CN102521593A - Affine invariant feature extraction method based on profile split - Google Patents

Abstract

The invention discloses an affine invariant feature extraction method based on profile split, belonging to the technical field of computer vision. The affine invariant feature extraction method comprises the following steps of: extracting a shape profile and carrying out equal-area sampling on the shape profile; taking a profile starting point as a segmentation starting point, carrying out equal-point-number segmentation on the profile to form K parts and obtaining K segmentation points; connecting the segmentation starting point and all the segmentation points, leading the shape to be divided into K regions, calculating the area of each region to form an ordered series till all the profile points are traversed in sequence, and obtaining an area ordered series corresponding to each profile point to form a matrix, wherein the number of lines is K and the number of rows is the number of the profile points; using shape-area normalization for the matrix to obtain a shape signing matrix in which each line is a discrete function; and applying discrete Fourier transformation, using amplitudes of transformation coefficients as a result, selecting former 20 transformation coefficients in each line, and finally forming a matrix with the number of lines being K and the number of rows being 20, i.e., a shape feature matrix.

Description

The affine invariant features method for distilling of cutting apart based on profile

Technical field

The present invention relates to a kind of affine invariant features method for distilling of cutting apart based on profile of being used for, belong to technical field of computer vision.

Background technology

Modern society's network has become the instrument that people use always, and still along with the continuous increase of quantity of information, how in vast as the open sea network information ocean, to search out required information is the huge problem that the modern network retrieval technique is faced.Comparatively ripe at present gopher all is search key with the text, but when carrying out picture search, uses text to have bigger subjectivity and be difficult to as term image is described comprehensively, directly retrieves so hope to hold through image energy.In fact, the CBIR instrument is also very immature now, and accuracy rate is lower, and main cause is that computer vision and people's intuition impression has very big gap.And in the feature extracting method of picture shape,, obtained now using widely the most near people's intuition impression based on the described target signature of the feature extracting method of shape profile.Fourier descriptor has good performance in the feature extracting method based on the shape profile; These class methods of experiment proof have unchangeability to shape translation, yardstick, rotational transform; And computation complexity is lower, so the feature extraction algorithm based on Fourier transform has obtained using widely in reality.Affined transformation is a kind of common conversion of shape, but for general fourier descriptor affined transformation is not had robustness, has reduced the accuracy of method in image retrieval is used like this.

Summary of the invention

To the deficiency of prior art, pin of the present invention provides a kind of affine invariant features method for distilling of cutting apart based on profile.

A kind of affine invariant features method for distilling of cutting apart based on profile may further comprise the steps:

The first step is cut apart the target object shape, extracts the profile of shape, and the shape profile is carried out the homalographic sampling, representes the shape profile with N point.

In second step, arbitrarily a bit as starting point, with the definite order of counter clockwise direction numbering, wherein this starting point is a point 1 to have a few on the shape profile on the selection profile, and following point is followed successively by 2,3 ... N.

In the 3rd step, profile etc. counted is divided into the K part.On profile, select point 1 for cutting apart starting point; And remember that this point is a cut-point 1; Counterclockwise profile etc. is counted from the beginning of this point and to cut apart, wherein be designated as cut-point 2 apart from cut-point 1 nearest cut-point, and the like to the last a cut-point order be cut-point K.Straight line connection source and each cut-point obtain the line between starting point and each cut-point; Computed segmentation starting point and cut-point 2 lines and silhouette edge are along the area in the zone that shape profile between the area that surrounds the zone, two adjacent connection lines and its corresponding cut-point is surrounded and cut apart starting point and the line of K cut-point and point-to-point transmission shape profile surround regional area successively; Area by above order computation becomes an ordered sequence.

The 4th step, on profile, select point 2 as cutting apart starting point, repeat the computation process in the 3rd step, obtain the area value ordered sequence of point 2 correspondences; Point all on profile are all calculated one time, obtain the corresponding area value ordered sequence of all point.

