CN103955906A - Criminisi image restoration method based on bat algorithm - Google Patents

Abstract

A Criminisi image restoration method based on a bat algorithm comprises the steps of calculating the priority of each pixel point of to-be-restored region edges of to-be-restored images, selecting a pixel point with maximum priority as a pixel point with restoration priority, performing searching and filling on best matching blocks in perfect regions of the to-be-restored images, wherein searching (as mentioned) is performed by adopting the bat algorithm according to a matching principle; updating the to-be-restored region edges, returning the edges and performing repeated circulating operation till the restoration of to-be-restored regions is finished and obtaining an image restoration result. The Criminisi image restoration method is wide in image restoration range, improves the restoration speed, reduces the time consumption and meets the visual demands of people on the premise that the restoration quality of the to-be-restored images having different emphasis areas is ensured. Therefore, the Criminisi image restoration method has important practical significance.

Description

Criminisi image repair method based on bat algorithm

Technical field

The invention belongs to image restoration field, particularly relate to Criminisi image repair method.

Background technology

In recent years, image repair is a study hotspot of computer vision, is widely used in the various fields such as the removing of reparation, redundancy object, compression of images, video display special technology making of image and video.Its essence is exactly the information of disappearance of recovering by the information existing in image to be repaired, makes image repair whole structure meet people's visual demand.

For image repair, there are at present two large technology: the image repair technology based on partial differential equation and based on the synthetic image repair technology of texture.The size of the selective basis restoring area of two kinds of restorative procedures is determined, the former core concept is the hot-fluid partial differential equation based in physics, thereby make the peripheral information of damaged area in image be diffused into the inside of damaged area, reach the object of image repair, the method of representative has BSCB model algorithm and TV model algorithm etc., is applicable to small size image repair; The latter's core concept is the small sample based on texture, centered by its pixel on impaired edge, by block of pixels existing in image, mates, and comes the impaired region of filling information, reaches the object of image repair, applicable picture of large image scale reparation.

Criminisi image repair algorithm is the representative of the image repair method based on texture, by people such as Criminisi, in 2004, is proposed, and its repair process is: right of priority calculating, optimum matching block search and filling, renewal degree of confidence.But the right of priority priority (p) in Criminisi image repair algorithm calculates and search and the filling of best matching blocks exist self-defect, that is: degree of confidence can make along with the increase of repairing number of times the information reduction of its Central Plains figure, the appearance that has the difference of the order of magnitude affects right of priority, in data item, there will be vertical phenomenon, causing priority is zero, the size of degree of confidence will be nonsensical to right of priority and more than one of optimum matching module, and system can be random chooses etc.Therefore nearly ten years, there is the improvement of the right of priority of many different emphasis and the search of best matching blocks and filling: in improvement image repair algorithm > > mono-literary composition at the interim < < of sensing technology journal the 25th volume the 3rd in 2012 based on variable-size template, keeping on the basis that Criminisi rudimentary algorithm framework is constant, designed piece matcher based on variable-size template so that template is searched for more accurate and flexible and reliability update mode is revised, and combining local searching is to improve the combination property of algorithm, rapid image mono-kind of the 2011 annual data collections < < interim with processing the 26th volume the 6th based on sample block is repaired in algorithm > > mono-literary composition, introduce new metric function and upgrade degree of confidence, make the calculating of priority more accurate, the screening strategy again of to be matched reduced and select the randomness of best matching blocks, repair sample block neighborhood and detect and avoided finding in global scope broken edges.The method has obtained good repairing effect, has improved the efficiency of algorithm simultaneously; In computer utility in 2012 and interim < < improved Criminisi image repair algorithm > > mono-literary composition of software the 29th volume the 9th, introduce curvature and decide the fill order of object block and the selection of best matching blocks, and to improve right of priority be every weighted sum, by changing weights, can obtain better repairing effect, avoided, due to the degree of confidence wrong fill order that decay brings rapidly, having obtained gratifying repairing effect simultaneously.To sum up, the emphasis of improving algorithm is the reliability of right of priority calculating and the searching method of optimum matching template, take repairing quality as center of gravity.Guaranteeing that right of priority determines to improve its computing formula under a constant prerequisite, promote its confidence level herein; Bat algorithm can merge global search and Local Search well efficiently, and its search speed is fast, and precision is high, is incorporated into optimum matching block search and filling in Criminisi algorithm.This patent algorithm guarantees to improve remediation efficiency under repairing quality prerequisite.

