CN103310425A - Large-scale image restoration achieving method based on image gradient prior model - Google Patents

Abstract

The invention relates to a large-scale image restoration achieving method based on an image gradient prior model. The method includes: reading-in information of an image to be restored, building an image noise model, a non-local mean value spatial domain probability model and an image gradient prior probability model according to the information of the image to be restored, building a Bayesian posterior probability model and calculating a posterior probability estimation value, confirming a value of the current pixel according to the posterior probability estimation value and saving the value of the current pixel, and then restoring the next pixel until restoration of the whole image is finished. The large-scale image restoration achieving method based on the image gradient prior model is used for solving the problem that texture boundary consistency is poor in the image restoration process in the prior art, making full use of redundancy of video images in a spatial domain, combining cooperative constraints of the non-local mean value spatial domain probability model and the image gradient prior probability model, and effectively restoring the information of images with large-scale damages. The large-scale image restoration achieving method based on the image gradient prior model is convenient to use and has a wider application range.

Description

Realize the method that large scale image is repaired based on the image gradient prior model

Technical field

The present invention relates to image processing field, relate in particular to large scale image and repair the field, specifically refer to a kind of method that realizes the large scale image reparation based on the image gradient prior model.

Background technology

Image repair is the processing procedure that breakage image is repaired.Along with the development of digital technology, increasing image document employing digital form gathers, Storage and Processing.In the digital picture use procedure, often can run into the damaged situation of digital picture local message, after digitizing, often there is cut such as old cinefilm in the image, losing of some data appearred in patch equivalent damages etc. in digital picture.When the picture editting, when the literal in removal or the moving images and object, can cause the Image blank district, after the personage of script was removed in image, the view data that stays bulk was blank.The disappearance of these information has reduced picture quality greatly.Image repair method based on Computer Processing is divided into usually: based on the method for spatial model with based on the method for sample image.The former is taking full advantage of image correlativity spatially.The latter then is the damaged information of repairing present image by the information of sample image.Existing image repair method generally all is based on space block of pixels matching strategy, and under the texture complicated situation, restoring area often can not obtain preferably visual effect.

Find through the literature search to prior art, the algorithm that adopts based on the image repair method of spatial model and sample image mainly is to utilize the redundant information of view data on space and sample space, such as Criminisi, Antonio, Patrick Perez, and Kentaro Toyama is at " IEEE Transactions on Image Processing " (the IEEE image is processed periodical) the 13rd volume, the 9th phase, the image repair method that calculates based on the priority of block of pixels is proposed in the 1200th page to 1212 pages " Region filling and object removal by exemplar-based image inpainting " (based on object removal and the area filling method of space analysis model) literary compositions of delivering.Hays, James, and Alexei A.Efros is at " ACM Transactions on Graphics " (ACM computer graphics periodical) the 26th volume, the 3rd phase, propose based on the characteristic matching of sample space and the image repair method of study in the 4th page of " Scene completion using millions of photographs " that delivers (scene based on the multisample space is recovered) literary composition.Said method is to come the repair deficiency image information by the redundant information in local image space or sample image space.Yet the shortcoming of these method maximums is to fail to consider the using texture homogeneity of restoring area and known region, and the blocks and optimal matching blocks that obtains in piece coupling priority computation process often can't obtain preferably visual effect, i.e. the continuity on border.To the restorative procedure based on the sample space, how obtaining suitable coupling sample still is problem demanding prompt solution.

Summary of the invention

The objective of the invention is to overcome the shortcoming of above-mentioned prior art, provide a kind of can realize overcoming texture in the image repair process and boundary constraint consistency problem, to large scale breakage image information have efficient recovery, easy to use, have a broader applications scope realize the method that large scale image is repaired based on the image gradient prior model.

To achieve these goals, the method based on the reparation of image gradient prior model realization large scale image of the present invention has following formation:

Should realize the method that large scale image is repaired based on the image gradient prior model, its principal feature is that described method may further comprise the steps:

(1) reads in image information to be repaired;

(2) set up image noise model and computed image noise probability distribution value;

(3) set up the spatial domain probability model of non-local mean and calculate the spatial domain probability distribution value of non-local mean;

(4) set up image gradient prior probability model and computed image gradient prior probability distribution value;

(5) set up Bayes posterior probability model and calculate the posterior probability estimation value;

(6) also preserve according to the value of described posterior probability estimation value affirmation current pixel point.

Preferably, between described step (1) and the step (2), further comprising the steps of:

(11) judge whether described image to be repaired finishes dealing with, if not, then continue step (2), if so, then finish to withdraw from.