In the 5th step, the area value ordered sequence that all point are corresponding is formed a matrix, each row of matrix corresponding the area value ordered sequence of a point, this matrix is carried out normalization, the matrix after normalization is called shape signature matrix.

In the 6th step, each line number value of shape signature matrix all is a discrete function; Each discrete function is done discrete Fourier transformation, and the Fourier coefficient that obtains is formed a matrix, i.e. shape facility matrix.

Among the present invention, parameter K, promptly the number of shape cut zone is a comparatively crucial value.Contrasted the performance situation in the different K values image retrieval among Fig. 7, following table is feature extraction required time and the retrieval accuracy table of K when getting different value.Balance movement speed and retrieval accuracy that can be best in the time of can finding out K=12.

The present invention is in the universal property that not only has fourier descriptor; Promptly have translation, yardstick, rotational invariance; And the shape facility that is extracted has stronger robustness for the affined transformation of shape, and recall ratio and the precision ratio aspect retrieval also is higher than general fourier descriptor simultaneously.

Description of drawings

Fig. 1 is the aircraft image

Fig. 2 is that the shape profile extracts synoptic diagram.

Fig. 3 is a shape point homalographic sampling synoptic diagram.

Fig. 4 a is that starting point is cut apart and cut apart to profile and the cut-point line calculates synoptic diagram at the shape internal area.

Fig. 4 b is that starting point is cut apart and cut apart to profile and the cut-point line does not calculate synoptic diagram at the shape internal area.

Fig. 5 a is the apple original image.

Fig. 5 b is an original image shape signature matrix table diagrammatic sketch.

Fig. 5 c is an original image shape facility matrix table diagrammatic sketch.

Fig. 6 a is an apple affine graph picture.

Fig. 6 b is affine picture shape signature matrix table diagrammatic sketch.

Fig. 6 c is that affine graph is as shape facility matrix table diagrammatic sketch.

Fig. 7 is the performance that feature extracting method of the present invention is retrieved in MPEG-7Shape CE-1 image library under the different parameters K.

Fig. 8 is different fourier descriptors and the inventive method are retrieved performance in MPEG-7Shape CE-1 image library contrast.

Fig. 9 is the framework synoptic diagram that the present invention is based on the affine invariant features extraction algorithm that profile cuts apart.

Figure 10 is the overall flow figure that the present invention is based on the affine invariant features extraction algorithm that profile cuts apart.

Embodiment

Embodiment:

A kind of affine invariant features method for distilling of cutting apart based on profile may further comprise the steps:

The first step is cut apart the target object shape, extracts the profile of shape, and the shape profile is carried out the homalographic sampling, representes the shape profile with N point.

In second step, arbitrarily a bit as starting point, with the definite order of counter clockwise direction numbering, wherein this starting point is a point 1 to have a few on the shape profile on the selection profile, and following point is followed successively by 2,3 ... N.

In the 3rd step, profile etc. counted is divided into the K part.On profile, select point 1 for cutting apart starting point; And remember that this point is a cut-point 1; Counterclockwise profile etc. is counted from the beginning of this point and to cut apart, wherein be designated as cut-point 2 apart from cut-point 1 nearest cut-point, and the like to the last a cut-point order be cut-point K.Straight line connection source and each cut-point obtain the line between starting point and each cut-point; Computed segmentation starting point and cut-point 2 lines and silhouette edge are along the area in the zone that shape profile between the area that surrounds the zone, two adjacent connection lines and its corresponding cut-point is surrounded and cut apart starting point and the line of K cut-point and point-to-point transmission shape profile surround regional area successively; Area by above order computation becomes an ordered sequence.

The 4th step, on profile, select point 2 as cutting apart starting point, repeat the computation process in the 3rd step, obtain the area value ordered sequence of point 2 correspondences; Point all on profile are all calculated one time, obtain the corresponding area value ordered sequence of all point.

In the 5th step, the area value ordered sequence that all point are corresponding is formed a matrix, each row of matrix corresponding the area value ordered sequence of a point, this matrix is carried out normalization, the matrix after normalization is called shape signature matrix.

