CN102930531A - Detection method for repetition structure of building surface image - Google Patents
Detection method for repetition structure of building surface image Download PDFInfo
- Publication number
- CN102930531A CN102930531A CN2012103669136A CN201210366913A CN102930531A CN 102930531 A CN102930531 A CN 102930531A CN 2012103669136 A CN2012103669136 A CN 2012103669136A CN 201210366913 A CN201210366913 A CN 201210366913A CN 102930531 A CN102930531 A CN 102930531A
- Authority
- CN
- China
- Prior art keywords
- pixel
- image
- stay
- degree
- place form
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 49
- 238000000034 method Methods 0.000 claims abstract description 48
- 239000013598 vector Substances 0.000 claims abstract description 28
- 230000003252 repetitive effect Effects 0.000 claims description 52
- 238000009499 grossing Methods 0.000 claims description 10
- 238000012545 processing Methods 0.000 claims description 8
- 230000017105 transposition Effects 0.000 claims description 5
- 239000013604 expression vector Substances 0.000 claims description 3
- 230000001105 regulatory effect Effects 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 description 24
- 230000006870 function Effects 0.000 description 14
- 238000012360 testing method Methods 0.000 description 11
- 238000005516 engineering process Methods 0.000 description 9
- 230000008569 process Effects 0.000 description 5
- 238000013519 translation Methods 0.000 description 5
- 239000004744 fabric Substances 0.000 description 4
- 238000003709 image segmentation Methods 0.000 description 4
- PXFBZOLANLWPMH-UHFFFAOYSA-N 16-Epiaffinine Natural products C1C(C2=CC=CC=C2N2)=C2C(=O)CC2C(=CC)CN(C)C1C2CO PXFBZOLANLWPMH-UHFFFAOYSA-N 0.000 description 3
- 230000008901 benefit Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 239000000284 extract Substances 0.000 description 3
- 230000000007 visual effect Effects 0.000 description 3
- 238000009825 accumulation Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000005286 illumination Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000007667 floating Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000001028 reflection method Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a detection method for a repetition structure of a building surface image. The method comprises the following steps: constructing a structure template according to a structure unit marked by a user; using colour of pixels included in the structure template to construct a feature vector of the structure template; estimating grade attachment that each pixel of an image to be detected belongs to the structure template; using the colour of the pixels to estimate the affinity between adjacent pixels; using the affinity of the adjacent pixels to smooth the grade attachment that each pixel of the image to be detected belongs to the structure template; and using a ranking search method to extract a repetition structure similar to the structure template from the image to be detected by virtue of the grade attachment that each pixel of the image to be detected belongs to the structure template. The detection method provided by the invention solves a problem that the detection precision is low in the prior art, and a problem that uncertainty appears while defining a plurality of different repetition structures in the prior art, and has an extensive application prospect.
Description
Technical field
The present invention relates to computer vision and mode identification technology, particularly building surface image repetitive structure detection field.
Background technology
Repetitive structure is present in the surface of buildings widely.Such as, at residential neighborhoods, door, window and the balcony of same building residential building present identical structural unit usually.Building surface image repetitive structure detects and refers to detect the structural unit that repeats from captured building surface image.It is the important means that realizes building surface compression of images, image-based buildings three-dimensional reconstruction, the identification of image-based city landmark thing that building surface image repetitive structure detects, and has important application in fields such as historical relic's protection, urban construction, city management (showing such as the community based on virtual reality technology), security monitorings.
Existing most of building surface image repetitive structure detection technique mainly is based on the Symmetry Detection method and realizes.The Symmetry Detection method comprises rotational symmetry detection technique, Rotational Symmetry detection technique, slides symmetrical detection technique and the symmetrical detection technique of translation.The rotational symmetry detection technique is mainly for detection of the axis of symmetry that is implied between the building surface image repetitive structure.This technology is pocket with image segmentation at first; Then, determine axis of symmetry according to textural characteristics and the locus of pocket; At last, detect by to similar pocket according to axis of symmetry, thereby obtain the structure that repeats.The rotational symmetry detection technique depends on cutting apart image.But the core technology of relevant image segmentation not yet makes a breakthrough, and the difficult point that image segmentation remains in the computer vision field at present studies a question.This point has affected the application of rotational symmetry detection technique in building surface image repetitive structure detects.The Rotational Symmetry detection technique is mainly for detection of the structural unit through still remaining unchanged after the rotation of certain angle.This technology at first is rotated with various angle the small images of suitable size, and extracts its postrotational textural characteristics; Then, based on the topmost anglec of rotation of the similarity measurement between the textural characteristics; At last, by each pocket of traversing graph picture, detect the structure that repeats in the image.The structure that the Rotational Symmetry detection technique detects should have rotational invariance, so this technology is only applicable to the specific buildings of minority.The symmetrical detection technique of sliding axis has adopted the technical step identical with the rotational symmetry detection technique.Different is, the rotational symmetry detection technique is only estimated an axis of symmetry to whole image, but the symmetrical detection technique of sliding axis can be estimated a plurality of axis of symmetry, and axis of symmetry is arranged along a curve.Than the rotational symmetry detection technique, the image range that the symmetrical detection technique of sliding axis be suitable for is larger, but this technology still needs Image Segmentation Using, so has weakened its practicality.The symmetrical detection technique of translation has adopted with Rotational Symmetry and has detected identical technical step.Different is that the symmetrical detection technique of translation realizes by structural unit is carried out translation in image.The symmetrical detection technique of translation is subject to illumination easily, block the impact with noise.Summary is got up, and above-mentioned Symmetry Detection method needs to specify in advance symmetrical type, and accuracy of detection depends on employed texture characteristic extracting method simultaneously.But, how to extract effective textural characteristics and remain at present a challenging difficult problem.
