CN102663442B - Automatic Matching Method of Irregular Regions Based on Straight-line Dichotomy - Google Patents

Abstract

The invention relates to a method for automatic matching of irregular areas in digital images based on position dichotomy, comprising: collecting images and inputting them into a computer; using existing area detection methods to detect areas; calculating the maximum symmetrical position of the area; and calculating the characteristics of each point in the area Vector; use the gradient amplitude extremum to calculate the key points of the region; calculate the region description vector corresponding to the region key point; obtain the region descriptor by calculating the mean vector and standard deviation vector of each region description vector; use the region descriptor to perform region matching. Compared with the existing area matching method, the method provided by the invention does not need to perform fixed shape fitting, nor does it need to perform block processing, and has better stability to image deformation.

Description

Irregular area automatic matching method based on straight line dichotomy

Technical field

The present invention relates to the automatic matching method of the image region of disorder characteristic of field in a kind of computer vision.

Background technology

Image Feature Matching technology has important application at numerous areas such as image retrieval, object identification, video tracking and augmented realities.In the last few years with yardstick invariant features conversion (Scale Invariant Feature Transform, abbreviation SIFT) proposition of technology is sign, greater advance has been obtained in Image Feature Matching field, formed the Feature Correspondence Algorithm of a collection of maturation, as SIFT, GLOH, Shape Context etc. (is specifically shown in document A performance evaluation of local descriptors.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2005,27 (10): 1615-1630).

The basic ideas that various existing matching process adopt when provincial characteristics is mated in image are as follows: first irregular area is fitted to regular domain (as circle, ellipse, square) or directly selection rule region as the supporting zone of structure description; Then on regular domain, carry out the subregion of fixed size and divide (supporting zone being divided into the subregion of several fixed sizes), finally by calculating subregion descriptor and carrying out Region Matching.It is different that the difference of the whole bag of tricks is mainly to choose the feature that supporting zone shape is different, subregion is divided difference and structure realm descriptor is chosen.

But, during due to photographic images, visual angle often changes, between one group of image to be matched, often there is deformation, the picture material that existing method utilizes solid shape will cause supporting zone to comprise as supporting zone is inconsistent, and it is also inconsistent to be fixed the big or small subregion division content that all subregion comprises afterwards, thereby reduced the accuracy of matching result, can say that existing various Region Matching methods are more responsive for image deformation.

Summary of the invention

The present invention is directed to the tender subject of existing Region Matching method to deformation in digital picture, object is to provide a kind of automated regional matching process deformation to better stability.In order to realize this object, the present invention is based on the irregular area automatic matching method of straight line dichotomy, comprise the following steps:

Step S1: gather image and input computing machine;

Step S2: utilize existing method for detecting area surveyed area;

Step S3: zoning maximum symmetric position;

Step S4: each point gradient vector in region and regional average value vector are carried out to the proper vector that computing obtains each point in region;

Step S5: utilize gradient magnitude extreme value zoning key point;

Step S6: the region description vector that zoning key point is corresponding;

Step S7: the mean vector and the standard deviation vector that calculate each region description vector obtain region description.

Step S8: utilize the Euclidean distance between region description to carry out Region Matching.

Irregular area automatic matching method based on straight line dichotomy provided by the invention, mainly utilized can remain unchanged when the deformation with respect to the position relationship of straight line this character of key point in image-region, first key point in the maximum symmetric position of definite area, the proper vector of calculating each point definite area, then utilize respectively each key point and maximum symmetric position to form straight line region is divided into two sub regions and calculates description vectors, finally by calculating the statistic of each description vectors, obtain region description and mate.Because point and the relative position relation of straight line remain unchanged under image deformation, therefore utilize the definite straight line of each key point and maximum symmetric position to divide with feature and be described under deformation and there is stability region.Method provided by the invention neither needs to carry out shape matching and does not need to be again fixed big or small subregion division in structure realm descriptor process, reduce the error causing due to image deformation, therefore aspect the stability of image deformation, be better than existing method.

Accompanying drawing explanation

Fig. 1 is the irregular area automatic matching method process flow diagram that the present invention is based on straight line dichotomy.

