-resolution low-frequency image of the input picture and high-resolution low-frequency image piece
Corresponding probability density in the probability density function obtains adaptivity prior image.
In the present embodiment, for unknown adaptivity prior image X, the image sheet x of the arbitrary extracting from XiWith
In the image sheet x ' of the middle corresponding position of image X 'iIt is coupled, obtains corresponding connection vectorIt gathers in training
Probability density function present in probability can indicate are as follows:
In turn, xiProbability meet a specific Gaussian Profile:
Wherein,For the mean vector of this Gaussian Profile, expression formula are as follows:
For the covariance matrix of this Gaussian Profile, expression formula is
Occur in formula (9), (10)Specific representation are as follows:
In turn, the form of priori knowledge is determined are as follows:
Wherein, N is the image sheet quantity extracted in unknown adaptivity prior image X, PiThe extraction set for one
Matrix, for extracting i-th of image sheet in X.The priori knowledge as described in formula (10) is to adaptivity priori figure
As the similitude of each image sheet of X is measured, therefore, self-similarity is maximized first between this priori is known as scale
It tests.
S4 is raw according to the low resolution fidelity image of the adaptivity prior image and the adaptivity prior image
At the output image of amplification.
In the present embodiment, while meeting global self-similarity priori, output image is dropped output image by image
It, should also be similar as far as possible to the low-resolution image of input after matter model treatment.Therefore, output image combining adaptive is first
It tests the high-resolution low-frequency image of image, low resolution fidelity image and input picture and generates, so that output picture structure is more
Completely, better quality.
Preferably, described to be protected according to the low resolution of the adaptivity prior image and the adaptivity prior image
True image generates the output image of amplification, specifically includes:
Adaptivity prior image is done into downsampled and gaussian filtering process, obtains low point of adaptivity prior image
Resolution image；
According to the input picture and the low-resolution image of adaptivity prior image, adaptivity prior image is obtained
Low resolution fidelity image；
According to the low resolution fidelity image, adaptivity prior image and ratio system of the adaptivity prior image
Number, generates the output image of amplification.
Preferably, the calculation formula of the output image of the amplification are as follows:
Wherein, formula (11) indicates that the value of X keeps the result of entire expression formula minimum,For fidelity term, Y
Indicate that input picture matrix, X indicate that adaptivity prior image, A indicate the matrix that degrades, the X that fidelity term formula (11) calculates is logical
It crosses similar as far as possible to the image Y of input after degrading processing；λ is proportionality coefficient,It is adaptive
Property prior image expression formula form, N is the image sheet quantity extracted in X, this makes in the high-definition picture being calculated
Each image sheet xiAll defer to one specifically, mean value isCovariance matrix isGaussian Profile.
It is available (12) to cost function equation (11) further abbreviation
The solution of equation (12) is relatively simple, directly carries out derivation and enables derivative to be zero to equation (12), can obtain
Further arrange the final result that can must export Image estimation
A kind of image magnification method based on self-similarity priori provided in an embodiment of the present invention, is obtained from input picture
Training data set is gathered without external image training, avoids the dependence to extraneous training dataset；It is mixed based on Gauss
Molding type obtains the probability density function of the self-training data acquisition system, avoids the neighbouring field search for being easy to introduce error,
Computer resource is saved；It is obtained according to the high-resolution low-frequency image of the probability density function of acquisition and input picture adaptive
Property prior image, and combine low resolution fidelity image generate amplification output image, using adaptivity as the intrinsic of image
Property, for constructing adaptivity prior image, algorithm is novel, efficient, while by high-resolution low-frequency image in view of output
In image, keep output picture structure more complete；The image magnification method based on self-similarity priori in the embodiment of the present invention, makes
The calculation method for exporting image is efficient, better quality.In addition, self-training data acquisition system is obtained according to input picture training, it can be with
It is adjusted automatically according to the difference of input picture, keeps the robustness of the amplification method more preferable, be highly suitable for DTV
Etc. computing capabilitys it is not strong, there is no an extra storage medium, and need the display equipment of quickly display amplification result, and easily scale to
The fields such as image denoising, image restoration, the scope of application are wider.
