CN1731449A - A method of image restoration - Google Patents

Abstract

The invention relates to an image restoring method which dose automatically restoring course after user enters into the restored area. The automatically restoring course uses the edge side as the unit of primary iteration packing and then iterates the following four steps until packs all the restored area: first step, computing all the precedent degree of the edge; second step, ascertaining the packing edge side of the current iteration by the precedent degree and the similarity degree criterion; third step, adopting line synthesis method and expends the known line and structure information; fourth step, updating the synthetic picture element ascertain value of the current iteration side.

Description

A kind of image repair method

Technical field

The invention belongs to computer virtual reality and computer graphics techniques field, specifically, is a kind of image repair technology---object removes on the reconstruct of damaged image and the image.

Background technology

Usually, image repair itself is a kind of art.As far back as the Renaissance, ancient times, the phenomenon of little breakage arm just appearred in sculpture, thereby, the sculptor not only self wants original work, also will repair simultaneously the sculpture in ancient times, it mainly is slight crack or the ditch of filling up on the artistic products to be occurred, as G.Emile-Male. " The Restorer ' s Handbook of Easel Painting " the .Van NostrandReinhold of document 1, New York, 1976 introductions.China is a historical relic big country, for the historical relic resource to preciousness is forever protected, reappears the original appearance of ancient character picture or old photo.China's ancient character draws the seal that face often is printed on the collector, perhaps because of the folding pleat trace that produces of the remote past; On the other hand, the photo of real scene or synthetic image often need remove some unnecessary big block objects on the photo, and the literal of suspension etc., particularly the pleat trace on some old photos need to repair.Therefore, we need seek a kind of image repair method of robotization.

The impaired part of image and remove the cavity that part can regard that all the information of losing forms as, therefore, the problem that the image repair technology mainly solves is how to fill these cavities.In the past, two kinds of typical image repair methods are arranged: the constraint synthetic (constrained synthesis) in the synthetic field of texture, A.Efros and T.Leung as document 2, " Texture synthesis by nonparametricsampling; " in Proc.Int.Conf.Computer Vision, Kerkyra, Greece, Sept.1999, pp.1033-1038 and L.Liang, C.Liu, Y.-Q.Xu, B.Guo, and H.-Y.Shum, " Real-time texture synthesis by patch-based sampling, " in ACMTrans.Graphics, 2001 that introduce and digital picture reparation (imageinpainting) methods mathematical computations, as the Bertalmio of document 3, M, Sapiro, G., Caselles, V., Ballester, C. " lmage lnpainting " .SIGGRAPH 2000, pages 417-424.

Preceding a kind of, be based on equine husband random field (Markov Random Field based on the constraint texture synthesis method in the texture synthesis method of master drawing, MRF) model that is: utilizes the repeatability and the randomness feature of texture image, fills the information of losing according to known image information on every side.A bit being that the square dough sheet (Patch) (size is defined by the user) at center is a match window on the regional edge to be filled, utilize the known portions that comprises in this window, in known image, search the dough sheet that mates most, fill the unknown portions in this window.This method is suitable for repairing the bigger affected area of area.But the reparation object of this method is confined to texture image, just seems powerless for the stronger image of linear structure.

The back is a kind of, and the digital picture restorative procedure of mathematical computations is a kind of digital picture restorative procedure based on physics hot-fluid partial differential equation (partial differential equations).Known colouring information (three passages of RGB) permeates expansion (utilizing the differential equation) along isophote (isophote) direction (perpendicular to the color gradient vector) to the area to be repaired.Thereby inwardly expand Given information when having realized preserving marginal information.Two dimension Laplacian conversion is used for realizing the local smooth change transition of color, and expands along the isophote direction.After the multiple processing of each rebuilding, all to carry out smoothing processing to the zone of being repaired.This method can keep the linear structure of original image effectively, is suitable for repairing the elongated zones such as literal of slight crack, suspension on the old photo.But when repairing big zone, adopt the method for infiltration expansion can lose interior details, and tangible blooming occurs.

Afterwards, image repair (exemplar-based image inpainting) method based on master drawing, the M.K.Leung and Y-H.Yang.First sight:a human body outlinelabeling system.IEEE Trans.Pattern Analysis and Machine Intelligence of document 5, (17) 4:359-377,1995 combine the advantage of above-mentioned two kinds of methods effectively, utilize texture synthesis method copy texture and structural information simultaneously, and the expansion of Structural Characteristics information depends on the order of filling.The core concept of restorative procedure is to utilize isophote (isophote) control chart as sampling process, and detailed process is as follows:

(a) calculate the priority of all edged faces: priority decision fill order is the key factor that texture synthesis method can be realized the expansion of linear structure information.