Embodiment

Be illustrated in figure 1 the irregular area automatic matching method process flow diagram that the present invention is based on straight line dichotomy, comprise step: gather image and input computing machine; Utilize existing method for detecting area surveyed area; Zoning maximum symmetric position; The proper vector of each point in zoning; Utilize gradient magnitude extreme value zoning key point; The region description vector that zoning key point is corresponding; By calculating mean vector and the standard deviation vector of each region description vector, obtain region description; Utilize region description to carry out Region Matching.The concrete implementation detail of each step is as follows:

Step S1: take from different perspectives Same Scene two width different images and input computing machine;

Step S2: utilize existing region detection technique to carry out region detection, as used MSER technology;

Step S3: the maximum symmetric position P that calculates each region G _c, concrete mode is, centered by the arbitrary position P in the G of region, to draw 18 straight line L _iby whole circumference equal dividing, it is 36 parts; Straight line L in note G _ithe pixel count of both sides is respectively N _l(i), N _r(i), definition asymmetry for P place; By the position P of asymmetry minimum in the G of region _cbe defined as the maximum symmetric position of region G;

Step S4: the proper vector of each point in the G of zoning, specifically mode is, utilizes the gradient vector of each point in Gauss's gradient template zoning G, Gauss's gradient that in note G, some X (x, y) locates is [d _x, d _y], in G, the average gradient of each point is [V _x, V _y], the proper vector s=[s that calculation level X (x, y) locates ₁, s ₂], s wherein ₁=d _xv _x+ d _yv _y, s ₂=d _xv _y-d _yv _x;

Step S5: utilize gradient magnitude extreme value zoning key point, concrete mode is, any point X (x in note region G, y) gradient magnitude of locating is E (x, y), under threshold value T constraint, the point that will be maximum value in 3 * 3 neighborhood inside gradient amplitudes, as the key point in the G of region, meets following condition:

E(x，y)＞T，E(x，y)＞E(x+1，y+1)，E(x，y)＞E(x-1，y-1)，

E(x，y)＞E(x-1，y)，E(x，y)＞E(x+1，y)，E(x，y)＞E(x，y-1)，

E (x, y) > E (x, y+1), E (x, y) > E (x-1, y+1), E (x, y) > E (x+1, y-1); The concrete of described threshold value T determines that method is: T=Mean (E)+kStd (E), and Mean (E) and Std (E) represent respectively average and the standard deviation of each point gradient magnitude in described region, the span of scale-up factor k is 2～3;

Step S6: region description vector corresponding to key point in the G of zoning, concrete mode is that the key point that note step S5 obtains region G is respectively P ₁, P ₂... P _n, key point P _isymmetric position P with region G _cdefinite straight line is divided into two sub regions by region G, and to the gradient magnitude summation of each point in two sub regions, note gradient magnitude is sued for peace more greatly, less subregion is respectively G _b, G _s; In two sub regions, distinguish the positive and negative each point proper vector component that step S4 is obtained and add up, can obtain key point P _i8 corresponding dimension region description vector V _i=[v _i1, v _i2..., v _i8], wherein $v_{i1}=\underset{s_1\left(X\right)>0\mathrm{andX}\&Element;G_B}{\mathrm{\&Sigma;}}{s}_1\left(X\right),$ $v_{i2}=\underset{s_1\left(X\right)<0\mathrm{andX}\&Element;G_B}{\mathrm{\&Sigma;}}{s}_1\left(X\right),$ $v_{i3}=\underset{s_2\left(X\right)>0\mathrm{andX}\&Element;G_B}{\mathrm{\&Sigma;}}{s}_2\left(X\right),$ $v_{i4}=\underset{s_2\left(X\right)<0\mathrm{andX}\&Element;G_B}{\mathrm{\&Sigma;}}{s}_2\left(X\right),$ ${\mathrm{<figure-callout\; id="5"\; label="v\; i"\; filenames="HSA00000685584300011.png"\; state="\{\{state\}\}">v</figure-callout>}}_i=\underset{s_1\left(X\right)>0\mathrm{andX}\&Element;G_S}{\mathrm{\&Sigma;}}{s}_1\left(X\right),$ ${\mathrm{<figure-callout\; id="6"\; label="v\; i"\; filenames="HSA00000685584300011.png"\; state="\{\{state\}\}">v</figure-callout>}}_i=\underset{s_1\left(X\right)<0\mathrm{andX}\&Element;G_S}{\mathrm{\&Sigma;}}{s}_1\left(X\right),$ $v_{i7}=\underset{s_2\left(X\right)>0\mathrm{andX}\&Element;G_S}{\mathrm{\&Sigma;}}{s}_2\left(X\right),$ $v_{i8}=\underset{s_2\left(X\right)<0\mathrm{andX}\&Element;G_S}{\mathrm{\&Sigma;}}{s}_2\left(X\right);$