Owing to have many limitations based on the building surface image repetitive structure detection technique of Symmetry Detection method, the related researcher begins to pay close attention to that development need not at first to carry out Symmetry Detection and direct-detection goes out the method for repetitive structure at present.At present, direct detecting method mainly comprises the probability density estimation technique, three-dimensional point cloud proof method and low-rank texture matrix method.The probability density estimation technique is at first estimated the probability density function about repetitive structure; Then, obtain the maximal value of probability density function, and the corresponding structural unit of maximal value is joined in the existing Markov random field as new node; At last, again carry out probability density and estimate, and detect iteratively in this way all repetitive structures.But, only have when image comprises the same structure of One's name is legion and probability density function and estimate that when more accurate, it is effective that the method just can become.The three-dimensional point cloud proof method is to utilize the image extraction of a plurality of visual angles shooting about the three-dimensional point cloud of buildings, and detects repetitive structure by the mutual checking between the three-dimensional point cloud subset.The method need to be estimated three-dimensional point cloud from several input pictures, calculated amount is very huge.Low-rank texture matrix method forms a texture matrix with the texture feature vector of image fritter, then finds the structure that repeats by analyzing low-rank submatrix that this matrix comprises.This method mainly contains two shortcomings: one, estimate that the order of all submatrixs is very consuming time; Its two, even if the submatrix of order minimum, usually might not be corresponding interested repetitive structure, thereby export incorrect testing result.
In sum, existing building surface image repetitive structure detection technique is ripe far away, and accuracy of detection is not high, calculated amount is large.The research that relevant building surface image repetitive structure detects also is in the experimental verification stage mostly, lacks practical repetitive structure detection technique.In addition, existing most of repetitive structure detection techniques are not limited the structure that detects, and usually causing the result who detects not is to be the desired repetitive structure that obtains of user.The main cause that causes this problem is that most of building surfaces comprise multiple different repetitive structure.In this case, during the repetitive structure of the number of different types in defining image to be detected, prior art can run into uncertain problem, thus the testing result that leads to errors.Therefore, in actual applications, the user wishes the structure identical with user annotation that can accurately obtain to comprise in the image by marking an interested structure to reach the effect that gets twice the result with half the effort.
Summary of the invention
The technical matters that (one) will solve
The object of the invention is to overcome deficiency of the prior art, so that the user utilizes the structural unit of its mark to extract repetitive structure rapidly and accurately from image to be detected.
(2) technical scheme
For solving the problems of the technologies described above, the invention provides a kind of building surface image repetitive structure detection method, the method comprises the steps:
Step 1: the structural unit according to user annotation makes up stay in place form;
Step 2: the proper vector of utilizing the color structure stay in place form of pixel that stay in place form comprises;
Step 3: estimate that each pixel of image to be detected belongs to the degree of membership of stay in place form;
Step 4: utilize the color of pixel to estimate affinity degree between neighbour's pixel;
Step 5: utilize the affinity degree between neighbour's pixel, treat the degree of membership that each pixel of detected image belongs to stay in place form and carry out smoothly;
Step 6: utilize each pixel of image to be detected to belong to the degree of membership of stay in place form, adopt the sorted search method from image to be detected, to extract the repetitive structure similar to stay in place form.
Preferably, in step 3, the degree of membership that described image pixel belongs to stay in place form is calculated as follows:
Wherein, s
pFor the pixel p of image to be detected belongs to the degree of membership of stay in place form, v
pBe the proper vector of pixel p, v is the proper vector of the stay in place form of step 2 acquisition, || || the expression vector is asked the long computing of mould, and subscript T represents transposition, and p is that natural number and span are 1 to w * h, and w is the width of image to be detected, and h is the height of image to be detected.
Preferably, in step 4, the affinity degree between neighbour's pixel is calculated as follows:
Wherein, a
P, qAffinity degree between expression pixel q and the pixel p; r
p, g
p, b
pThe redness, green and the blue component value that represent respectively pixel p; r
q, g
q, b
qThe redness, green and the blue component value that represent respectively pixel q; σ is a positive adjustment factor, is used for regulating the span of affinity; The exponent arithmetic of function exp () expression take natural number as the truth of a matter; Q is natural number, and its span is the index value of neighbour's pixel of pixel p; P is that natural number and span are 1 to w * h, and w is the width of image to be detected, and h is the height of image to be detected.
Preferably, the acquiescence value of described adjustment factor σ is 30.