Summary of the invention

The object of the invention is to, guarantee, on the basis of repairing quality, to improve the speed of repairing.

Technical scheme provided by the invention is a kind of Criminisi image repair method based on bat algorithm, comprise the following steps, step 1, the right of priority of each pixel of edge, area to be repaired of calculating image to be repaired is as follows, the pixel of choosing right of priority maximum is the preferential pixel of repairing

priority(p)＝C(p)×D(p)

Wherein, the right of priority that priority (p) is edge pixel point p, C (p) is degree of confidence, D (p) is data item;

Step 2, for the preferential pixel of repairing of step 1 gained, carries out search and the filling of best matching blocks in the intact region of image to be repaired, search adopts bat algorithm to carry out according to SSD matching principle, and described SSD matching principle is as follows,

${\mathrm{\&psi;}}_{\hat{q}}=\arg \underset{{\mathrm{\&psi;}}_q\&Element;\mathrm{\&phi;}}{\min }d({\mathrm{\&psi;}}_{\hat{p}},{\mathrm{\&psi;}}_q)$

Wherein, represent the preferential corresponding object block to be repaired of pixel of repairing of step 1 gained, ψ _qrepresent the sample block in intact region φ, represent piece ψ _qthe quadratic sum of known pixels point color difference, for object block to be repaired best matching blocks;

Step 3, upgrades edge, area to be repaired, returns to step 1 and carries out repetitive cycling operation, until has repaired area to be repaired, obtains image repair result.

And, in step 1,

$C\left(p\right)=\frac{{\mathrm{\&Sigma;}}_{{i\&Element;\mathrm{\&psi;}}_p}\&cap;{\mathrm{\&phi;}}^{C\left(i\right)}}{\left|{\mathrm{\&psi;}}_p\right|}$

Wherein, variable C (i) is defined as C (i)=0, φ is intact region, and Ω is area to be repaired; | ψ _p| represent ψ _parea, ψ _prepresent the corresponding object block to be repaired of edge pixel point p;

$D\left(p\right)=\frac{|\&dtri;I\stackrel{\&perp;}{p}\&CenterDot;n_p|}{\mathrm{\&alpha;}}$

Wherein, n _pit is the border in area to be repaired at edge pixel point p unit outside normal vector; the direction and intensity that waits irradiation line that represents edge pixel point p place, its expression formula is i _x, I _yrepresent respectively the partial differential of edge pixel p in x, y direction; α is normalized parameter.

The wide range of the applicable image repair of the present invention,, consumes during reduction guaranteeing, under the prerequisite of repairing quality, to improve reparation speed for the image to be repaired of different emphasis, meets people's visual demand.Therefore, the present invention uses and has important practical significance the enforcement of Criminisi image repair algorithm.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the embodiment of the present invention.

Fig. 2 is the pixel schematic diagram at the edge, area to be repaired of the embodiment of the present invention.

Fig. 3 is the searching route schematic diagram of the bat algorithm of the embodiment of the present invention.

Embodiment

First the present invention carries out the calculating of right of priority, right of priority maximum be the piece of first repairing, then bat algorithm is incorporated into selection and the filling of optimal Template, final updating degree of confidence, that goes round and begins again carries out aforesaid operations until has repaired area to be repaired.Technical solution of the present invention can adopt computer software mode to support automatic operational scheme.Below in conjunction with drawings and Examples, describe technical solution of the present invention in detail.

Referring to Fig. 1, the flow process of embodiment is as follows:

Step 1, calculates the right of priority priority (p) of each pixel at the edge, area to be repaired of image to be repaired, determines the preferential pixel of repairing.

During concrete enforcement, generally can adopt a kind of specific color to carry out manually mark in advance to the area to be repaired of image to be repaired, can select voluntarily area to be repaired to carry out handmarking by user.In advance mark after the image to be repaired input of area to be repaired, the flow process that adopts the embodiment of the present invention to provide is processed.Image to be repaired in the embodiment of the present invention adopts white.For example, in certain personage's photo, in background, there is supernumerary group, adopt white to cover supernumerary group region.