(b) expansion of texture and structural information: at first, select the highest dough sheet of priority as in this iteration with the edged faces of filling; Then, be similar to general texture synthesis method, the distance between interior two dough sheets of color space in the whole known region of original image, is searched with it dough sheet that mates (distance is minimum) most as the measurement criterion of similarity; At last, the unknown pixel color value in this edged faces directly duplicates the color value in the corresponding match block.

(c) certainty factor (confidence) value of all synthetic pixels is updated to the certainty factor value of this edged faces in this iteration.

The extended mode of the certainty factor value of area to be repaired interior pixel is similar to the extended mode of known color information in the digital picture repair process of mathematical computations; And actual color that should the zone interior pixel obtains by the texture synthesis method based on master drawing.The better effects if that this method is repaired, but also there are some problems: the first, in step (b), utilize colouring information to weigh similarity merely; The second, in step (b), fill order is just determined according to priority simply, and has been ignored the similarity principle; The 3rd, in step (c), during calculating, the pixel certainty factor do not consider the influence of matching error.These problems affect have arrived the robustness of this method and final repairing effect, because can not prevent a ubiquitous problem---spreading of error message in the texture building-up process effectively.

Summary of the invention

Technology of the present invention is dealt with problems: overcome the deficiencies in the prior art, a kind of robustness height, image repair method that repairing effect is good are provided, the user only need specify the area to be repaired, and the ancient character that this method can be repaired in real pictures, composograph and the digital museum with complex texture and architectural feature is automatically drawn.

Technical solution of the present invention: a kind of image repair method, on basis based on image repair (the exemplar-based image inpainting) thought of master drawing, its characteristics are: the user selects zone to be repaired on the image, and the scope in definite localized source zone, promptly be to carry out automatic repair process behind the master drawing sample area window size at center with the edged faces, automatically repair process is with the unit of edged faces as an iteration filling, four steps below repeating, up to filling up whole area to be repaired: the first step, the priority of all dough sheets on the edge calculation; In second step, determine that according to priority and similarity criterion this iteration is with the edged faces of filling; The 3rd step, adopt texture synthesis method, expand known texture and structural information simultaneously; In the 4th step, upgrade the pixel certainty factor value that is synthesized in the dough sheet of this iterative processing.

Image repair method of the present invention is with the advantage that existing restorative procedure is compared: reparation effective, repair process more approaches the repaired by hand process, the robustness height, prevent the continuous expansion of error message effectively, successfully be applied to multiple the have real pictures of complex texture and architectural feature or the reparation of composograph.In addition, the present invention also is applied to the reparation of digital museum's calligraphy and painting in the middle ancient times.

Description of drawings

Fig. 1 illustrates symbol definition involved among the present invention;

Fig. 2 illustrates limited texture building-up process;

Fig. 3 illustrates in certain iterative process the process flow diagram of determining edged faces to be filled according to priority and similarity criterion;

Fig. 4 illustrates the importance that the fill order that proposes among the present invention is adjusted strategy;

Fig. 5 illustrates the present invention introduces gradient information in the similarity judgment criterion necessity.

Embodiment

At first, define some symbols, as shown in Figure 1: image I to be repaired; Zone Ω is area to be repaired (zone to be filled), is called the target area, is specified by the user, and shape is not subjected to any restriction; Ω is meant the edge line of current not fill area, and along with the carrying out of repairing, Ω ceaselessly changes; Φ is the known portions (I-Ω) in the image, is the sample space of filling the Ω zone, is called source region; P, Q are any 2 points on the Ω, Ψ _p, Ψ q is respectively to be the edged faces (Patches) at center with P, Q, w is the edged faces window size.Obviously, along with the carrying out of repairing, regional Φ constantly enlarges, and Ω will diminish, and Ω also changes thereupon.

Then, introduce some related notions and method:

1. edged faces: Ψ as shown in Figure 1 _p, Ψ q dough sheet, to be the center a bit on the edge, be the unit of Given information expansion.The window size of edged faces as the W among Fig. 1, has very big influence to the result who repairs.If it is excessive that window size is got, lost the local detail feature, produce significantly discontinuous; If window size is too small, lose textural characteristics completely, produce a large amount of fragments.This size should be a bit larger tham the size of a line unit (texel, discernible texel) in theory, and default value is 7 among the present invention, and the user also can do suitable adjusting.