Preferably, the described degree of membership that each pixel of image is belonged to stay in place form is carried out smoothing processing and is realized by finding the solution following minimum model:
Wherein, t
p, p=1,2 ..., w * h, the expression pixel p is variable to be found the solution through the degree of membership that belongs to stay in place form after level and smooth; s
pThe degree of membership that does not belong to stay in place form through level and smooth pixel p for step 3 acquisition; a
P, qBe the affinity degree between pixel p and the pixel q; N
pFor by the four index value set that are communicated with neighbour's pixel of the pixel p that syntoples obtain; λ is a given positive balance parameters, is used for the contribution of two of these minimum model objective function front and back of balance; Q is natural number, and its span is defined in index value set N
pThe index value of neighbour's pixel of the middle pixel p that comprises; t
qFor pixel q through the degree of membership that belongs to stay in place form after level and smooth; P is that natural number and span are 1 to w * h; W is the width of image to be detected; H is the height of image to be detected; Min represents that this model is minimum model.
Preferably, the acquiescence value of described balance parameters λ is 0.001.
(3) beneficial effect
Method provided by the present invention can detect the repetitive structure that comprises in the building surface image rapidly and accurately, is mainly reflected in the following aspects:
(1) utilizes user annotation, avoided the uncertain problem that when the repetitive structure that defines number of different types exists simultaneously, occurs, and the repetitive structure of user's appointment can have been separated from image-region;
(2) repetitive structure is detected the computational problem that is considered as each pixel degree of membership of estimation, have intuitively, be easy to the characteristics such as programming;
(3) only depend on the pixel color of image to be detected, need not extra Visual Feature Retrieval Process method, can realize that repetitive structure detects fast and accurately;
In addition, method provided by the present invention has broken through prior art and has been difficult to determine the interested repetitive structure of user from the repetitive structure of number of different types, has broken through simultaneously prior art and has been difficult to many restrictions that selected appropriate visual signature is described repetitive structure.The method is used as stay in place form with the structural unit of user annotation, comes level and smooth each pixel of image to be detected to belong to the degree of membership of stay in place form by global optimization's model, has improved the precision that repetitive structure detects.Building surface image repetitive structure detection method provided by the invention, accuracy of detection is high, calculating is quick, has broad application prospects.
Description of drawings
Fig. 1 is the process flow diagram according to a kind of building surface image repetitive structure detection method of the present invention;
Fig. 2 is about stay in place form and belongs to the numbering of the pixel of this stay in place form;
Fig. 3 a is the image of a width of cloth external walls of building and a window of user annotation;
Fig. 3 b has showed the resulting window testing result of the method according to this invention, and wherein, the border of each detected window all is labeled out;
Fig. 4 a is the image of a width of cloth external walls of building and an arched door of user annotation;
Fig. 4 b has showed the resulting arched door testing result of the method according to this invention, and wherein, the border of each detected arched door all is labeled out.
Embodiment
For making the purpose, technical solutions and advantages of the present invention clearer, below in conjunction with specific embodiment, and with reference to accompanying drawing, the present invention is described in further detail.
Fig. 1 is the process flow diagram of building surface image repetitive structure detection method provided by the present invention.As shown in Figure 1, the step of building surface image repetitive structure detection method provided by the invention comprises:
Step S1: the structural unit according to user annotation makes up stay in place form;
Step S2: the proper vector of utilizing the color structure stay in place form of pixel that stay in place form comprises;
Step S3: estimate that each pixel of image to be detected belongs to the degree of membership of stay in place form;
Step S4: utilize the color of pixel to estimate affinity degree between neighbour's pixel;
Step S5: utilize the affinity degree between neighbour's pixel, treat the degree of membership that each pixel of detected image belongs to stay in place form and carry out smoothly;
Step S6: utilize each pixel of image to be detected to belong to the degree of membership of stay in place form, adopt the sorted search method from image to be detected, to extract the repetitive structure similar to stay in place form.
The below is described in detail above steps.
Step S1: the structural unit according to user annotation makes up stay in place form.
A given coloured image to be detected is supposed that the repetitive structure that this image comprises has identical size, and is not had affine and perspective distortion each other.For having affine and situation perspective distortion, can be by affine and perspective distortion bearing calibration commonly used to correct image.
For the image of pending detection, suppose that the user has marked out the border of an interested structural unit by mouse from this image.In this way, the user has defined the structural unit that needs extraction.This structural unit will be used as benchmark architecture, be used for being limited to image to be detected and need the structural unit that extracts.
From image to be detected, take out the structural unit of user annotation, draw the width of structural unit by its difference that comprises the minimum and maximum space horizontal ordinate of pixel, draw the height of structural unit by its difference that comprises the minimum and maximum space ordinate of pixel.The width of note structural unit is w
sIndividual pixel, the height of structural unit are h
sIndividual pixel.
Then, the pixel that the structural unit of user annotation is comprised is numbered.Particularly, the pixel that from top to bottom repetitive structure is comprised from left to right by the mode of line scanning is numbered continuously since 1.Remember the n that is numbered of last pixel.That is to say that structural unit comprises altogether n pixel.After finishing above-mentioned numbering, we have obtained a stay in place form.Therefore, stay in place form is a rectangular area, and its width is w
sIndividual pixel highly is h
sIndividual pixel; This rectangular area comprises altogether n pixel, and each pixel has the unique integer numbering from 1 to n.In this way, stay in place form defined the structural unit that need to from image to be detected, extract size and with stay in place form with the rectangular area of size in which pixel belong to this structural unit.Fig. 1 provided a stay in place form and comprise the synoptic diagram of the numbering of pixel.
