CN101996393B - Super-resolution method based on reconstruction - Google Patents

Abstract

The invention belongs to the technical field of computer multimedia technology and digital image processing, which relates to a super-resolution method based on reconstruction. On the basis of assumption of a boundary linear model, a low-resolution image is regarded as a combination of objects with polygon boundaries; a boundary segment expression of an original object is extracted from the low-resolution image; accordingly, in a high-resolution image, pixels on the boundaries are accurately colored; and other pixels not on the boundaries are colored according to a traditional method. Compared with traditional popular quick treating methods PR and LI, the super-resolution method of the invention can be used for obtaining a more accurate boundary so that the imaging effect is better and the image is clearer to the largest degree.

Description

A kind of based on the super-resolution method of rebuilding

Technical field

The invention belongs to Computer Multimedia Technology and digital image processing techniques field, relate to a kind of based on the super-resolution method of rebuilding.

Background technology

Super-resolution technique is meant, generates the technology of high-definition picture from the low-resolution image of input.The super-resolution technique applied scene is quite extensive, and for example, we need show same width of cloth image on the display of different resolution, televisor, mobile phone; And for example, we hope to improve the photographic quality that a pair is taken by low-resolution camera; Perhaps we hope to improve the resolution characteristic of image identification system through improving resolution; Also can upgrade to HDTV form image more clearly to the image of DVD form.

The notion of super-resolution technique is put forward by Tsai and Huang.Most basic super-resolution method comprises Pixel replication (PR).This method only zooms into single pixel the pixel of the same race of a square formation.Generally, the effect of PR and bad.Most super-resolution methods have all used space interpolation (Spatial interpolation).Wherein, linear interpolation Linear interpolation (LI) method is the most common.This method is set at a cum rights with the pixel of centre and heavily is the average of 4 pixel distances of this position and near adjacency.The list of references relevant with the present invention has:

1.T.Acharya?and?P.-S.Tsai.Computational?foundations?of?image?interpolation?algorithms.ACMUbiquity，8，2007.

2.C.B.Atkins，C.A.Bouman，and?J.P.Allebach.Optimal?image?scaling?using?pixel?classification.In?Proceedings?of?the?2001?International?Conference?on?Image?Processing(ICIP’01)，volume?3，pages?864-867，2001.

3.S.Baker?and?T.Kanade.Limits?on?super-resolution?and?how?to?break?them.In?Proceedings?of?the2000?IEEE?Conference?on?Computer?Vision?and?Pattern?Recognition(CPVR’00)，volume?2，pages372-379，2000.

4.P.Blomgren，G.Papanicolaou，and?H.Zhao.Super-resolution?in?time?reversal?acoustics.Journalof?the?Acoustical?Society?of?America，111：230-248，2002.

5.T.Blu，P.Th′evenaz，and?M.Unser.Generalized?interpolation：higher?quality?at?no?additionalcost.In?Proceedings?of?the?1999?International?Conference?on?Image?Processing(ICIP’99)，volume?3，pages?667-671，1999.

6.F.M.Candocia?and?J.C.Principe.Superresolution?of?images?with?learned?multiplereconstruction?kernels.In?L.Guan，S.Y.Kung，and?J.Larsen，editors，Multimedia?Image?and?VideoProcessing，chapter?4，pages?219-243.CRC?Press，New?York，2000.

7.A.Corduneanu?and?J.C.Platt.Learning?spatially-variable?filters?for?superresolution?of?text.InProceedings?of?the?2010?International?Conference?on?Image?Processing(ICIP’10)，volume?1，pages849-852，2005.

8.L.Davis.A?survey?of?edge?detection?techniques.Computer?Graphics?and?Image?Processing，4：248-270，1975.

9.T.K.Dey，K.Mehlhorn，and?E.A.Ramos.Curve?reconstruction：connecting?dots?with?goodreason.Computational?Geometry：Theory?and?Applications，10：289-303，2000.

10.S.Farsiu，M.Elad，and?P.Milanfar.A?practical?approach?to?super-resolution.In?Proceedings?ofthe?40th?Asilomar?Conference?on?Signals，Systems?and?Computers，2006.Invited?paper.