Below by taking the color image amplification process of the image magnification method based on self-similarity priori as an example, provide it is a kind of compared with
Good way of example please refers to the color image amplification process of the image magnification method based on self-similarity priori in Fig. 3
Schematic diagram.
Preferably, the amplification method amplifies for color image, comprising:
Color image is transformed into yuv space from rgb space；
The channel U and V of the color image is amplified to target size according to bicubic sampling method；
The channel Y of the color image is amplified to the target size according to above-mentioned image magnification method；
The channel Y, U and V of the color image target size is merged, and is converted into the output of rgb space image.
As shown in figure 3, input picture is the color image of rgb space, the rgb space of color image is first transformed into YUV
Space.Since human eye is relatively sluggish to the variation of Color Channel, more strong to the variation impression of luminance channel；Therefore, with original
It manages bicubic algorithm that is simple and can completing in real time and the channel U and V is directly amplified to specified target size, using above-mentioned base
The channel Y is amplified in the image magnification method of self-similarity priori, then the channel U, V and Y is merged, and be converted into rgb space figure
As output.
In the present embodiment, the luminance channel Y of human eye sensitivity in color image is only based on self-similarity using above-mentioned
The image of priori amplifies, and directly amplifies specified target ruler using bicubic algorithm to the channel U and V of not human eye sensitivity
It is very little, it both ensure that the quality of color image after amplification, also reduce the calculating of computer.
Due to self-similarity meeting with the reduction to input picture scale, between input picture and this low scale image
Gradually weaken；Therefore, amplify lesser multiple every time, such as 1.25 times, then using the result of amplification as input next time, after
Continuous 1.25 times of amplification, until reaching target multiple.If in last time iteration, target image size and present image size are discontented
Current input image is then used bicubic method to be amplified to target image size as the low frequency of output image by 1.25 times of relationships of foot
Version is further continued for being enlarged.
As Fig. 4 amplifies 3 times based on Glasner method and the image magnification method shown in fig. 5 based on self-similarity priori
Image and ideal enlarged drawing as shown in FIG. 6；It can be seen from the figure that the image based on self-similarity priori amplifies
Method is obviously better than the picture quality based on Glasner method, also closer to ideal enlarged drawing.
As shown in fig. 7, the embodiment of the invention also provides a kind of image enhancement system based on self-similarity between scale, packet
It includes:
Self-training data acquisition system obtains module, for being obtained from training data set using input picture；
Probability density function computing module, for obtaining the probability of the self-training data acquisition system based on gauss hybrid models
Density function；
Adaptivity prior image determining module, for the high score according to the probability density function and the input picture
Resolution low-frequency image obtains adaptivity prior image；
Image generation module is exported, for according to the adaptivity prior image and the adaptivity prior image
Low resolution fidelity image generates the output image of amplification.
In an alternate embodiment of the invention, the self-training data acquisition system obtains module, specifically includes:
Sampling unit, for by after s times of down-sampling of input picture bicubic of result, then it is upward to the result bicubic
S times of sampling obtains the low-frequency image of the input picture, and wherein s is positive number；
Gather acquiring unit, all image sheets of the low-frequency image for extracting the input picture and the input picture
And corresponding connection, it is obtained from training data set.
In an alternate embodiment of the invention, adaptivity prior image determining module, specifically includes:
High-resolution low-frequency image acquiring unit, for input picture bicubic to be obtained the input for s times to up-sampling
The high-resolution low-frequency image of image；
Adaptivity prior image unit, for all images according to the high-resolution low-frequency image of the input picture
Piece and high-resolution low-frequency image the piece corresponding probability density in the probability density function obtain adaptivity priori figure
Picture.