Embodiment, according to formula priority (p)=C (p) * D (p), calculates the right of priority of each pixel at edge, area to be repaired, chooses the pixel of right of priority maximum and preferentially repairs.

Referring to Fig. 2, wherein:

$C\left(p\right)=\frac{{\mathrm{\&Sigma;}}_{{i\&Element;\mathrm{\&psi;}}_p}\&cap;{\mathrm{\&phi;}}^{C\left(i\right)}}{\left|{\mathrm{\&psi;}}_p\right|}$

Wherein, variable C (i) is defined as C (i)=0, the region that φ represents is unlabelled region, i.e. intact region; The region that Ω represents is the region of mark, i.e. area to be repaired.At the pixel i in intact region, variable C (i) assignment is 1, the pixel i in area to be repaired, and variable C (i) assignment is 0.| ψ _p| represent ψ _parea, ψ _prepresent the corresponding object block to be repaired of edge pixel point p.

$D\left(p\right)=\frac{|\&dtri;I\stackrel{\&perp;}{p}\&CenterDot;n_p|}{\mathrm{\&alpha;}}$

Wherein, n _pit is the border in area to be repaired at edge pixel point p unit outside normal vector; the direction and intensity that waits irradiation line that represents edge pixel point p place, its expression formula is i _x, I _yrepresent respectively the partial differential of edge pixel p in x, y direction; α is normalized parameter, and in gray-scale map, value is 255.

Wherein, the right of priority that priority (p) is edge pixel point p, C (p) is degree of confidence, the larger right of priority of numerical value is higher, has reflected that the information of the former graph region containing is many, should give preferential repairing; D (p) is data item, and the larger right of priority of numerical value is higher, has reflected that evolution surface linear structural strength is high, should give preferential repairing.

Step 2, this patent to this step, is determined bat algorithm application after preferential repairing pixel point, is carried out search and the filling of best matching blocks at intact region φ.Can adopt the way of search of bat algorithm to carry out search and the filling of optimum matching template according to SSD matching principle, wherein SSD matching principle be the condition that bat is distinguished object and barrier.

The SSD matching principle of embodiment is shown below:

${\mathrm{\&psi;}}_{\hat{q}}=\arg \underset{{\mathrm{\&psi;}}_q\&Element;\mathrm{\&phi;}}{\min }d({\mathrm{\&psi;}}_{\hat{p}},{\mathrm{\&psi;}}_q)$

Wherein, represent object block to be repaired, corresponding with the preferential pixel of repairing of step 1 gained, be generally all pixels of getting centered by preferential repairing pixel point in default size windows, for example 3 * 3,9 * 9 window; ψ _qthe sample block that represents intact region; the quadratic sum that represents the known pixels point color difference of two pieces; SSD matching principle represents when object block to be repaired with sample block ψ in intact region _qknown pixels point color difference quadratic sum hour, this sample block is best matching blocks and then according to best matching blocks to object block to be repaired fill.

Bat algorithm is incorporated in the search and filling of optimum matching template, while getting minimum value with bat algorithm search SSD matching principle, obtains corresponding block of pixels as best matching blocks, fill it into object block accordingly to be repaired in area to be repaired.

Bat algorithm can well merge global search and Local Search efficiently.And this algorithm has good adaptability and robustness, can, guaranteeing, under the prerequisite of repairing quality, to promote the speed of repairing, during reduction, consume.

Step 3, upgrades edge, area to be repaired.The piece of repairing is all on edge, area to be repaired, and the reparation that therefore often completes a texture block is upgraded with regard to edge.After renewal, from step 1, carry out repetitive cycling operation, constantly circulation is until area to be repaired Ω has repaired, and image repair completes, and obtains the Criminisi image repair result based on bat algorithm.During concrete enforcement, can first judge whether area to be repaired Ω has repaired, be process ends, otherwise upgrade edge, area to be repaired, return to step 1, each pixel at the edge, area to be repaired based on remaining, selects new preferential repairing pixel point to process equally.