2. localized source zone: regional θ as shown in Figure 1 is a square area at center with the edged faces.According to the local similar principle of image, it is in the master drawing sample area at center that the search of match block is limited in the edged faces, can improve the efficient of algorithm.The big or small default value of localized source regional window is 5 times of edged faces window, and the user also can do suitable adjusting.

3. similarity judgment criterion:

(B) distance between two dough sheets in space is a module commonly used in the texture synthesis method as the judgment criterion of similarity for R, G to adopt color.Distance is big more, and similarity is more little, otherwise similarity is big.Be defined as follows:

$D=\sqrt{\frac{1}{A}\underset{i=1}{\overset{i=A}{\mathrm{\Σ}}}{\left((R_{\mathrm{pi}}-R_{\mathrm{qi}}\right)}^2+{(G_{\mathrm{pi}}-G_{\mathrm{qi}})}^2+{(B_{\mathrm{pi}}-B_{\mathrm{qi}})}^2}---\left(1\right)$

$d_{\max }=\sqrt{\frac{1}{A}\underset{p\∈({\mathrm{\Ψ}}_p\∩\stackrel{\‾}{\mathrm{\Ω}})}{\mathrm{\Σ}}({(\∈\mathrm{Rp})}^2+{\left(\mathrm{\ϵGp}\right)}^2+{\left(\mathrm{\ϵBp}\right)}^2)}---\left(2\right)$

D represents two edged faces Ψ _pWith the distance of Ψ q at color space, A represents Ψ _pThe number of interior known pixels, R, G, B represent three Color Channels respectively.Dmax represents the maximum match error when front piece allowed, between the two apart from the similarity constraint condition that is called less than dmax in the RGB color space, ε is a consumer premise justice parameter, similarity and synthetic result between Controlling Source master drawing and the composograph, the ε value is bigger than normal then can occur significantly discontinuous, the ε continuous repetition that then can cause same texture less than normal.

4. mate: two dough sheets are in the RGB color space, and distance is called both less than dmax and is complementary.

5. candidate item chained list list: to a certain edged faces, be used for specially depositing and its distance in the RGB color space less than all coupling dough sheets of dmax.

6. limited texture is synthetic: black partly is target area to be filled as shown in Figure 2, edged faces Ψ _p(yellow line collimation mark will is the center with red pixel point on the edge) is the current part that is filled.At first, according to the similarity judgment criterion of formula (1) (2), in known region, all dough sheets of the similarity constraint condition in the RGB color space are satisfied in search, shown in the dotted yellow line frame among Fig. 2, deposit candidate item chained list list in.Then, optional dough sheet Ψ q from list, edged faces Ψ _pIn the unknown pixel color value directly copy the color of respective pixel in the Ψ q.Usually, limited texture and Chengdu are according to certain synthetic order, and document 2 adopts the sweep trace order, then fill according to snail (onion-skin) mode ecto-entad is successively synthetic in the document 3.

Image repair method of the present invention begins automatic repair process after the selected area to be repaired of user and some parameters of initialization, will introduce automatic repair process in detail below.

Texture synthesis method can be realized the expansion of linear structure information, depends on the order of filling to a great extent.In the image repair method that the present invention proposes, fill order determines by two factors---priority and similarity, priority are principal elements, and similarity is the adjusting factor.Automatically repair process, with on the edge a bit be the dough sheet at center as a unit that recirculates and fill, four steps below repeating, all pixel in filling up whole target area.

The first step: the priority of all dough sheets on the edge calculation

Introduce according to document 5, priority is determined by two factors: certainty factor and the inside tensor that shrinks.

Degree of certainty: the known quantity of information and the correctness thereof that comprise in the expression edged faces.The degree of certainty factor guarantees to be positioned on the peripheral edge, comprises the abundant dough sheet of Given information and repairs at first, tends to the mode that ecto-entad is successively repaired.

Inwardly shrink tensor: the size of the component of isophote direction vector on normal vector of expression central pixel point, the dynamics that decision is permeated in restoring area, relatively large as the inside contraction tensor that P point and Q among Fig. 1 are ordered.Inwardly shrink tensor and impel on the edge line Ω direction, the dough sheet of change color big (the image gradient value is big) is repaired at first, keeps the continuity of original image internal linear structure.