Step S7: average and the standard deviation of calculating each key point region description vector obtain region description, and concrete mode is to remember key point P in G ₁, P ₂... P _ndefinite region description vector is respectively V ₁, V ₂..., V _n, by V ₁, V ₂..., V _nthe average of each component forms a vector and is normalized the 8 dimension mean vector V that obtain G _m=[v _m1, v _m2... v _m8]/|| v _m1, v _m2..., v _m8||, wherein || || represent vectorial modulo operation; By V ₁, V ₂..., V _nthe standard deviation of each component forms a vector and is normalized the 8 dimension standard deviation vector V that obtain region G _s=[v _s1, v _s2..., v _s8]/|| v _s1, v _s2..., v _s8||, wherein by mean vector V _mwith standard deviation vector V _sform a vector and be normalized the sub-D=[V of 16 dimension region description that can obtain region G _m, V _s]/|| [V _m, V _s] ||;

Step S8: utilize region description to carry out Region Matching, concrete mode is to remember region G to be matched in the 1st width image ₁₁, G ₁₂..., G _1mregion description be respectively D ₁₁, D ₁₂..., D _1m, the region G to be matched in the 2nd width image ₂₁, G ₂₂..., G _2ndescriptor be respectively D ₂₁, D ₂₂..., D _2n, for D ₁₁, D ₁₂..., D _1min the sub-D of arbitrary region description _1i, find D ₁₁, D ₂₂..., D _2nin with D _1ithe sub-D of region description of Euclidean distance minimum _2jif, D _1ialso be D simultaneously ₁₁, D ₁₂..., D _1min with D _2jregion description of Euclidean distance minimum, G _1iwith D _2jfor a pair of matching area.

Irregular area automatic matching method based on straight line dichotomy provided by the invention, mainly utilized can remain unchanged when the deformation with respect to the position relationship of straight line this rule of key point in image-region, first key point in the maximum symmetric position of definite area, the proper vector of calculating each point definite area, then utilize successively each key point and maximum symmetric position to form straight line region is divided into two sub regions and calculates description vectors, finally by calculating the statistic of each description vectors, obtain region description and mate.Because point and the relative position relation of straight line remain unchanged under image deformation, therefore utilize the definite straight line of each key point and maximum symmetric position to be divided under deformation and to there is stability region.Method provided by the invention neither needs to carry out shape matching and does not need to be again fixed big or small subregion division in structure realm descriptor process, reduce the error causing due to image deformation, therefore aspect the stability of image deformation, be better than existing method.

Claims (1)