Bat algorithm can be with reference to existing techniques in realizing, and for the purpose of those skilled in the art's implementation step 2, the embodiment of the present invention provides corresponding specific implementation to be described as follows.

Bat algorithmic formula is based on following three principles:

(1) bat adopts specific mode to differentiate object and barrier, and utilizes echolocation mode perceived distance;

(2) X at an arbitrary position _iplace, bat is with speed V _iflight arbitrarily, to determine big or small frequency f _minwavelength X, loudness A with variable size ₀look for object, and according to object and the distance of self far and near adjusting wavelength or frequency automatically, and near object time, adjust exomonental frequency r ∈ [0,1].

(3) variation of acquiescence loudness is from maximal value A ₀to minimum value A _min, A ₀be necessary on the occasion of.

Based on mentioned above principle, the step of bat algorithm is as follows:

Step 1: parameter initialization: objective function f (X), wherein dimension set X=(x ₁, x ₂, x ₃... x _d) ^t, d is dimension, x ₁, x ₂, x ₃... x _dfor variable number, the individuality of bat colony adopts i sign, and the individuality of initialization bat colony is at X _ithe speed V at place _iand define X _ithe pulsed frequency f at place _i, initialization pulse speed r _iwith loudness A _i; In the present invention, objective function f (X) adopts SSD matching principle.

Step 2: produce new solution renewal speed and position by adjusting frequency;

Step 3: according to rand1 > r _ijudge, select a solution from the set of optimum solution, in optimum solution, form a local solution around, wherein rand1 represents to produce a random number.

Step 4: determine a new solution by uncertain flight;

Step 5: according to Rule of judgment rand2 < A _iaMP.AMp.Amp f (x _i) < f (X ^*) judge, if condition meets, increase r _ireduce A _i, then enter step 6, wherein rand2 represents to produce a random number.If do not meet and directly enter step 6.

Step 6: arrange bat and find optimum solution X ^*;

Step 7: result: if termination condition meets, finishing iteration is exported optimum solution; Do not meet and return to second step and carry out iteration.

Wherein, in the space of d dimension search,

Generally need to carry out repeatedly iteration, termination condition is generally the default iterations upper limit.When the t time iteration is carried out step 2, the renewal of its medium velocity, position is shown below.

f _i＝f _min+(f _max-f _min)β (1)

$V_i^t=V_i^{t-1}+(X_i^t-X^{*})\&times;f_i---\left(2\right)$

$X_i^t=X_i^{t-1}+V_i^t---\left(3\right)$

Wherein, be the position X of the t time iteration _i, be that the t time iteration is at X _ithe speed V at place _i.The initial value of t is 1, and each step 7 of carrying out is carried out t=t+1 before returning to step 2, upgrades, until meet termination condition to continue to press above formula in step 2.When t=1, adopt initialization acquired results.

Can determine pulsed frequency f according to the field size of problem _ivalue, such as getting f _min=0, f _max=100, every bat Random assignment frequency during beginning, frequency can be from [f _min, f _max] on average draw.F _minfor minimum pulse frequency, f _maxfor maximum impulse frequency, can preset voluntarily value by those skilled in the art.β ∈ [0,1] is random vector, X ^*optimum solution for the current overall situation.At fixing λ _i(f _i) time change f _i(λ _i), and in the process of Local Search, the new position of every bat draws nearby, wherein λ _ithe wavelength that represents respective pulses frequency.As shown in formula (4).

X _new＝X _old+εA ^t (4)

Wherein, ε ∈ [1,1] is random number, X _oldrepresent every bat carry out Local Search time old position, X _newrepresent every bat carry out Local Search time reposition, A ^tfor the mean value of all bats in this iteration.

Along with the speed of bat and the renewal of position, the loudness A of impulse ejection _iwith speed r _ialso can upgrade.Once object is found conventionally, the transmitting loudness A of its pulse _ican reduce the emission rate r of pulse _ican raise, its expression formula is as follows.