More preferably, between described step (5) and the step (6), further comprising the steps of:

(51) judge whether described posterior probability estimation value is local maximum, if so, then continues step (6), if not, then continues step (11).

Preferably, described image noise model and the computed image noise probability distribution value set up is specially:

According to described image information to be repaired, set up image noise model and computed image noise probability distribution value, described image noise model formula is as follows:

P(Δ _n|s)∝N(Δ _n|μ,σ)；

Wherein, P (Δ _n| s) be described image noise model, σ is the variance of Gaussian distribution, and μ is the average of Gaussian distribution, Δ _nNeighbor for current pixel point.

More preferably, the variance yields of described Gaussian distribution is 0.5, and the average of described Gaussian distribution is 0.7.

Preferably, the described spatial domain probability distribution value of setting up the spatial domain probability model of non-local mean and calculating non-local mean is specially:

According to described image information to be repaired, to set up the spatial domain probability model of non-local mean and calculate the spatial domain probability distribution value of non-local mean, the spatial domain probability model formula of described non-local mean is as follows:

$P\left({\mathrm{\θ}}_n\right|s)\∝\exp \{-\frac{{\left|\right|s-{\mathrm{\θ}}_n\left(\mathrm{\Ω}\right)\left|\right|}_2^2}{h^2}\};$

Wherein, P (θ _n| s) be the spatial domain probability model, Ω is the neighbor pixel set, and s is the current pixel dot information, and h is decay control parameter.

More preferably, described decay control parameter value is 1.6.

Preferably, described image gradient prior probability model and the computed image gradient prior probability distribution value set up is specially:

According to described image information to be repaired, set up image gradient prior probability model and computed image gradient prior probability distribution value, described image gradient prior probability model formation is as follows:

$P\left(s\right)\∝\frac{{\mathrm{\γ\λ}}^{\left(\frac{1}{\mathrm{\γ}}\right)}}{2\mathrm{\Γ}\left(\frac{1}{\mathrm{\γ}}\right)}\exp (-\mathrm{\λ}{\left|s\right|}^{\mathrm{\γ}});$

Wherein, P (s) is the prior probability model, and Γ is that gamma distributes, and λ and γ are the shape parameters that gamma distributes.

More preferably, the shape parameters λ value of described gamma distribution is that 4, γ value is 0.8.

Preferably, describedly set up the Bayes posterior probability model and calculate the posterior probability estimation value, be specially:

Based on described image noise model, spatial domain probability model and prior probability model, to set up the Bayes posterior probability model and calculate the posterior probability estimation value, described Bayes posterior probability model formation is as follows:

P(s|Δ _n,θ _n)∝P(Δ _n|s)·P(θ _n|s)·P(s)；

Wherein, P (s| Δ _n, θ _n) be the Bayes posterior probability model, P (Δ _n| s) be image noise model, P (θ _n| s) be the spatial domain probability model of non-local mean, P (s) is the image gradient prior model, and s is the current pixel dot information.

Adopt the method based on the reparation of image gradient prior model realization large scale image in this invention, had following beneficial effect:

(1) utilize image noise model, based on the image space restricted model of non-local mean and based on the prior probability model of image gradient, set up the Bayes posterior probability Unified frame of image repair, overcome the poor problem of Texture Boundaries consistance in the image repair process.

(2) the present invention takes full advantage of the redundancy of video image on the spatial domain, by realized the image repair of large scale based on Bayes's Unified frame; Because the present invention is summed up as image repair the constraint of uniting of spatial domain model and prior model, thereby has overcome the insurmountable problem of traditional restorative procedure, has realized the efficient recovery that has to large scale breakage image information, has widely range of application.

Description of drawings

Fig. 1 is the method that realizes the large scale image reparation based on the image gradient prior model of the present invention

Embodiment

In order more clearly to describe technology contents of the present invention, conduct further description below in conjunction with specific embodiment.

The present invention is directed to above-mentioned the deficiencies in the prior art, proposed a kind of image repair method based on the image gradient prior model, the method adopts Markov image gradient prior model, in order to guarantee that image is in borderline consistance; And set up unified Bayesian inference framework, utilize non-local mean model in the framework to come the space constraint of Description Image, adopt at last the greatest hope algorithm to finish whole posterior probability estimation process.

Image repair principle of the present invention is: by setting up unified Bayes posterior probability framework, introduce the distribution of spatial domain probability model, Gaussian image noise model and image gradient prior model based on non-local mean, effectively solved the reparation problem under the damaged condition of image large scale; Under Unified frame, realize the damaged information reparation of image, and adopted spatial domain constraint and prior model to overcome the inconsistent problem of restoring area texture; The speed of the estimation by condition iterative optimization techniques Effective Raise posterior model.Therefore, by having set up a kind of unified Bayesian inference model framework, can realize the image information reparation of effective large scale.