At last, to be calculated as width be w to the geometric center point of stay in place form
sWith length h
sThe center of rectangle.Particularly, the horizontal ordinate of the geometric center point of stay in place form is [(w
s+ 1)/2], the ordinate of the geometric center point of stay in place form is [(h
s+ 1)/2], wherein [] represents rounding operation.
Step S2: the proper vector of utilizing the color structure stay in place form of pixel that stay in place form comprises.
At first, the stay in place form to step S1 obtains takes out the pixel that it comprises by number in turn, according to redness, green and the blue component value of each pixel, forms the proper vector of stay in place form.The proper vector of note stay in place form is v, and calculates as follows:
v＝[r
1，g
1，b
1，r
2，g
2，b
2，…，r
n，g
n，b
n]
T，
Wherein, r
1, g
1, b
1Be illustrated respectively in the redness, green and the blue component value that are numbered 1 pixel in the stay in place form; r
2, g
2, b
2Be illustrated respectively in the redness, green and the blue component value that are numbered 2 pixel in the stay in place form; r
n, g
n, b
nBe illustrated respectively in redness, green and the blue component value of the pixel that is numbered n in the stay in place form; N represents the number of the pixel that stay in place form comprises that step S1 obtains; Subscript T represents transposition.Like this, the proper vector v of stay in place form is a column vector that length is 3n.
Step S3: estimate that each pixel of image to be detected belongs to the degree of membership of stay in place form.
If the width of image to be detected is w pixel, highly be h pixel.Mode by line scanning is numbered each pixel from left to right from top to bottom.Like this, each pixel of image all obtains one and is positioned at 1 unique integer index number between w * h.
At first, distribute a behavior h to classify the matrix of w as, each pixel that is used for document image belongs to the degree of membership of stay in place form, remembers that this matrix is A, and the element of matrix A is initialized as zero.
Secondly, to pixel p, p=1,2 ..., w * h, the geometric center point of the stay in place form that step S1 is obtained is positioned over the pixel p place of image to be detected, and obtaining thus a width is w
sIndividual pixels tall is h
sThe image fritter of individual pixel.Here, w
sBe the width of the stay in place form that obtains among the step S1, h
sHeight for the stay in place form that obtains among the step S1.If stay in place form exceeds the border of image to be detected, fill the zone that is beyond the boundary in this image fritter by the pixel along edge reflection method duplicating image commonly used in the image processing.For convenience of description, remember that this image fritter is I
pAs seen, image fritter I
pLarge young pathbreaker and stay in place form big or small identical.
The numbering of the pixel that then, comprises according to stay in place form is to image fritter I
pThe pixel of middle same position is numbered.Particularly, be numbered 1 the residing position of pixel in the reference structure template, with image fritter I
pIn be in same position pixel number be 1; Be numbered 2 the residing position of pixel in the reference structure template, with image fritter I
pIn be in same position pixel number be 2; According to said method analogize.Because stay in place form comprises n pixel, so image fritter I
pIn only have n pixel to obtain numbering.According to above-mentioned numbering, make up the proper vector of pixel p.The proper vector of note pixel p is v
p, and calculate as follows:
Wherein,
Be illustrated respectively in image fritter I
pIn be numbered redness, green and the blue component value of 1 pixel;
Be illustrated respectively in image fritter I
pIn be numbered redness, green and the blue component value of 2 pixel;
Be illustrated respectively in image fritter I
pIn be numbered redness, green and the blue component value of the pixel of n; N represents the number of the pixel that stay in place form comprises that obtains among the step S1; Subscript T represents transposition.As seen, the proper vector v of pixel p
pIt is a column vector that length is 3n.
Then, calculating pixel p belongs to the degree of membership of stay in place form.Note s
pFor pixel p belongs to the degree of membership of stay in place form, and be calculated as follows:
Wherein, v
pBe the proper vector of pixel p, v is the proper vector of the stay in place form of step S2 acquisition, || || the expression vector is asked the long computing of mould, and subscript T represents transposition.Because the proper vector v of pixel p
pWith each element of proper vector v of stay in place form all greater than 0, so degree of membership s
pTo be [0, a 1] interval interior floating number.As seen, degree of membership s
pProper vector v by pixel p
pAnd the similarity between the proper vector v of stay in place form is calculated.Therefore, degree of membership s
pIn fact measured image fritter I
pSimilarity degree with stay in place form.
At last, with degree of membership s
pBe assigned to equally image fritter I
pIn the numbered pixel of all tools, and the degree of membership that is recorded in matrix A adjusted.Particularly, at image fritter I
pNumbered each pixel of middle tool if this pixel does not exceed image boundary, checks that this pixel is recorded in the degree of membership of matrix A, if its value is less than the degree of membership s of current acquisition
p, then use numerical value s
pWith its replacement; Otherwise, do not change the degree of membership that this pixel has recorded in matrix A.By aforesaid operations, image fritter I
pIn the numbered pixel of each tool all obtained identical degree of membership, thereby take full advantage of the pixel distribution characteristic of stay in place form, help to improve the precision of detection.