$A_i^{t+1}={\mathrm{\&alpha;A}}_i^t---\left(5\right)$

$r_i^{t+1}=r_i^0(1-e^{-\mathrm{\&gamma;t}})---\left(6\right)$

In above formula, alpha, gamma is constant, and meets 0 < α < 1, γ > 0; E=2.71828.Under the span definition of alpha, gamma, along with the increase of iterations, the loudness A of impulse ejection _ican be tending towards gradually 0, the speed r of impulse ejection _ithe emission rate that can be tending towards gradually initialization pulse (its numerical value is selected near 0 conventionally), meets the principle of bat algorithmic formula.

This patent adopts the Rosenbroke detection function of standard to test bat algorithm, and its function expression is as follows.

$f\left(X\right)=\underset{i=1}{\overset{d-1}{\mathrm{\&Sigma;}}}{(1-x_i^2)}^2+100{(x_{i+1}-x_i^2)}^2---\left(7\right)$

Wherein, this paper value d=2, application scenario is two dimension, number individual in the population of bat is 25, i.e. 1≤i≤25 and i ∈ N ^*, α=γ=0.9.N ^*for positive integer.

Utilize bat algorithm to find the minimum value of f (X) herein, from mathematical theory, f (X), when (1,1), can obtain minimum value 0.

The bat searching route of formula (7) as shown in Figure 3.

From Fig. 3, can find out that bat algorithm is definite by region of search by 25 bats, then in the region of search among a small circle of determining, carry out the search of optimum solution, present on the whole a kind of search procedure of optimum solution of disorder to order.The point coordinate of last optimum solution is that (1.0006,1.0011) are more or less the same with theoretical value.

Specific embodiment described herein is only to the explanation for example of the present invention's spirit.Those skilled in the art can make various modifications or supplement or adopt similar mode to substitute described specific embodiment, but can't depart from spirit of the present invention or surmount the defined scope of appended claims.

Claims (2)

1. the Criminisi image repair method based on bat algorithm, is characterized in that, comprise the following steps,

Step 1, the right of priority of each pixel of edge, area to be repaired of calculating image to be repaired is as follows, and the pixel of choosing right of priority maximum is the preferential pixel of repairing,

priority(p)＝C(p)×D(p)

Wherein, the right of priority that priority (p) is edge pixel point p, C (p) is degree of confidence, D (p) is data item;

Step 2, for the preferential pixel of repairing of step 1 gained, carries out search and the filling of best matching blocks in the intact region of image to be repaired, search adopts bat algorithm to carry out according to SSD matching principle, and described SSD matching principle is as follows,

${\mathrm{\&psi;}}_{\hat{q}}=\arg \underset{{\mathrm{\&psi;}}_q\&Element;\mathrm{\&phi;}}{\min }d({\mathrm{\&psi;}}_{\hat{p}},{\mathrm{\&psi;}}_q)$

Wherein, represent the preferential corresponding object block to be repaired of pixel of repairing of step 1 gained, ψ _qrepresent the sample block in intact region φ, represent piece ψ _qthe quadratic sum of known pixels point color difference, for object block to be repaired best matching blocks;

Step 3, upgrades edge, area to be repaired, returns to step 1 and carries out repetitive cycling operation, until has repaired area to be repaired, obtains image repair result.

2. the Criminisi image repair method based on bat algorithm according to claim 1, is characterized in that: in step 1,

$C\left(p\right)=\frac{{\mathrm{\&Sigma;}}_{{i\&Element;\mathrm{\&psi;}}_p}\&cap;{\mathrm{\&phi;}}^{C\left(i\right)}}{\left|{\mathrm{\&psi;}}_p\right|}$

Wherein, variable C (i) is defined as C (i)=0, φ is intact region, and Ω is area to be repaired; | ψ _p| represent ψ _parea, ψ _prepresent the corresponding object block to be repaired of edge pixel point p;

$D\left(p\right)=\frac{|\&dtri;I\stackrel{\&perp;}{p}\&CenterDot;n_p|}{\mathrm{\&alpha;}}$

Wherein, n _pit is the border in area to be repaired at edge pixel point p unit outside normal vector; the direction and intensity that waits irradiation line that represents edge pixel point p place, its expression formula is i _x, I _yrepresent respectively the partial differential of edge pixel p in x, y direction; α is normalized parameter.