Based on above principle, method of the present invention may further comprise the steps: image information is read in; To the Spatial Probability model profile of the non-local mean that calculates current pixel point, picture noise probability distribution and gradient prior probability distribution; Calculate the posterior probability estimation value under posteriority Bayes Unified frame, if the posterior probability estimation value is local maximum, then current pixel point is obtained by current neighbor information calculations; Otherwise, then need to continue iterative computation.

As shown in Figure 1, be the process flow diagram of the reparation image of the present embodiment.

Be that the image of 512 * 512 pixels is done image repair to the image size in the present embodiment, flow process comprises the steps: as shown in Figure 1

(1) reads in current image information to be repaired.

(2) pixel in the upper left corner is set up rectangular coordinate system as initial point in the image, to current pixel point computed image noise probability distribution value, i.e. and pixel noise probability model Distribution Value, the formula of concrete image noise model is as follows:

P(Δ _n|s)∝N(Δ _n|μ,σ)；

In following formula, N is the Gaussian distribution probability model, and parameter σ and μ are variance and the averages of Gaussian distribution.In the present invention, our value is 0.5 and 0.7 in the present invention.Δ _nNeighbor for current pixel.

(3) to the spatial domain constraint probability of current pixel point calculating non-local mean, concrete spatial domain probability model formula is as follows

$P\left({\mathrm{\θ}}_n\right|s)\∝\exp \{-\frac{{\left|\right|s-{\mathrm{\θ}}_n\left(\mathrm{\Ω}\right)\left|\right|}_2^2}{h^2}\};$

Spatial domain probability P (θ _n| s) represented the similarity degree of current pixel and neighbor.In following formula, Ω represents the neighbor pixel set, and s represents current pixel, and h is decay control parameter, and the h value is 1.6 in the present embodiment.

(4) current pixel point being calculated Markov gradient prior probability distribution is the image gradient prior probability, and concrete prior probability model is as follows:

$P\left(s\right)\∝\frac{{\mathrm{\γ\λ}}^{\left(\frac{1}{\mathrm{\γ}}\right)}}{2\mathrm{\Γ}\left(\frac{1}{\mathrm{\γ}}\right)}\exp (-\mathrm{\λ}{\left|s\right|}^{\mathrm{\γ}});$

In following formula, Γ represents that gamma distributes, and λ, γ are the shape parameters that gamma distributes.In the present embodiment, the λ value is that 4, γ value is 0.8.

(5) (2) step, (3) step and (4) step being obtained the following Bayes's posteriority of estimated value substitution expression formula calculates, be about in image noise model description, the description of non-local mean Space category model and the image gradient prior probability model substitution Bayes posterior probability framework, calculating by posterior probability, thereby obtain the estimation to current pixel point, if the posterior probability estimation value is local maximum, then current pixel point is obtained by current neighbor information calculations; Otherwise, then need to continue iterative computation.In present image, calculate by constantly updating posteriority result and iteration optimization, finally obtain the Pixel Information of damaged area.

The process of describing framework of setting up Bayesian inference model posterior probability is as follows:

P(s|Δ _n,θ _n)∝P(Δ _n,θ _n|s)/P(Δ _n,θ _n)

∝P(Δ _n,θ _n|s)P(s)

∝P(Δ _n|s)·P(θ _n|s)·P(s)；

In the description of whole Bayes posterior probability model, P (Δ _n| s) be image noise model, P (θ _n| s) be the spatial domain probability model, P (s) is the image gradient prior model.S is damaged area pixel information, finish a current pixel point after, do the pixel of next damaged area, until finish the reparation of whole image damaged area information.

Adopt the method based on the reparation of image gradient prior model realization large scale image in this invention, had following beneficial effect:

(1) utilize image noise model, based on the image space restricted model of non-local mean and based on the prior probability model of image gradient, set up the Bayes posterior probability Unified frame of image repair, overcome the poor problem of Texture Boundaries consistance in the image repair process.

(2) the present invention takes full advantage of the redundancy of video image on the spatial domain, by realized the image repair of large scale based on Bayes's Unified frame; Because the present invention is summed up as image repair the constraint of uniting of spatial domain model and prior model, thereby has overcome the insurmountable problem of traditional restorative procedure, has realized the efficient recovery that has to large scale breakage image information, has widely range of application.

In this instructions, the present invention is described with reference to its specific embodiment.But, still can make various modifications and conversion obviously and not deviate from the spirit and scope of the present invention.Therefore, instructions and accompanying drawing are regarded in an illustrative, rather than a restrictive.