In an alternate embodiment of the invention, the output image generation module, specifically includes:
Low-resolution image acquiring unit, for adaptivity prior image to be done downsampled and gaussian filtering process,
Obtain the low-resolution image of adaptivity prior image；
Low resolution fidelity image acquiring unit, for according to low point of the input picture and adaptivity prior image
Resolution image obtains the low resolution fidelity image of adaptivity prior image；
Image generation unit is exported, for according to the low resolution fidelity image of the adaptivity prior image, adaptive
Answering property prior image and proportionality coefficient generate the output image of amplification.
Suitable for this system embodiment, this system embodiment is implemented content in above method embodiment
Function is identical as above method embodiment, and the beneficial effect reached and above method embodiment beneficial effect achieved
It is identical.
As shown in figure 8, the embodiment of the invention also provides a kind of data processing equipments, comprising:
At least one processor；
At least one processor, for storing at least one program；
When at least one described program is executed by least one described processor, so that at least one described processor is realized
The data processing method.
Suitable for this system embodiment, this system embodiment is implemented content in above method embodiment
Function is identical as above method embodiment, and the beneficial effect reached and above method embodiment beneficial effect achieved
It is identical.
In addition, the embodiment of the invention also provides a kind of storage mediums, wherein being stored with the executable instruction of processor, institute
The executable instruction of processor is stated when executed by the processor for executing the data processing method.
In some selectable embodiments, the function/operation mentioned in a block diagram can not be mentioned according to operational illustrations
The sequence arrived occurs.For example, depending on related function/operation, two boxes continuously shown can actually be by substantially
On simultaneously execute or the box can be performed sometimes with reverse order.In addition, presented in flow chart of the invention and
The embodiment of description is provided in an illustrative manner, and it is an object of the present invention to provide technology is more completely understood.Disclosed method is not
It is limited to operation presented herein and logic flow.Selectable embodiment is it is contemplated that the wherein sequence quilt of various operations
The sub-operation of a part for changing and being wherein described as larger operation is executed independently.
Although in addition, describing the present invention under the background of functional module and being illustrated in the form of block diagram
It is bright, but it is to be understood that, unless otherwise indicated, one or more of the function and/or feature can be collected
At in single physical device and/or software module or one or more functions and/or feature can be filled in individual physics
Set or software module in be implemented.It will also be appreciated that the practical realization in relation to each module is discussed in detail for understanding
The present invention is unnecessary.More specifically, it is contemplated that the attribute of various functional modules, function in device disclosed herein
In the case where internal relations, it will understand that the practical realization of the module in the routine techniques of engineer.Therefore, this field skill
Art personnel can realize this illustrated in detail in the claims hair with ordinary skill in the case where being not necessarily to undue experimentation
It is bright.It will also be appreciated that disclosed specific concept is merely illustrative, it is not intended to limit the scope of the present invention, this
The range of invention is determined by the full scope of the appended claims and its equivalent program.
It is to be illustrated to preferable implementation of the invention, but the present invention is not limited to the embodiment above, it is ripe
Various equivalent deformation or replacement can also be made on the premise of without prejudice to spirit of the invention by knowing those skilled in the art, this
Equivalent deformation or replacement are all included in the scope defined by the claims of the present application a bit.
Claims (10)
1. a kind of image magnification method based on self-similarity between scale characterized by comprising
Training data set is obtained from using input picture；
The probability density function of the self-training data acquisition system is obtained based on gauss hybrid models；
Adaptivity prior image is obtained according to the high-resolution low-frequency image of the probability density function and the input picture；
It is generated and is amplified according to the low resolution fidelity image of the adaptivity prior image and the adaptivity prior image
Output image.