By this step, after each pixel of traversing graph picture, each pixel that we have obtained image to be detected belongs to respectively the degree of membership of stay in place form, and is recorded in the matrix A.
Step S4: utilize the color of pixel to estimate affinity degree between neighbour's pixel.
At first, treat the pixel p of detected image, p=1,2 ..., w * h is communicated with syntople by four and takes out the pixel that is adjacent.Here, w is the width of image to be detected, and h is the height of image to be detected.If pixel p is the interior pixels of image, it will have 4 neighbour's pixels to be communicated with syntople by four; If pixel p is the boundary pixel of image, it will have 2 or 3 neighbour's pixels.Note N
pIndex value set for all neighbour's pixels of pixel p.As seen, according to the residing position of pixel p, index value set N
pThe index value that comprises 2,3 or 4 neighbour's pixels, and each index value is an integer between 1 to w * h.
Secondly, the affinity degree of calculating pixel p and its neighbour's pixel.Note q is index value set N
pIn an index value.This index value is corresponding to the pixel q of image to be detected.Affinity degree between note pixel q and the pixel p is a
P, q, and calculate as follows:
Wherein, r
p, g
p, b
pThe redness, green and the blue component value that represent respectively pixel p; r
q, g
q, b
qThe redness, green and the blue component value that represent respectively pixel q; σ is a positive adjustment factor, is used for regulating the span of affinity; The exponent arithmetic of function exp () expression take natural number as the truth of a matter; Q is natural number, and its span is defined in index value set N
pThe index value of neighbour's pixel of the middle pixel p that comprises; P is that natural number and span are 1 to w * h; W is the width of image to be detected; H is the height of image to be detected.
As seen, the color of pixel p and pixel q is more approaching, and affinity degree between the two can be higher.
In the method provided by the present invention, adjustment factor σ is set as 30.Considering that the value of color component of image pixel is the highest can reach 255, sets a less adjustment factor here and can make the affinity degree of two not identical pixels of color close to zero.
Step S5: utilize the affinity degree between neighbour's pixel, treat the degree of membership that each pixel of detected image belongs to stay in place form and carry out smoothly.
At first, make up the optimal model that belongs to the degree of membership of stay in place form about each pixel of image to be detected, facilitate the use this model degree of membership is carried out smoothly.Remember that each pixel is t through the degree of membership after level and smooth
p, p=1,2 ..., w * h.Here, w is the width of image to be detected, and h is the height of image to be detected.Introduce the degree of membership that following minimum model comes each pixel of image to be belonged to stay in place form and carry out smoothing processing:
Wherein, t
p, p=1,2 ..., w * h, the expression pixel p is variable to be found the solution through the degree of membership that belongs to stay in place form after level and smooth; s
pBe recorded in the degree of membership that pixel p in the matrix A belongs to stay in place form for what step S3 obtained, be pixel p not through the degree of membership that belongs to stay in place form of smoothing processing; a
P, qBe the pixel p that obtains among the step S4 and the affinity degree between the pixel q; N
pFor in step S4 by the four index value set that are communicated with neighbour's pixel of the pixel p that syntoples obtain; λ is a given positive balance parameters, is used for the contribution of two of the above-mentioned minimum model objective function of balance front and back; Q is natural number, and its span is defined in index value set N
pThe index value of neighbour's pixel of the middle pixel p that comprises; t
qFor pixel q through the degree of membership that belongs to stay in place form after level and smooth; P is that natural number and span are 1 to w * h; W is the width of image to be detected; H is the height of image to be detected; Min represents that this model is minimum model.
In the method provided by the present invention, balance parameters λ is set as 0.001.Here, set a less balance parameters and be in order to increase first contribution to optimal model of objective function, both can not be partially too far away afterwards before level and smooth and smoothly to guarantee each pixel belonging to the degree of membership of stay in place form.
We are further explained above-mentioned minimum model.The objective function of this minimum model consists of by two.First of objective function by each pixel before level and smooth and the quadratic sum of the difference of the degree of membership smoothly consist of.Under the minimization of object function meaning, this implication can be regarded as each pixel, and numerically both can not be partially too far away before optimization and through the degree of membership after optimizing.Second quadratic sum by the difference of the degree of membership of each pixel and its neighbour's pixel consists of.This implication can be explained as follows: if the color similarity of two neighbour's pixels, through after the smoothing processing, their degree of membership also should be close, and the quadratic sum of the difference of its degree of membership could be close to zero like this, thereby obtains minimum target function value.With regard to whole structure, the effect of above-mentioned minimum model is to utilize the affinity between neighbour's pixel that the degree of membership of each pixel of step S3 acquisition is carried out smoothly.Different from the smoothing methods such as employed medium filtering, gaussian kernel convolution in the existing image processing, the inventive method takes full advantage of the affinity between neighbour's pixel.Therefore, the degree of membership smoothing method based on neighbour's pixel affinity that the present invention introduces is a kind of smoothing processing method that drives based on view data, thereby can obtain more accurate testing result.