Claims (10)

1. a method that realizes the large scale image reparation based on the image gradient prior model is characterized in that, described method may further comprise the steps:

(1) reads in image information to be repaired;

(2) set up image noise model and computed image noise probability distribution value;

(3) set up the spatial domain probability model of non-local mean and calculate the spatial domain probability distribution value of non-local mean;

(4) set up image gradient prior probability model and computed image gradient prior probability distribution value;

(5) set up Bayes posterior probability model and calculate the posterior probability estimation value;

(6) also preserve according to the value of described posterior probability estimation value affirmation current pixel point.

2. the method that realizes the large scale image reparation based on the image gradient prior model according to claim 1 is characterized in that, and is between described step (1) and the step (2), further comprising the steps of:

(11) judge whether described image to be repaired finishes dealing with, if not, then continue step (2), if so, then finish to withdraw from.

3. the method that realizes the large scale image reparation based on the image gradient prior model according to claim 2 is characterized in that, and is between described step (5) and the step (6), further comprising the steps of:

(51) judge whether described posterior probability estimation value is local maximum, if so, then continues step (6), if not, then continues step (11).

4. the method that realizes the large scale image reparation based on the image gradient prior model according to claim 1 is characterized in that, described image noise model and the computed image noise probability distribution value set up is specially:

According to described image information to be repaired, set up image noise model and computed image noise probability distribution value, described image noise model formula is as follows:

P(Δ _n|s)∝N(Δ _n|μ,σ)；

Wherein, P (Δ _n| s) be described image noise model, σ is the variance of Gaussian distribution, and μ is the average of Gaussian distribution, Δ _nNeighbor for current pixel point.

5. the method that realizes the large scale image reparation based on the image gradient prior model according to claim 4 is characterized in that, the variance yields of described Gaussian distribution is 0.5, and the average of described Gaussian distribution is 0.7.

6. the method that realizes the large scale image reparation based on the image gradient prior model according to claim 1 is characterized in that, the described spatial domain probability distribution value of setting up the spatial domain probability model of non-local mean and calculating non-local mean is specially:

According to described image information to be repaired, to set up the spatial domain probability model of non-local mean and calculate the spatial domain probability distribution value of non-local mean, the spatial domain probability model formula of described non-local mean is as follows:

$P\left({\mathrm{\θ}}_n\right|s)\∝\exp \{-\frac{{\left|\right|s-{\mathrm{\θ}}_n\left(\mathrm{\Ω}\right)\left|\right|}_2^2}{h^2}\};$

Wherein, P (θ _n| s) be the spatial domain probability model, Ω is the neighbor pixel set, and s is the current pixel dot information, and h is decay control parameter.

7. the method that realizes the large scale image reparation based on the image gradient prior model according to claim 6 is characterized in that, described decay control parameter value is 1.6.

8. the method that realizes the large scale image reparation based on the image gradient prior model according to claim 1 is characterized in that, described image gradient prior probability model and the computed image gradient prior probability distribution value set up is specially:

According to described image information to be repaired, set up image gradient prior probability model and computed image gradient prior probability distribution value, described image gradient prior probability model formation is as follows:

$P\left(s\right)\∝\frac{{\mathrm{\γ\λ}}^{\left(\frac{1}{\mathrm{\γ}}\right)}}{2\mathrm{\Γ}\left(\frac{1}{\mathrm{\γ}}\right)}\exp (-\mathrm{\λ}{\left|s\right|}^{\mathrm{\γ}});$

Wherein, P (s) is the prior probability model, and Γ is that gamma distributes, and λ and γ are the shape parameters that gamma distributes.

9. the method that realizes the large scale image reparation based on the image gradient prior model according to claim 8 is characterized in that, the shape parameters λ value that described gamma distributes is that 4, γ value is 0.8.

10. according to claim 1ly realize the method that large scale image is repaired based on the image gradient prior model, it is characterized in that, describedly set up the Bayes posterior probability model and calculate the posterior probability estimation value, be specially:

Based on described image noise model, spatial domain probability model and prior probability model, to set up the Bayes posterior probability model and calculate the posterior probability estimation value, described Bayes posterior probability model formation is as follows:

P(s|Δ _n,θ _n)∝P(Δ _n|s)·P(θ _n|s)·P(s)；

Wherein, P (s| Δ _n, θ _n) be the Bayes posterior probability model, P (Δ _n| s) be image noise model, P (θ _n| s) be the spatial domain probability model of non-local mean, P (s) is the image gradient prior model, and s is the current pixel dot information.

Images