2. the image magnification method according to claim 1 based on self-similarity between scale, it is characterised in that: described to be directed to
Input picture is obtained from training data set, specifically includes:
It is described to being obtained after up-sampling s times after s times of down-sampling of input picture bicubic of result, then to the result bicubic
The low-frequency image of input picture, wherein s is positive number；
All image sheets of the low-frequency image of the input picture and the input picture and corresponding connection are extracted, self-training is obtained
Data acquisition system.
3. the image magnification method according to claim 2 based on self-similarity between scale, it is characterised in that: the probability
The expression formula of density function is as follows:
Wherein, p indicates that probability, T indicate transposition operation, yjWith y 'jIt indicates respectively from the input picture and the input picture
J-th of the image sheet extracted in low-frequency image, K are the number for introducing Gaussian function, πk、μk、∑kRespectively k-th of Gaussian function
Weight coefficient, mean vector and covariance matrix,Indicate k-th of Gaussian function.
4. the image magnification method according to claim 3 based on self-similarity between scale, it is characterised in that: the basis
The high-resolution low-frequency image of the probability density function and the input picture obtains the step for adaptivity prior image,
It specifically includes:
By input picture bicubic to up-sampling s times after obtain the high-resolution low-frequency image of the input picture；
All image sheets according to the high-resolution low-frequency image of the input picture are corresponding in the probability density function
Probability density obtains adaptivity elder generation after handling all image sheets of the high-resolution low-frequency image of the input picture
Test image.
5. the image magnification method according to claim 1 based on self-similarity between scale, it is characterised in that: the basis
The low resolution fidelity image of the adaptivity prior image and the adaptivity prior image generates the output figure of amplification
Picture specifically includes:
Adaptivity prior image is done into downsampled and gaussian filtering process, obtains the low resolution of adaptivity prior image
Image；
According to the input picture and the low-resolution image of adaptivity prior image, the low of adaptivity prior image is obtained
Resolution ratio fidelity image；
It is raw according to the low resolution fidelity image, adaptivity prior image and proportionality coefficient of the adaptivity prior image
At the output image of amplification.
6. the image magnification method according to claim 5 based on self-similarity between scale, it is characterised in that: the amplification
Output image calculation formula are as follows:
Wherein,For fidelity term, Y indicates that input picture matrix, X indicate adaptivity prior image, and A expression degrades
Matrix, λ are proportionality coefficient,For adaptivity prior image expression formula form, N is to take out in X
The image sheet quantity taken, Pi、AndRespectively indicate the probability of i-th of image sheet in X, the mean vector of Gaussian function and association
Variance matrix.
7. the image magnification method according to claim 1 based on self-similarity between scale, it is characterised in that: the input
Image is obtained by following steps:
Color image is transformed into yuv space from rgb space；
Using Y channel image as the input picture.
This method is further comprising the steps of:
U the and V channel image of the color image is amplified to target size according to bicubic sampling method；
Using Y channel image as input picture, and amplified according to the above-mentioned image magnification method based on self-similarity between scale
To the target size；
U, V channel image for being amplified to target size and the channel Y output image are merged, and are converted into the output of rgb space image,
As output image.
8. a kind of image enhancement system based on self-similarity between scale characterized by comprising
Self-training data acquisition system obtains module, for being obtained from training data set using input picture；
Probability density function computing module, for obtaining the probability density of the self-training data acquisition system based on gauss hybrid models
Function；
Adaptivity prior image determining module, for the high-resolution according to the probability density function and the input picture
Low-frequency image obtains adaptivity prior image；
Image generation module is exported, for according to low point of the adaptivity prior image and the adaptivity prior image
Resolution fidelity image generates the output image of amplification.
9. a kind of image enhancement system based on self-similarity between scale, it is characterised in that: include:
At least one processor；
At least one processor, for storing at least one program；
When at least one described program is executed by least one described processor, so that at least one described processor is realized as weighed
Benefit requires the image magnification method based on self-similarity between scale described in any one of 1-7.