1. a method for automatic matching of irregular regions based on straight line dichotomy in a digital image, characterized in that, comprising the steps: Step S1: Take two different images of the same scene from different angles and input them into the computer; Step S2: using the existing area detection technology to perform area detection; Step S3: Calculate the maximum symmetrical position P _C of each area G. The specific method is to take any position P in the area G as the center, draw 18 straight lines L _{i and} divide the entire circumference into 36 parts; record the straight line in G The number of pixels on both sides of L _i are respectively N _L (i), N _R (i), defined is the asymmetry at P; the position P _C with the smallest asymmetry in the region G is determined as the maximum symmetrical position of the region G; Step S4: Calculate the eigenvectors of each point in the region G. The specific method is to use the Gaussian gradient template to calculate the gradient vector of each point in the region G, and record the Gaussian gradient at the point X(x, y) in G as [d _x , d _y ], the average gradient of each point in G is [V _x , V _y ], calculate the feature vector s=[s ₁ , s ₂ ] at the point X(x, y), where s ₁ =d _x ·V _x +d _y ·V _y , s ₂ =d _x ·V _y -d _y ·V _x ; Step S5: Calculate the key points of the region by using the extreme value of the gradient magnitude. The specific method is to record the gradient magnitude at any point X(x, y) in the region G as E(x, y), and under the constraint of the threshold T, set The point with the maximum gradient amplitude in the 3×3 neighborhood is used as the key point in the region G, that is, the following conditions are met: E(x, y)>T, E(x, y)>E(x+1, y+1), E(x, y)>E(x-1, y-1), E(x,y)>E(x-1,y), E(x,y)>E(x+1,y), E(x,y)>E(x,y-1), E(x,y)>E(x,y+1), E(x,y)>E(x-1,y+1), E(x,y)>E(x+1,y-1 ); The specific determination method of the threshold T is: T=Mean(E)+k Std(E), Mean(E) and Std(E) respectively represent the mean value and standard deviation of the gradient amplitude values of each point in the region, The value range of the proportional coefficient k is 2 to 3; Step S6: Calculating the region description vector corresponding to the key points in the region G, the specific way is to note that the key points of the region G obtained in step S5 are P ₁ , P ₂ ... P _n , and the symmetry between the key point P _i and the region G The straight line determined by the position P _C divides the area G into two sub-areas, and sums the gradient amplitude values of each point in the two sub-areas, and records the sub-areas with larger and smaller gradient amplitude sums as _GB and G _S respectively ; Distinguish between positive and negative in the two sub-regions, accumulate the feature vector components of each point obtained in step S4, and obtain the 8-dimensional region description vector V _i =[v _i1 , v _i2 ,..., v corresponding to the key point P _{i i8} ], where $v_{i1}=\underset{{\mathrm{the\; s}}_1\left(x\right)>0\mathrm{andX}\&Element;G_B}{\mathrm{\&Sigma;}}{\mathrm{the\; s}}_i\left(x\right),v_{i2}=\underset{{\mathrm{the\; s}}_1\left(x\right)<0\mathrm{andX}\&Element;G_B}{\mathrm{\&Sigma;}}{\mathrm{the\; s}}_1\left(x\right),$ $v_{i3}=\underset{{\mathrm{the\; s}}_2\left(x\right)>0\mathrm{andX}\&Element;G_B}{\mathrm{\&Sigma;}}{\mathrm{the\; s}}_2\left(x\right),v_{i4}=\underset{{\mathrm{the\; s}}_2\left(x\right)<0\mathrm{andX}\&Element;G_B}{\mathrm{\&Sigma;}}{\mathrm{the\; s}}_2\left(x\right),v_{i5}=\underset{{\mathrm{the\; s}}_1\left(x\right)>0\mathrm{andX}\&Element;G_S}{\mathrm{\&Sigma;}}{\mathrm{the\; s}}_1\left(x\right),$ $v_{i6}=\underset{{\mathrm{the\; s}}_1\left(x\right)<0\mathrm{andX}\&Element;G_S}{\mathrm{\&Sigma;}}{\mathrm{the\; s}}_1\left(x\right),v_{i7}=\underset{{\mathrm{the\; s}}_2\left(x\right)>0\mathrm{andX}\&Element;G_S}{\mathrm{\&Sigma;}}{\mathrm{the\; s}}_2\left(x\right),v_{i8}=\underset{{\mathrm{the\; s}}_2\left(x\right)<0\mathrm{andX}\&Element;G_S}{\mathrm{\&Sigma;}}{\mathrm{the\; s}}_2\left(x\right);$ Step S7: Calculate the mean value and standard deviation of the region description vectors of each key point to obtain the region descriptor. The specific method is to record the region description vectors determined by the key points P ₁ , P ₂ ... P _n in G as V ₁ , V ₂ ,...,V _n , compose the mean value of each component of V ₁ , V ₂ ,...,V _n into a vector and normalize to obtain the 8-dimensional mean value vector V _M of G =[v _M1 , v _M2 ,...,v _M8 ]/||v _M1 , v _M2 ,...,v _M8 ||, where

||·|| represents a vector modulo operation; the standard deviation of each component of V ₁ , V ₂ , ..., V _n is composed into a vector and normalized to obtain the 8-dimensional standard deviation vector V _S =[ v _S1 , v _S2 , ..., v _S8 ]/||v _S1 , v _S2 , ..., v _S8 ||, where

Combining the mean vector V _M and the standard deviation vector V _S into one vector and performing normalization can obtain the 16-dimensional region descriptor D=[V _M , V _S ]/||[V _M , V _S ]|| ; Step S8: Use region descriptors to perform region matching. The specific method is to record the region descriptors of regions G ₁₁ , G ₁₂ , ..., G _1m to be matched in the first image as D ₁₁ , D ₁₂ , .. ., D _1m , the descriptors of the regions to be matched G ₂₁ , G ₂₂ , ..., G _2n in the second image are D ₂₁ , D ₂₂ , ..., D _2n , for D ₁₁ , D ₁₂ , ..., any region descriptor D _1i in D _1m , find the region descriptor D 2j with the smallest Euclidean distance to D _1i in _{D 21} , D ₂₂ , ..., D _2n , if D _1i is also D ₁₁ , D ₁₂ , ..., the region descriptors with the smallest Euclidean distance between D _1m and D _2j , then G _1i and G _2j are a pair of matching regions.

Images