In the 6th step, each line number value of shape signature matrix all is a discrete function; Each discrete function is done discrete Fourier transformation, and the Fourier coefficient that obtains is formed a matrix, i.e. shape facility matrix.

Use the present invention below piece image is carried out Shape Feature Extraction, performing step is following:

1) image binaryzation is formed bianry image, make that target object partly is 1, all the other background parts are 0.Use the Canny operator to binary Images Processing, extract the closed contour of target object.Like Fig. 2 is the closed contour of shape.

2) select on the profile arbitrfary point as starting point expression with

, from this point beginning counterclockwise successively the mark point for wherein M be that point is total.The arbitrfary point

(i=1,2 ..., coordinate figure M) is (x _i, y _i), x wherein _iRepresent the columns of this pixel in image array, y _iThe expression line number.Profile can use orderly some set representations: $C_o=\{P_1^o,P_2^o,...,P_M^o\}.$

3) the center G of calculating shape carries out homalographic to the shape profile and resamples, and supposes that sampling number is N.Point ordered set after the sampling that obtains so is C={P ₁, P ₂..., P _N.Calculate the total area A of shape.Homalographic sampling promptly is to make consecutive point P on the profile _i, P _I+1Triangle P with center G composition _iGP _I+1Area be A/N, i=1 wherein, 2 ..., N.Because profile is closed then P _i=P _I+NFig. 3 is the shape profile diagram through the homalographic sampling.

4) select profile starting point P ₁As cutting apart starting point, profile waited to count cut apart, suppose and need profile be divided into the K part; Profile between the then adjacent cut-point is counted and is N/K; And on the profile except that cutting apart starting point, also have K-1 cut-point, comprise and cut apart starting point; Is that starting point makes up an orderly point set counterclockwise with the whole cut-point on the profile to cut apart starting point, and establishing this point set is C _s={ X ₁, X ₂..., X _K, X wherein ₁For cutting apart starting point P ₁, all the other points are the cut-point on the profile.Respectively with X ₁With all the other cut-point (X ₂To X _K) connect.Shape is divided into K zone by these lines so, and wherein first zone is by line segment X ₁X ₂With an X ₁With X ₂Between the shape profile surrounded, its area is designated as S ₁(P ₁), be by line segment X then for i zone (7＜i＜K, i ∈ N) ₁X _iWith line segment X ₁X _I+1And some X _iWith an X _I+1Between the shape profile surrounded, its area is designated as S _i(P ₁), and K zone is by line segment X ₁X _KAnd some X ₁With an X _KBetween the shape profile surrounded, its area is designated as S _K(P ₁).For each zone all is closed with X ₁Point is for starting point, regional form an orderly point set a little by counterclockwise forming, and the terminal point of this point set is X ₁The orderly point set of supposing the regional i of expression is SP={D ₁, D ₂..., D _L, wherein L representes the number of a centrostigma.The form that is expressed as coordinate is SP={ (x ₁, y ₁), (x ₂, y ₂) ..., (x _L, y _L), x wherein _L=x ₁, y _L=y ₁Can use the area S of formula

zoning i.In Fig. 3 a, represented as line segment X ₁X _iComputed segmentation region area synoptic diagram when shape is inner has been represented among Fig. 3 b as line segment X fully ₁X _iIncomplete computed segmentation region area synoptic diagram when shape is inner.So for a P ₁As cutting apart starting point, can obtain the area in K zone.By this K of counterclockwise series arrangement zone, the area of Zone Full is formed ordered sequence a: A ₁={ S ₁(P ₁), S ₂(P ₁) ..., S _K(P ₁).