11.W.T.Freeman，T.R.Jones，and?E.C.Pasztor.Example-based?super-resolution.IEEE?ComputerGraphics?and?Applications，pages?56-65，2002.

12.T.L.Friedman.The?World?is?Flat：A?Brief?History?of?the?Twenty-First?Century.Farrar，Straussand?Giroux，2005.

13.M.Irani?and?S.Peleg.Super?resolution?from?image?sequences.In?Proceedings?of?the?10thInternational?Conference?on?Pattern?Recognition(ICPR’90)，volume?2，pages?115-120，1990.

14.Z.Jiang，T.-T.Wong，and?H.Bao.Practical?super-resolution?from?dynamic?video?sequences.InProceedings?of?the?2003?IEEE?Conference?on?Computation?Vision?and?Pattern?Recognition(CVPR’03)，volume?2，pages?549-554，2003.

15.T.M.Lehmann，C.

and?K.Spitzer.Survey：interpolation?methods?in?medical?imageprocessing.IEEE?Transactions?on?Medical?Imaging，18(11)：1049-1075，1999.

16.T.Lengauer?and?K.Mehlhorn.The?HILL?system：a?design?environment?for?the?hierarchicalspecification，compaction，and?simulation?of?integrated?circuit?layouts.In?Jr.Paul?Penfield，editor，Proceedings?of?the?MIT?VLSI?Conference.Artech?House，Inc.，1984.

17.X.Li?and?M.T.Orchard.New?edge-directed?interpolation.IEEE?Transactions?on?ImageProcessing，10(10)：1521-1527，2001.

18.Z.Lin?and?H.-Y.Shum.Fundamental?limits?of?reconstruction-based?superresolution?algorithmsunder?local?translation.IEEE?Transactions?on?Pattern?Analysis?and?Machine?Intelligence，26(1)：83-97，2004.

19.K.Mehlhorn.On?the?size?of?sets?of?computable?functions.In?In?Proceedings?of?the?14th?IEEESymposium?on?Automata?and?Switching?Theory，pages?190-196，1973.

20.K.Mehlhorn?and?S.

The?LEDA?Platform?for?Combinatorial?and?Geometric?Computing.Cambridge?University?Press，Cambridge，England，1999.

21.E.Meijering.A?chronology?of?interpolation：from?ancient?astronomy?to?modern?signal?andimage?processing.Proceedings?of?the?IEEE，90(3)：319-342，2002.

22.B.Mitra.Gaussian-based?edge-detection?methods：a?survey.IEEE?Transactions?on?Systems，Man，and?Cybernetics-Part?C：Applications?and?Reviews，32(3)，2002.

Summary of the invention

The present invention aims to provide a kind of based on the super-resolution method of rebuilding, and from low-resolution image, extracts the boundary sections expression formula of original objects, then borderline pixel is carried out accurately painted, thereby obtain high-definition picture.

Technical scheme of the present invention is: a kind of based on the super-resolution method of rebuilding, it comprises the steps:

Step 1: in given low-resolution image, at first seek those and have pixel, handle these pixels with copied pixels PR method then, obtain the gray-scale value C of each newly-generated corresponding high-resolution pixel in abutting connection with same color;

Step 2: distinguish boundary pixel and gradual change pixel, boundary pixel is calculated the expression formula of the boundary sections of passing their pixels; The gradual change pixel is gone to step 5 to be handled;

Step 3: the boundary line segment table is reached formula carry out approximate processing, make the intersection point of adjacent boundary line segment overlap, guarantee to couple together that the border that forms is continuous, smooth; If the boundary sections intersection point on the common edge does not overlap, then with the mid point of two intersection points as the border intersection point of two boundary pixels on this limit, go to step 2, recomputate the boundary line expression formula of this position;

Step 4: utilize the boundary line expression formula of gained, and the gray-scale value of both sides, border object, directly calculate the gray-scale value of respective pixel in the high-definition picture;

Step 5: be labeled as the gradual change pixel to all remaining pixels, handle with linear interpolation LI method.

Wherein said differentiation boundary pixel of step 2 and gradual change pixel, the method step that boundary pixel is calculated the expression formula of passing their pixel boundary line segments is following:

Step 21: the approximate match in the boundary line of two different pixels is become the straight-line segment with the surface level oblique, and the left and right sides of said straight-line segment is the color of two different pixels, and its gray-scale value is respectively C _lAnd C _rFor the boundary line of level, can said image perhaps be rotated counterclockwise 90 ° clockwise;

Step 22: get the unit grid square for said straight-line segment, the length of side of each grid is 0.5, makes 3 abscissa value that straight-line segment and adjacent two grids intersect be respectively a, b and c;

Step 23: set up two grey scale pixel value C that straight-line segment passes through adjacent two along slope coordinate lattice _bAnd C _tFormula;

Step 24: utilize the geometric relationship of a, b and c, try to achieve a, b and c value; If 0＜a, b, c＜1, then said two different pixels are boundary pixel, thereby obtain the formula that embodies of boundary straight line section; Otherwise said two different pixels are the gradual change pixel.

The super-resolution method that the present invention is based on reconstruction has some application, such as the resolution of the photo that can use super-resolution technique raising experience recognition of face instrument to shoot, thereby improves the success ratio of recognition of face; And for example can adopt this method to improve the resolution that mobile phone is taken pictures, make photograph more clear; And for example can use super-resolution technique that the DVD image is promoted the image as HDTV.

Compare with existing popular immediate processing method PR and LI, super-resolution method of the present invention can obtain border more accurately, makes imaging effect better, can make to a great extent that image becomes more clear.

Description of drawings

Fig. 1 is the FB(flow block) of super-resolution method of the present invention.

Fig. 2 is the border expression formula computing method synoptic diagram of super-resolution method of the present invention.

Fig. 3 (a) is an original low-resolution image;

Fig. 3 (b) is the image after super-resolution method of the present invention is handled;

Fig. 3 (c) is the image after the PR algorithm process;

Fig. 3 (d) is the image after the LI method is handled.

Fig. 4 (a) is original olympic logo pattern;

Fig. 4 (b) is the high resolving power olympic logo pattern after super-resolution method of the present invention is handled;

Fig. 4 (c) is the emblem partial enlarged drawing of Fig. 4 (a);

Fig. 4 (d) is the emblem partial enlarged drawing after super-resolution method of the present invention is handled;

Fig. 4 (e) is the emblem partial enlarged drawing after the PR algorithm process;

Fig. 4 (f) is the emblem partial enlarged drawing after the LI method is handled.

Embodiment

The thinking of super-resolution method of the present invention is: from the hypothesis based on the border linear model; Can regard low-resolution image as object have the combination of Polygonal Boundary; From low-resolution image, extract the boundary sections expression formula of original objects, in high-definition picture, borderline pixel is carried out accurately painted then; All the other non-boundary pixels carry out painted according to the conventional method.

Below in conjunction with accompanying drawing super-resolution method of the present invention is elaborated.

1.Exact super-resolution method

The flow process of super-resolution method Exact of the present invention is as shown in Figure 1, and concrete steps are following:

The first step to given low-resolution image, at first finds out those to have the pixel in abutting connection with the same color neighbours, handles these pixels with the PR method;

In second step, find the boundary pixel of one of all several kinds of situation that meet above description, and calculate the expression formula of the boundary line of passing these pixels.

The 3rd step reached formula to the boundary line segment table and carries out approximate processing, and it is continuous, smooth to make them couple together the border that forms.To all (horizontal direction, vertical direction) adjacent boundary pixels,, the mid point of two intersection points as the border intersection point of two boundary pixels on this limit, is recomputated the boundary line expression formula of this position if the boundary sections intersection point on the common edge does not overlap.

The 4th goes on foot, and utilizes the boundary line expression formula of gained, directly calculates the gray-scale value of respective pixel in the high-definition picture;

The 5th step was labeled as the gradual change pixel to all remaining pixels, handled with the LI method;

2. calculate on the border

For the expression formula of said computation bound line of second step, all remaining pixels after the first step handled are done following check, confirm whether this pixel is boundary pixel, suppose the straight line boundary line be about two different color gray-scale values be respectively C _lAnd C _r, as shown in Figure 2.Then the gray-scale value of several pixels of running through of this boundary line (like figure) is by following formula decision.

$c_b=\frac{a+b}{2}\·c_l+(1-\frac{a+b}{2})\·c_r.$

$c_t=\frac{b+c}{2}\·c_l+(1-\frac{b+c}{2})\·c_r.$

It is known again,

$b=\frac{a+c}{2},$

The substitution following formula obtains,

$a=\frac{3c_b-c_t-2c_r}{2(c_l-c_r)}$

$b=\frac{c_b+c_t-2c_r}{2(c_l-c_r)}.$

Promptly according to given C _l, C _r, C _bAnd C _t, can calculate a, b, c.If a of gained, b, c are not less than 0, are not more than 1, think that then this pixel is boundary pixel and the formula that embodies that obtains boundary straight line.

Intersect situation for other boundary sections and pixel, for example, linear barrier's level pass pixel, the present invention can be revolved the example that provides and turn 90 degrees.

Adopt Exact super-resolution method of the present invention that image in the real world has been done many experiments, list several results at this.

Pattern shown in Figure 3 is light and dark polygon, so all pattern boundaries all are straight lines, through different disposal routes, can find the image that high-definition picture quality that the Exact super-resolution method obtains obtains apparently higher than classic method.

The olympic logo pattern that employing distinct methods shown in Figure 4 is handled is not though the border of this pattern is a straight line.Yet after the Exact super-resolution method is handled, still can obviously be superior to the visual effect of classic method.Reason is under high definition case, and the polygonal profile section that fine and closely woven short line segment is formed is the approximating curve profile very.

Claims (1)

1. the super-resolution method based on reconstruction is characterized in that it comprises the steps:

Step 1: in given low-resolution image, at first seek those and have pixel, handle these pixels with copied pixels PR method then, obtain the gray-scale value C of each newly-generated corresponding high-resolution pixel in abutting connection with same color;

Step 2: distinguish boundary pixel and gradual change pixel, boundary pixel is calculated the boundary straight line section expression formula of passing their pixels; The gradual change pixel is gone to step 5 to be handled; Described differentiation boundary pixel and gradual change pixel, the method step that boundary pixel is calculated the boundary straight line section expression formula of passing their pixels is following:

Step 21: the approximate match in the boundary line of two different pixels is become the straight-line segment with the surface level oblique, and the left and right sides of said straight-line segment is the color of two different pixels, and its gray-scale value is respectively C ₁And C _rFor the boundary line of level, can said image perhaps be rotated counterclockwise 90 ° clockwise;

Step 22: get the unit grid square for said straight-line segment, the length of side of each grid is 0.5, makes 3 horizontal ordinates that straight-line segment and adjacent two grids intersect be respectively a, b and c;

Step 23: set up two grey scale pixel value C that straight-line segment passes through adjacent two along slope coordinate lattice _bAnd C _tFormula:

$C_b=\frac{a+b}{2}\×C_l+(1-\frac{a+b}{2})\×C_r$

$C_t=\frac{b+c}{2}\×C_l+(1-\frac{b+c}{2})\×C_r$

Known;

substitution following formula obtains:

$a=\frac{{3C}_b-C_t-2C_r}{2(C_l-C_r)}$

$b=\frac{C_b+C_t-2C_r}{2(C_l-C_r)}$

Step 24: according to given C _l, C _r, C _bAnd C _t, try to achieve a, b, c value; If 0＜a, b, c＜1, then said two different pixels are boundary pixel, thereby obtain boundary straight line section expression formula; Otherwise said two different pixels are the gradual change pixel;

Step 3: border straight-line segment expression formula is carried out approximate treatment, make the intersection point of adjacent boundary line segment overlap, guarantee to couple together that the border that forms is continuous, smooth; If the boundary sections intersection point on the common edge does not overlap, then with the mid point of two intersection points as the border intersection point of two boundary pixels on this limit, go to step 2, recomputate boundary straight line section expression formula;

Step 4: utilize the boundary straight line section expression formula of gained, and the gray-scale value of both sides, border object, directly calculate the gray-scale value of respective pixel in the high-definition picture;

Step 5: be labeled as the gradual change pixel to all remaining pixels, handle with linear interpolation LI method.

Images