Above-mentionedly change most minimodel to comprise altogether the variable that w * h needs optimize (be t
p, p=1,2 ..., w * h).Each quadratic term of extended target function can find that this objective function only comprises once item and the quadratic term about variable to be optimized.In addition, this minimum model does not contain Prescribed Properties.Therefore, this minimum model is a nothing constraint quadratic programming model.According to Optimum Theory, this model has a globally optimal solution, and what employing was commonly used can obtain this globally optimal solution rapidly based on conjugate gradient iterative method.
At last, by the index value of each pixel, will be through the degree of membership t after level and smooth
p, p=1,2 ..., w * h, the behavior h of being filled up to classify as among the matrix of w.The note matrix H is used for each pixel of record through the degree of membership after the smoothing processing.Particularly, matrix H upper left corner element has recorded the degree of membership of image top left corner pixel to be detected; The first row secondary series element of matrix H has recorded the degree of membership of the first row secondary series pixel of image to be detected; By that analogy, the lower right corner element of matrix H has recorded the degree of membership of the lower right corner pixel of image to be detected.
Step S6: utilize each pixel of image to be detected to belong to the degree of membership of stay in place form, adopt the sorted search method from image to be detected, to extract the repetitive structure similar to stay in place form.
At first, distribute the column vector that length is w * h, this vector will be for record degree of membership accumulation result.Remember that this vector is b.Adopt identical mark among the step S3, w is the width of image to be detected, and h is the height of image to be detected.
Secondly, by the mode of line scanning, the element of the matrix H that from top to bottom step S5 is obtained from left to right is numbered.Like this, each element of matrix H all obtains one and is positioned at 1 unique integer index number between w * h.
Then, to the element i of matrix H, i=1,2 ..., w * h, the geometric center point of the stay in place form that step S1 is obtained is positioned over the element i place of matrix H.If stay in place form exceeds the scope of matrix H, i the element of vectorial b is set to-1.To this situation, owing to by zone that stay in place form the covered size less than stay in place form, therefore do not consider this pixel.If stay in place form is placed in the matrix H fully, from matrix H, take out a behavior h
sClassify w as
sSubmatrix.Here, w
sBe the width of the stay in place form that obtains among the step S1, h
sHeight for this stay in place form.Stay in place form as mask, is added up to the element in this submatrix.Particularly, in this submatrix, will be added up by the element that stay in place form covered, and with accumulation result divided by the number of pixels that stay in place form comprises, obtain average membership, and this average membership be assigned to i the element of vectorial b.
All w * h element of Ergodic Matrices H, so vectorial b has recorded the average membership of each pixel of image to be detected and stay in place form.Obtain owing to the average membership of each pixel calculates as mask with stay in place form, so it has been measured centered by this pixel and with stay in place form and has had the image-region of formed objects and the possibility that stay in place form belongs to same structure.
At last, w * h the element of vectorial b sorted, find out element value maximum among the vectorial b.If should be worth greater than 0.85, find the position of its corresponding pixel in image to be detected by the index value of this element.Then, the geometric center point of stay in place form is positioned over this pixel place, takes out the pixel that stay in place form covers, obtain a structural unit identical with stay in place form.Then, take out by the pixel that stay in place form covered, according to the call number of these pixels the value of the element of correspondence position among the vectorial b is set to-1.By above-mentioned steps, carry out repeatedly above-mentioned searching vector b the maximal value element, detect maximal value whether greater than 0.85 and the element of vectorial b carried out again the computation process of assignment, until can not detect again till the repetitive structure that makes new advances.
Detect in the process of repetitive structure in above-mentioned iterative manner, consider that may there be the impact of the factors such as noise and illumination in image, so the present invention adopts 0.85 rather than detect repetitive structure with higher numerical value as threshold value.
Introduced after the ins and outs in the embodiment, the following describes test effect of the present invention.In order to verify validity of the present invention, test with the building surface image.
Fig. 3 has provided the image of a width of cloth external walls of building.This image comprises the window that repeatedly occurs.Fig. 3 a has provided a window of user annotation, and this window will as a basic structural unit, be used to specify the interested repetitive structure of user.Therefore, the task of present embodiment is to detect other identical with the window of user annotation in this image window.According to the window of user annotation, obtain the repetitive structure testing result by carrying out method provided by the present invention.Fig. 3 b has provided detected window, and wherein, the border of each detected window all is labeled out.As seen, all windows are all correctly detected.
Fig. 4 has provided the image of another width of cloth external walls of building.This image comprises window, arched door and other structure that repeats.If these dissimilar repetitive structures are not added restriction, the structure that it is the desired detection of user that existing algorithm will be difficult to clear and definite any structure.Fig. 4 a has provided an arched door of user annotation.By its interested structural unit of user annotation, reduced when multiple different structure all repeatedly repeats existing algorithm and be difficult to judge the difficult problem that to export which kind of structure.Therefore, the task of present embodiment is to detect other identical with the arched door of user annotation in this image arched door.According to the arched door of user annotation, obtain the repetitive structure testing result by carrying out method provided by the present invention.Fig. 4 b has provided detected arched door, and wherein, the border of each detected arched door all is labeled out.As seen, all arched doors are all correctly detected.