The account form of priority is as follows: with 1 edged faces Ψ p that P is the center on the border, priority is P (p), is determined by degree of certainty function C (p) and toe-in tensor function D (p):

P(p)＝C(p)*D(p) (3)

$C\left(p\right)=\frac{{\mathrm{\Σ}}_{q\∈({\mathrm{\Ψ}}_p\∩\stackrel{\‾}{\mathrm{\Ω}})}C\left(q\right)}{\left|{\mathrm{\Ψ}}_p\right|}---\left(4\right)$

$D\left(p\right)=\frac{\left|{\▿}_p^{\⊥}g{n}_p\right|}{\mathrm{\α}}---\left(5\right)$

For general (0-255) gray-scale map, the α value in the formula (5) is 255, n _pExpression boundary line Ω is at the normal vector at center point P place, _p ^⊥Expression isophote direction is (perpendicular to Ψ _pThe color gradient vector of interior two component absolute value sum maximums _p), as shown in Figure 1.

Obviously, the colouring information in the Φ zone is that former figure has, and does not change the degree of certainty height afterwards in the repair process.When initialization, the degree of certainty functional value of the pixel in the Φ is 1, and the pixel degree of certainty functional value in the Ω is minimum, is 0.

Second step: determine that according to priority and similarity criterion this iteration is with the edged faces of filling

As shown in Figure 3, determine this iteration, need to introduce a counter, prevent that similarity constraint condition is too harsh in the RGB color space, and enter endless loop the edged faces of filling.Process is as follows:

(1) select the highest edged faces of priority (to use Ψ as current dough sheet to be filled _pExpression),

With Ψ _pFor the regional θ at center is current localized source zone.Simultaneously with counter O reset.

(2) search in source region θ, according to the similarity judgment criterion, all dough sheets of similarity constraint condition deposit among the candidate item chained list list in the RGB color space with satisfying.

(3) if candidate item chained list non-NULL, then this edged faces Ψ _pBe defined as the dough sheet to be filled of this iteration; Otherwise, illustrate in current known source sampling figure not have the master drawing piece that is complementary with this edged faces, through further repairing, just might occur its similar.Therefore, continue (4) the following step.

(4) counter adds 1.

(5) if counter does not reach user-defined number of times (this method is in experimentation, and getting this value is 10), continue the 6th) step, otherwise increasing the ε value is original 1.5 times, turns back to for (1) step.

(6) this dough sheet is hung up, selected the highest dough sheet of residue edged faces medium priority (to use Ψ _pExpression), the dough sheet to be filled as current returns to (2).

Be illustrated in figure 4 as certain iteration, the dough sheet Patch to be filled (1) that priority is the highest obviously is a kind of special dough sheet---can not find match block in the drawings.And next edged faces Patch (n) of priority can find match block under the similarity constraint condition in current RGB color space.If adopt the method in the document 5, do not consider the similarity problem, directly directly fill with a little dough sheet of similarity, can cause artificial cutter trade.(a) being the filling process of this method according to similarity adjustment fill order, (b) is single filling process that relies on priority decision fill order.Relatively both as seen, the fill order of this method more approaches artificial order of repairing, that is: for the target area shown in the figure, shrink to the centre gradually along the trend of current known curve from two examples up and down, is compiled in a bit at last.

The 3rd step:, expand known texture and structural information simultaneously based on the texture synthesis method of master drawing

Employing is expanded known texture and structural information simultaneously based on the texture synthesis method of master drawing, and process is as follows:

(1) adds gradient information in the similarity measurement criterion

For the stronger zone of grain (shown in Fig. 5 (a), the part in the green wire frame), adopt the judgment criterion of the distance of formula (1) definition as similarity, just can find approximate coupling dough sheet.But, for the zone in the purple wire frame among Fig. 5 (a), amplify the back for scheming (b) (c), there are the continuous boundary lines (respectively as the edge that yellow line indicated of figure (b) in (c)) of two kinds of different trends in this zone, the edged faces of figure (b) blue wire frame correspondence, and weights are higher (according to the computing method of above-mentioned weights, isophote is parallel to the normal vector of central point), be filled at first, keep the continuity at the edge that yellow line is indicated in the original image, than being easier to.But for the pairing edged faces of figure (c) sky blue wire frame, weights lower (the isophote direction is perpendicular to the central point normal vector), the very difficult continuity that keeps the edge line of yellow line segment mark in the original image.In order to address this problem, the present invention introduces another prominent feature---the gradient information in this range of linearity, and the measurement criterion of similarity is the distance in color and two spaces of gradient, is defined as follows:

$D=\sqrt{\frac{1}{A}\underset{i=1}{\overset{i=A}{\mathrm{\Σ}}}({(C_{\mathrm{pi}}-C_{\mathrm{qi}})}^2+{(G_{\mathrm{pi}}-G_{\mathrm{qi}})}^2)}---\left(6\right)$

$d_{\max }=\sqrt{\frac{1}{A}\underset{p\∈({\mathrm{\Ψ}}_p\∩\stackrel{\‾}{\mathrm{\Ω}})}{\mathrm{\Σ}}({\left(\mathrm{\ϵCp}\right)}^2+{\left(\mathrm{\ϵ}{G}_p\right)}^2)}---\left(7\right)$

D represents two edged faces Ψ _pAnd the distance between the Ψ q, A represents Ψ _pThe number of interior known pixels, C represents the RGB color vector, brightness step vector in the G presentation video.Dmax represents the maximum match error that current edged faces allows, and is called the similarity constraint condition of this edged faces in color and two spaces of gradient less than dmax, and ε is a consumer premise justice parameter (with the formula of document 2 introductions).