10. a kind of storage medium, wherein being stored with the executable instruction of processor, it is characterised in that: the processor is executable
Instruction when executed by the processor for executing as of any of claims 1-7 based on self-similarity between scale
Image magnification method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910433432.4A CN110223229A (en) | 2019-05-23 | 2019-05-23 | Based on the image magnification method of self-similarity, system and storage medium between scale |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910433432.4A CN110223229A (en) | 2019-05-23 | 2019-05-23 | Based on the image magnification method of self-similarity, system and storage medium between scale |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110223229A true CN110223229A (en) | 2019-09-10 |
Family
ID=67818241
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910433432.4A Pending CN110223229A (en) | 2019-05-23 | 2019-05-23 | Based on the image magnification method of self-similarity, system and storage medium between scale |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110223229A (en) |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105354804A (en) * | 2015-10-23 | 2016-02-24 | 广州高清视信数码科技股份有限公司 | Maximization self-similarity based image super-resolution reconstruction method |
-
2019
- 2019-05-23 CN CN201910433432.4A patent/CN110223229A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105354804A (en) * | 2015-10-23 | 2016-02-24 | 广州高清视信数码科技股份有限公司 | Maximization self-similarity based image super-resolution reconstruction method |
Non-Patent Citations (1)
| Title |
|---|
| JIANHONG LI等: "Maximizing Nonlocal Self-Similarity Prior for Single Image Super-Resolution", 《MATHEMATICAL PROBLEMS IN ENGINEERING》 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Mahapatra et al. | Image super-resolution using progressive generative adversarial networks for medical image analysis | |
| CN105046672B (en) | A kind of image super-resolution rebuilding method | |
| Zhao et al. | Data-driven enhancement of blurry retinal images via generative adversarial networks | |
| Bao et al. | Undersampled MR image reconstruction using an enhanced recursive residual network | |
| Li et al. | A multiscale double-branch residual attention network for anatomical–functional medical image fusion | |
| CN105631807A (en) | Single-frame image super resolution reconstruction method based on sparse domain selection | |
| CN107133933B (en) | Mammary X-ray image enhancement method based on convolutional neural network | |
| CN104376568A (en) | Method for processing DICOM (digital imaging and communications in medicine) medical images on basis of formats | |
| Aghabiglou et al. | Projection-Based cascaded U-Net model for MR image reconstruction | |
| Yang et al. | Generative Adversarial Networks (GAN) Powered Fast Magnetic Resonance Imaging--Mini Review, Comparison and Perspectives | |
| Chen et al. | A sparse representation and dictionary learning based algorithm for image restoration in the presence of Rician noise | |
| Wang et al. | A two-step automated quality assessment for liver MR images based on convolutional neural network | |
| CN101799919A (en) | Front face image super-resolution rebuilding method based on PCA alignment | |
| Wang et al. | MHAN: Multi-Stage Hybrid Attention Network for MRI reconstruction and super-resolution | |
| CN104392452B (en) | A kind of DICOM medical image processing method based on application | |
| Martins et al. | A multi-object statistical atlas adaptive for deformable registration errors in anomalous medical image segmentation | |
| Yin et al. | Super resolution reconstruction of CT images based on multi-scale attention mechanism | |
| CN110223229A (en) | Based on the image magnification method of self-similarity, system and storage medium between scale | |
| Tan et al. | An endoscope image enhancement algorithm based on image decomposition | |
| CN116071239A (en) | CT image super-resolution method and device based on mixed attention model | |
| Zhang et al. | SS-SSAN: a self-supervised subspace attentional network for multi-modal medical image fusion | |
| CN115131361A (en) | Training of target segmentation model, focus segmentation method and device | |
| CN110895790B (en) | Scene image super-resolution method based on posterior degradation information estimation | |
| Hu et al. | Single image super resolution of 3D MRI using local regression and intermodality priors | |
| Wan et al. | Multi-focus color image fusion based on quaternion multi-scale singular value decomposition |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190910 |