Experiment shows that the method according to this invention can detect the structure of user's appointment effectively.The present invention adopts each pixel of computed image to belong to the degree of membership of stay in place form and based on the degree of membership smoothing method of the global optimum of neighbour's pixel affinity, can obtain high-precision repetitive structure testing result.Aspect operation time, experiment shows, to the image of 800 * 600 pixel sizes, on the computing machine of 3.0GHz CPU, 2GB internal memory, the inventive method only needs can provide the repetitive structure testing result in about 4 seconds in C language computing environment.
Above-described specific embodiment; purpose of the present invention, technical scheme and beneficial effect are further described; be understood that; the above only is specific embodiments of the invention; be not limited to the present invention; within the spirit and principles in the present invention all, any modification of making, be equal to replacement, improvement etc., all should be included within protection scope of the present invention.
Claims (6)
1. building surface image repetitive structure detection method, the method comprises the steps:
Step 1: the structural unit according to user annotation makes up stay in place form;
Step 2: the proper vector of utilizing the color structure stay in place form of pixel that stay in place form comprises;
Step 3: estimate that each pixel of image to be detected belongs to the degree of membership of stay in place form;
Step 4: utilize the color of pixel to estimate affinity degree between neighbour's pixel;
Step 5: utilize the affinity degree between neighbour's pixel, treat the degree of membership that each pixel of detected image belongs to stay in place form and carry out smoothly;
Step 6: utilize each pixel of image to be detected to belong to the degree of membership of stay in place form, adopt the sorted search method from image to be detected, to extract the repetitive structure similar to stay in place form.
2. the method for claim 1 is characterized in that, in described step 3, the degree of membership that described image pixel belongs to stay in place form is calculated as follows:
Wherein, s
pFor the pixel p of image to be detected belongs to the degree of membership of stay in place form, v
pBe the proper vector of pixel p, v is the proper vector of the stay in place form of step 2 acquisition, || || the expression vector is asked the long computing of mould, and subscript T represents transposition, and p is that natural number and span are 1 to w * h, and w is the width of image to be detected, and h is the height of image to be detected.
3. the method for claim 1 is characterized in that, in described step 4, the affinity degree between neighbour's pixel is calculated as follows:
Wherein, a
P, qAffinity degree between expression pixel q and the pixel p; r
p, g
p, b
pThe redness, green and the blue component value that represent respectively pixel p; r
q, g
q, b
qThe redness, green and the blue component value that represent respectively pixel q; σ is a positive adjustment factor, is used for regulating the span of affinity; The exponent arithmetic of function exp () expression take natural number as the truth of a matter; Q is natural number, and its span is the index value of neighbour's pixel of pixel p; P is that natural number and span are 1 to w * h, and w is the width of image to be detected, and h is the height of image to be detected.
4. method as claimed in claim 3 is characterized in that, the acquiescence value of described adjustment factor σ is 30.
5. the method for claim 1 is characterized in that, the described degree of membership that each pixel of image is belonged to stay in place form is carried out smoothing processing and realized by finding the solution following minimum model:
Wherein, t
p, p=1,2 ..., w * h, the expression pixel p is variable to be found the solution through the degree of membership that belongs to stay in place form after level and smooth; s
pThe degree of membership that does not belong to stay in place form through level and smooth pixel p for step 3 acquisition; a
P, qBe the affinity degree between pixel p and the pixel q; N
pFor by the four index value set that are communicated with neighbour's pixel of the pixel p that syntoples obtain; λ is a given positive balance parameters, is used for the contribution of two of these minimum model objective function front and back of balance; Q is natural number, and its span is defined in index value set N
pThe index value of neighbour's pixel of the middle pixel p that comprises; t
qFor pixel q through the degree of membership that belongs to stay in place form after level and smooth; P is that natural number and span are 1 to w * h; W is the width of image to be detected; H is the height of image to be detected; Min represents that this model is minimum model.