5) next one point P on the selected shape profile ₂As cutting apart starting point, the process of repeating step 4 obtains a P ₂Corresponding area ordered sequence: A ₂={ S ₁(P ₂), S ₂(P ₂) ..., S _K(P ₂).Continue the point on the selected shape profile successively afterwards, as a Pi (i=3,4 ..., N) corresponding area ordered sequence is: A _i={ S ₁(P _i), S ₂(P _i) ..., S _K(P _i).On the shape profile have a few and all calculated its corresponding area ordered sequence.

6) all area ordered sequences are formed a matrix according to the order of point, the corresponding profile of this matrix column number count the number of the corresponding cut zone of line number.Represent this matrix with M:

$M=\left(\begin{array}{cccc}{S}_1\left(P_1\right)& S_1\left(P_2\right)& L& S_1\left(P_N\right)\\ S_2\left(P_1\right)& S_2\left(P_2\right)& L& S_2\left(P_N\right)\\ M& M& O& M\\ S_K\left(P_1\right)& S_K\left(P_2\right)& L& S_K\left(P_N\right)\end{array}\right)$

The characteristic of matrix M reaction shape, but this character representation does not have the yardstick unchangeability, and responsive to the shape rotational transform, and data scale is bigger.

7) in order to make the described shape facility of matrix M have the yardstick unchangeability; Use the shape total area that calculates in the step 3 to carry out normalization to matrix M; And because the point of shape disperses, put us for each and use its order as parametric representation, the matrix that obtains after the normalization is M _n:

$M_n=\frac{1}{A}\left(\begin{array}{cccc}{S}_1\left(1\right)& S_1\left(2\right)& L& S_1\left(N\right)\\ S_2\left(1\right)& S_2\left(2\right)& L& S_2\left(N\right)\\ M& M& O& M\\ S_K\left(1\right)& S_K\left(2\right)& L& S_K\left(N\right)\end{array}\right)=\left(\begin{array}{cccc}{S}_1^n\left(1\right)& S_1^n\left(2\right)& L& S_1^n\left(N\right)\\ S_2^n\left(1\right)& S_2^n\left(2\right)& L& S_2^n\left(N\right)\\ M& M& O& M\\ S_K^n\left(1\right)& S_K^n\left(2\right)& L& S_K^n\left(N\right)\end{array}\right)$

Through normalized matrix M _nThe shape signature matrix that is called shape.

8) for the characteristic that makes extraction has rotational invariance, and the data scale of reduction eigenmatrix, each row of matrix is used discrete Fourier transformation.For matrix M _nThe capable vector of t can use discrete function

T=1 wherein, 2 ..., K.f _tDiscrete Fourier transformation be defined as

T=1 wherein, 2 ..., K, n=0,1,2 ..., N-1.Fourier transform coefficient f then _t(n) then as the characteristic of shape,, ignore the phase information of Fourier transform coefficient, only choose its amplitude information in order to realize rotational invariance | F _t(n) |.And because f _tIt is real-number function; So its Fourier transform result is the Fourier coefficient that can only choose half of symmetry; In order to guarantee that carrying out the later shape facility of Fourier transform still has the yardstick unchangeability, use DC component promptly simultaneously to all Fourier transform coefficients | F _t(0) | normalization, the sequence that finally obtains the Fourier coefficient composition does ${\mathrm{\Φ}}_t=(\frac{\left|F_t\left(1\right)\right|}{\left|F_t\left(0\right)\right|},\frac{\left|F_t\left(2\right)\right|}{\left|F_t\left(0\right)\right|},\frac{\left|F_t\left(3\right)\right|}{\left|F_t\left(0\right)\right|},\·\·\·,\frac{\left|F_t(N/2)\right|}{\left|F_t\left(0\right)\right|}),$ So it is last that eigenmatrix is M _f:

$M_f=\left(\begin{array}{cccc}\left|F_1\left(1\right)\right|/\left|F_1\left(0\right)\right|& \left|F_1\left(2\right)\right|/\left|F_1\left(0\right)\right|& L& \left|F_1\left(\frac{N}{2}\right)\right|/\left|F_1\left(0\right)\right|\\ \left|F_2\left(1\right)\right|/\left|F_2\left(0\right)\right|& \left|F_2\left(2\right)\right|/\left|F_2\left(0\right)\right|& L& \left|F_2\left(\frac{N}{2}\right)\right|/{|F}_2\left(0\right)|\\ M& M& O& M\\ \left|F_K\left(1\right)\right|/\left|F_K\left(0\right)\right|& \left|F_K\left(2\right)\right|/F_K\left(0\right)|& L& \left|F_K\left(\frac{N}{2}\right)\right|/\left|F_K\left(0\right)\right|\end{array}\right)=\left(\begin{array}{cccc}{\mathrm{\Ψ}}_1\left(1\right)& {\mathrm{\Ψ}}_1\left(2\right)& L& {\mathrm{\Ψ}}_1(N/2)\\ {\mathrm{\Ψ}}_2\left(1\right)& {\mathrm{\Ψ}}_2\left(2\right)& L& {\mathrm{\Ψ}}_2(N/2)\\ M& M& O& M\\ {\mathrm{\Ψ}}_K\left(1\right)& {\mathrm{\Ψ}}_K\left(2\right)& L& {\mathrm{\Ψ}}_K(N/2)\end{array}\right)$

In practical application, do not need whole Fourier transform coefficients just can well describe the characteristic of shape, through experimental verification, only need to use preceding 20 Fourier transform coefficients just can realize complete description to shape facility.So common eigenmatrix M _fBe expressed as:

$M_f=\left(\begin{array}{cccc}{\mathrm{\Ψ}}_1\left(1\right)& {\mathrm{\Ψ}}_1\left(2\right)& L& {\mathrm{\Ψ}}_1\left(20\right)\\ {\mathrm{\Ψ}}_2\left(1\right)& {\mathrm{\Ψ}}_2\left(2\right)& L& {\mathrm{\Ψ}}_2\left(20\right)\\ M& M& O& M\\ {\mathrm{\Ψ}}_K\left(1\right)& {\mathrm{\Ψ}}_K\left(2\right)& L& {\mathrm{\Ψ}}_K\left(20\right)\end{array}\right)$

Claims (1)

1. affine invariant features method for distilling of cutting apart based on profile is characterized in that method for distilling may further comprise the steps:

The first step is cut apart the target object shape, extracts the profile of shape, and the shape profile is carried out the homalographic sampling, representes the shape profile with N point;

In second step, arbitrarily a bit as starting point, with the definite order of counter clockwise direction numbering, wherein this starting point is a point 1 to have a few on the shape profile on the selection profile, and following point is followed successively by 2,3 ... N;

The 3rd step; Profile etc. counted be divided into the K part; On profile, select point 1 for cutting apart starting point, and note this be cut-point 1, begin counterclockwise profile etc. counted from this point and cut apart; Wherein be designated as cut-point 2 apart from cut-point 1 nearest cut-point, and the like to the last a cut-point order be cut-point K; Straight line connection source and each cut-point obtain the line between starting point and each cut-point; Computed segmentation starting point and cut-point 2 lines and silhouette edge are along the area in the zone that shape profile between the area that surrounds the zone, two adjacent connection lines and its corresponding cut-point is surrounded and cut apart starting point and the line of K cut-point and point-to-point transmission shape profile surround regional area successively; Area by above order computation becomes an ordered sequence;

The 4th step, on profile, select point 2 as cutting apart starting point, repeat the computation process in the 3rd step, obtain the area value ordered sequence of point 2 correspondences; Point all on profile are all calculated one time, obtain the corresponding area value ordered sequence of all point;

In the 5th step, the area value ordered sequence that all point are corresponding is formed a matrix, each row of matrix corresponding the area value ordered sequence of a point, this matrix is carried out normalization, the matrix after normalization is called shape signature matrix;

In the 6th step, each line number value of shape signature matrix all is a discrete function; Each discrete function is done discrete Fourier transformation, and the Fourier coefficient that obtains is formed a matrix, i.e. shape facility matrix.

Images