(2) double coupling

In order to improve the efficient of calculating, the present invention proposes a kind of new matching algorithm---double matching algorithm.

First weight is searched in the θ of localized source zone, according to the similarity judgment criterion of formula (1) and (2).Front second goes on foot according to the process of searching match block in the similarity adjusting fill order process and finishes, and here, directly the candidate item chained list list that preserves is exported as it, is equipped with second and heavily mates usefulness.

Second weight as new sampling source, according to formula (6) (7), is searched the master drawing piece of distance less than the dmax value with candidate item chained list list, forms new candidate item chained list.Be similar among Fig. 5 (a) the stronger image of grain in the green square zone, through first heavily coupling just can obtain similar dough sheet, but just need second heavily to mate for complex image, obtain the master drawing dough sheet that has more similar features.

(3) fill

In new candidate item chained list, select a Ψ q, Ψ arbitrarily _pThe color of middle unknown pixel is filled with the color of respective pixel among the Ψ q.

The 4th step: revise (Ψ pI Ω) interior pixel certainty factor value

From above-mentioned information expansion process as can be known, filling to being filled between the dough sheet can only be within the specific limits similar, rather than identical, therefore have matching error to exist, and along with the carrying out of repairing, the increase of cumulative errors can cause spreading of error message.Be the influence of reflection matching error, the pixel certainty factor function of the present invention's design is determined by two factors: the correctness of the certainty factor of place edged faces and Fill Color during filling.The same formula of definition (3) of edged faces certainty factor function; The correctness function definition of Fill Color is exp (kD ²), the same formula of the definition of D (6), k are adjustment factors.Therefore, the certainty factor value defined of arbitrary pixel is in the Ψ pI Ω:

C(p′)＝C(p)exp(-kD ²)p′∈(Ψ _pIΩ) (8)

From formula (8) as can be known, the closer to the core of Ω, the colouring information correctness that is filled is low more, and pixel qualitative value really is more little, conforms to the actual conditions of repairing image.

Claims (4)

1, a kind of image repair method is characterized in that may further comprise the steps:

(1) user selects zone to be repaired on the image;

(2) automatic repair process, this automatic repair process are the units of filling as iteration with edged faces, and in four steps below repeating, up to filling up the whole all pixels in the zone that are repaired, its step is as follows:

The first step, the priority of all dough sheets on the edge calculation;

In second step, determine that according to priority and similarity criterion this iteration is with the edged faces of filling;

The 3rd step, adopt texture synthesis method, expand known texture and structural information simultaneously;

In the 4th step, upgrade the pixel certainty factor value that is synthesized in the dough sheet of this iterative processing.

2, image repair method according to claim 1, it is characterized in that: second step, described definite this iterative process with the strategy that the edged faces of filling adopts was: fill order is by priority and two factors decisions of similarity, priority is principal element, similarity is the adjusting factor, and its filling process is as follows:

(a) select the highest edged faces of priority as current dough sheet to be filled;

(b) in the RGB color space, select the match block that satisfies this space similarity constraint condition, store in the chained list,, otherwise continue (c) if the chained list non-NULL then is defined as this edged faces this iteration with the dough sheet of filling;

(c) this dough sheet is hung up, selected the highest dough sheet of residue edged faces medium priority,, get back to (b) as current dough sheet to be filled.

3, image repair method according to claim 1, it is characterized in that: described employing texture synthesis method of the 3rd step, expand simultaneously in known texture and the structural information process and to adopt double matching algorithm: first weight, directly utilize the chained list that obtains in claim 2 (b) step output result as this heavy coupling; Second weight in color and two spaces of gradient, in the match block that goes out from first gravity treatment, is selected the match block that satisfies color and two spaces of gradient similarity constraint condition.

4, image repair method according to claim 1 is characterized in that: the certainty factor value of the edged faces at place and the matching error of this dough sheet determined jointly when the certainty factor value of employing pixel was filled by pixel in the certainty factor computing method of the 4th step pixel.

Images