6. method as claimed in claim 5 is characterized in that, the acquiescence value of described balance parameters λ is 0.001.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201210366913.6A CN102930531B (en) | 2012-09-28 | 2012-09-28 | Detection method for repetition structure of building surface image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201210366913.6A CN102930531B (en) | 2012-09-28 | 2012-09-28 | Detection method for repetition structure of building surface image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN102930531A true CN102930531A (en) | 2013-02-13 |
| CN102930531B CN102930531B (en) | 2015-06-10 |
Family
ID=47645321
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201210366913.6A Active CN102930531B (en) | 2012-09-28 | 2012-09-28 | Detection method for repetition structure of building surface image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102930531B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107945165A (en) * | 2017-11-24 | 2018-04-20 | 常州大学 | Textile flaw detection method based on peak value coverage values and areal calculation |
| CN107977961A (en) * | 2017-11-24 | 2018-05-01 | 常州大学 | Textile flaw detection method based on peak value coverage values and composite character |
| CN116563171A (en) * | 2023-07-11 | 2023-08-08 | 深圳大学 | Point cloud enhancement method and related equipment for building repeated structure |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101625755A (en) * | 2009-08-06 | 2010-01-13 | 西安电子科技大学 | Image division method based on watershed-quantum evolution clustering algorithm |
| WO2011091717A1 (en) * | 2010-01-29 | 2011-08-04 | The Hong Kong University Of Science And Technology | Architectural pattern detection and modeling in images |
-
2012
- 2012-09-28 CN CN201210366913.6A patent/CN102930531B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101625755A (en) * | 2009-08-06 | 2010-01-13 | 西安电子科技大学 | Image division method based on watershed-quantum evolution clustering algorithm |
| WO2011091717A1 (en) * | 2010-01-29 | 2011-08-04 | The Hong Kong University Of Science And Technology | Architectural pattern detection and modeling in images |
Non-Patent Citations (4)
| Title |
|---|
| CHANGCHANG WU ET AL.: "Detecting Large Repetitive Structures with Salient Boundaries", 《11TH EUROPEAN CONFERENCE ON COMPUTER VISION》, 11 September 2010 (2010-09-11), pages 142 - 155, XP019150561 * |
| GRANT SCHINDLER ET AL.: "Detecting and Matching Repeated Patterns for Automatic Geo-tagging in Urban Environments", 《IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION,2008》, 28 June 2008 (2008-06-28), pages 1 - 7, XP031297019 * |
| NIANJUAN JIANG ET AL.: "Multi-view Repetitive Structure Detection", 《2011 IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION(ICCV)》, 13 November 2011 (2011-11-13), pages 535 - 542, XP032101238, DOI: doi:10.1109/ICCV.2011.6126285 * |
| 李伟伟等: "3D模型中重复结构的多尺度快速检测算法", 《第六届和谐人机环境联合学术会议（HHME2010)、第19届全国多媒体学术会议（NCMT2010）、第6届全国人机交互学术会议（CHCI2010）、第5届全国普适计算学术会议（PCC2010）论文集》, 31 December 2010 (2010-12-31), pages 1 - 6 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107945165A (en) * | 2017-11-24 | 2018-04-20 | 常州大学 | Textile flaw detection method based on peak value coverage values and areal calculation |
| CN107977961A (en) * | 2017-11-24 | 2018-05-01 | 常州大学 | Textile flaw detection method based on peak value coverage values and composite character |
| CN107945165B (en) * | 2017-11-24 | 2019-10-11 | 常州大学 | Textile flaw detection method based on peak value coverage values and areal calculation |
| CN116563171A (en) * | 2023-07-11 | 2023-08-08 | 深圳大学 | Point cloud enhancement method and related equipment for building repeated structure |
| CN116563171B (en) * | 2023-07-11 | 2023-11-28 | 深圳大学 | Point cloud enhancement method and related equipment for building repeated structure |
Also Published As
| Publication number | Publication date |
|---|---|
| CN102930531B (en) | 2015-06-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US9619691B2 (en) | Multi-view 3D object recognition from a point cloud and change detection | |
| Xiong et al. | Automatic creation of semantically rich 3D building models from laser scanner data | |
| CN107093205B (en) | A kind of three-dimensional space building window detection method for reconstructing based on unmanned plane image | |
| CN106296693B (en) | Based on 3D point cloud FPFH feature real-time three-dimensional space-location method | |
| Tang et al. | Automatic reconstruction of as-built building information models from laser-scanned point clouds: A review of related techniques | |
| CN105405133B (en) | A kind of remote sensing image variation detection method | |
| US9036915B2 (en) | Architectural pattern detection and modeling in images | |
| CN105809651B (en) | Image significance detection method based on the comparison of edge non-similarity | |
| CN105513070A (en) | RGB-D salient object detection method based on foreground and background optimization | |
| CN104850850A (en) | Binocular stereoscopic vision image feature extraction method combining shape and color | |
| CN103226821A (en) | Stereo matching method based on disparity map pixel classification correction optimization | |
| CN101398886A (en) | Rapid three-dimensional face identification method based on bi-eye passiveness stereo vision | |
| CN103080979B (en) | From the system and method for photo synthesis portrait sketch | |
| CN104200212A (en) | Building outer boundary line extraction method based on onboard LiDAR (Light Detection and Ranging) data | |
| Yuan et al. | Learning to count buildings in diverse aerial scenes | |
| CN105139379A (en) | Airborne Lidar point cloud building top surface gradual extraction method based on classifying and laying | |
| CN103345760B (en) | A kind of automatic generation method of medical image object shapes template mark point | |
| US20150131871A1 (en) | Floor plan space detection | |
| CN102446356A (en) | Parallel and adaptive matching method for acquiring remote sensing images with homogeneously-distributed matched points | |
| Guo et al. | Exploring GIS knowledge to improve building extraction and change detection from VHR imagery in urban areas | |
| CN104866852B (en) | Extract the method and device of windy and sandy soil information in remote sensing image | |
| CN102930531B (en) | Detection method for repetition structure of building surface image | |
| CN105354845B (en) | A kind of semi-supervised change detecting method of remote sensing image | |
| CN108447084B (en) | Stereo matching compensation method based on ORB characteristics | |
| CN107038456A (en) | A kind of image classification method of the probability linear discriminant analysis based on L1 norms |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant |