CN101697232B - SIFT characteristic reducing method facing close repeated image matching - Google Patents

Abstract

The invention discloses an SIFT characteristic reducing method facing close repeated image matching, comprising the following steps of carrying out gaussian kernel convolution processing on each image of an image library to obtain image key points; carrying out gaussian uniformization on the contrast ratio and the principal curvature ratio of the key points and linear weighting to obtain significance degree; sequencing from small to large according to the significance degree of the key points, and selecting a number of the key points specified by uses to realize reduction; generating descriptors for the reduced key points according to the positions, the dimensions and the directional information of the reduced key points to obtain SIFT characteristics; establishing an image library index for all SIFT characteristic sets by a partial sensitivity hash technique, and providing query function of the close repeated image matching. The invention utilizes the research and the realization results of an image partial characteristic technique and the partial sensitivity hash technique, can conveniently and fast provide the query ability of the close image matching and enable the users to regulate the weighting coefficient of a reduction algorithm and the upper limit threshold value of the number of the SIFT characteristics according to application requirements so as to provide the best property.

Description

SIFT characteristic cutting method towards nearly multiimage coupling

Technical field

The present invention relates to technical field of image processing, particularly relate to a kind of SIFT characteristic cutting method towards nearly multiimage coupling.

Background technology

Nearly multiimage coupling is the major issue that is present in a plurality of applications such as image copyright monitoring, image retrieval.Different fields is different for the definition of nearly multiimage.The nearly multiimage that this patent is directed against is by generations such as the rotation of source images process, convergent-divergent, shearing, brightness and contrast's variation, color adjustment.Along with the birth of various image editing tools, this also becomes the main source of nearly multiimage on the internet.

In recent years, the SIFT technology has obtained many achievements in nearly multiimage coupling field, but still exists some problems.The SIFT technology is set forth in object identification field the earliest, is characterized in keeping certain unchangeability to three-dimensional view transformation, illumination projection or even object distortion, and object identification is needed just for these characteristics.Compare with object identification, nearly multiimage coupling is more paid attention to rotation, convergent-divergent, affined transformation, brightness and contrast's variation etc. are maintained the invariance, and it is much little that its difficulty is wanted relatively, and often the subclass of image SIFT characteristic just can reach higher matching accuracy.On the other hand, the SIFT key point number of image is very big usually, and is difficult to accurate control.If the key point number is excessive, not only calculate big leap ahead of time of key point descriptor, the efficient of inquiry also can reduce greatly, is difficult to the stable of the system that guarantees.Therefore, how effectively strategic point quantity is screened the most effective crucial point set, when not losing matching accuracy, improves the performance of system, is a problem demanding prompt solution.

But, in images match research before, often ignored the importance that key point is reduced, normally rule of thumb, adjust contrast lower limit and principal curvatures, roughly the quantity of strategic point and the unsettled key point of filtration than upper limit threshold; Perhaps, obtain relatively effectively key point of controlled quantity just simply through ordering to the key point contrast.These are too simply handled, and the key point matching capacity that possibly cause screening is not strong, and the accuracy of system is caused unnecessary loss.

Summary of the invention

The object of the present invention is to provide a kind of SIFT characteristic cutting method towards nearly multiimage coupling; Contrast and principal curvatures through to the SIFT key point are weighed its matching capacity than weighting; Thereby realize reducing, can solve the problem that exists in the background effectively.

The technical scheme that the present invention solves its technical matters employing is:

1) each width of cloth image in the image library is carried out the gaussian kernel process of convolution, use the difference of gaussian operator to detect extreme point, be called key point at the image multiscale space that obtains;

2) the key point contrast and the key point principal curvatures ratio that image are extracted carry out Gaussian normalization respectively;

3) linear weighted function of contrast after the employing Gaussian normalization and principal curvatures ratio is weighed the matching capacity of key point, is called significance;

4) set of keypoints that step 3) is obtained sorts according to the significance of key point from small to large, chooses the key point that the user specifies number, and realizes reducing;

5) to the key point after the step 4) reduction, generate descriptor, obtain the SIFT characteristic according to its position, yardstick and directional information;

6) all the SIFT characteristic sets through reducing that the image library all images extracted use the local sensitivity salted hash Salted to set up the image library index, and nearly multiimage matching inquiry function is provided.

The local feature of key point presentation video in the said step 1), piece image comprises a plurality of key points.

Said step 2) the Gaussian normalization method of selecting for use in guarantees that the span of the ordered series of numbers element after the normalization is [0,1], and this normalization formula is:

${\hat{R}}_i=(\frac{R_i-\stackrel{\‾}{R}}{\sqrt{\underset{i=0}{\overset{n-1}{\mathrm{\Σ}}}{(R_i-\stackrel{\‾}{R})}^2}}+1)/2$

${\hat{C}}_i^{\′}=(\frac{C_i^{\′}-{\stackrel{\‾}{C}}^{\′}}{\sqrt{\underset{i=0}{\overset{n-1}{\mathrm{\Σ}}}{(C_i^{\′}-{\stackrel{\‾}{C}}^{\′})}^2}}+1)/2$

R wherein _iBe the principal curvatures ratio of certain key point,

Be R _iValue after the normalization, R is all R _iAverage, C _iBe the contrast of certain key point, Be C _iValue after the normalization, C is all C _iAverage.

Significance in the said step 3) is designated as Sig (X _i), the matching capacity of the more little expression key point of its value is strong more, and its computing formula is:

$\mathrm{Sig}\left(X_i\right)=\mathrm{\α}\×{\hat{C}}_i^{\′}+(1-\mathrm{\α})\×{\hat{R}}_i$

Weighting coefficient α ∈ [0 wherein; 1], the weight of expression

.

Said step 5) generates descriptor according to image local zone, key point place, and each SIFT characteristic is one 128 dimensional vector.

Index establishment step in the said step 6) is following:

1) sets up the LSH index with the SIFT unit of being characterized as, and be stored in the disk;

2) according to the relation of image and SIFT characteristic, set up inverted index from the SIFT characteristic to image.

Nearly multiimage matching inquiry functional steps in the step 6) is following:

1) use the SIFT characteristic to reduce the SIFT characteristic that algorithm extracts user's input picture;

2) through the LSH search index SIFT feature set close with each SIFT characteristic of input picture, and they merging;

3), find out step 2 according to inverted index) the pairing image of SIFT characteristic in the result set, thus obtain the set of diagrams picture;

4) in the filter result image, with the SIFT number of features of user's input picture coupling image less than assign thresholds;

5), use the random sampling unification algorism further to verify to the filter result of step 4).

The beneficial effect that the present invention has is:

Make full use of the existing research of image local feature technology and local sensitivity salted hash Salted and realized achievement; Nearly multiimage query capability can conveniently be provided; The user is according to the weighting coefficient and the SIFT number of features upper limit threshold of application demand adjustment reduction algorithm, with the performance that offers the best.

Description of drawings

Accompanying drawing is based on the nearly multiimage matching system principle of work synoptic diagram of SIFT characteristic cutting method.

Embodiment

Combine accompanying drawing and embodiment that the present invention is described further below at present.

Shown in accompanying drawing, practical implementation process of the present invention and principle of work are following:

1) each width of cloth image in the image library is carried out the gaussian kernel process of convolution, use the difference of gaussian operator to detect extreme point, be called key point at the image multiscale space that obtains;

2) the key point contrast and the key point principal curvatures ratio that image are extracted carry out Gaussian normalization respectively;

3) linear weighted function of contrast after the employing Gaussian normalization and principal curvatures ratio is weighed the matching capacity of key point, is called significance;

4) set of keypoints that step 3) is obtained sorts according to the significance of key point from small to large, chooses the key point that the user specifies number, and realizes reducing;

5) to the key point after the step 4) reduction, generate descriptor, obtain the SIFT characteristic according to its position, yardstick and directional information;

6) all the SIFT characteristic sets through reducing that the image library all images extracted use the local sensitivity salted hash Salted to set up the image library index, and nearly multiimage matching inquiry function is provided.

The local feature of key point presentation video in the step 1), piece image comprises a plurality of key points.

In practical implementation, the Gaussian convolution method is:

One width of cloth input picture, (x, y), the metric space under different scale representes that (x, y σ), can obtain through the gaussian kernel convolution, that is: to be designated as L to be designated as I

L(x，y，σ)＝G(x，y，σ)*I(x，y)

Wherein two-dimensional Gaussian function G (x, y, σ) definition as follows:

$G(x,y,\mathrm{\σ})=\frac{1}{{2\mathrm{\π\σ}}^2}{e}^{-(x^2+y^2)/2{\mathrm{\σ}}^2}$

Wherein (x, the y) location of pixels of presentation video, σ are the metric space factor, also are the variances of Gauss normal distribution.

Step 2) the Gaussian normalization method of selecting for use in can guarantee that the span of the ordered series of numbers element after the normalization is [0,1], and this normalization formula is:

${\hat{R}}_i=(\frac{R_i-\stackrel{\‾}{R}}{\sqrt{\underset{i=0}{\overset{n-1}{\mathrm{\Σ}}}{(R_i-\stackrel{\‾}{R})}^2}}+1)/2$

${\hat{C}}_i^{\′}=(\frac{C_i^{\′}-{\stackrel{\‾}{C}}^{\′}}{\sqrt{\underset{i=0}{\overset{n-1}{\mathrm{\Σ}}}{(C_i^{\′}-{\stackrel{\‾}{C}}^{\′})}^2}}+1)/2$

R wherein _iBe the principal curvatures ratio of certain key point,

Be R _iValue after the normalization, R is all R _iAverage, C _iBe the contrast of certain key point,

Be C _iValue after the normalization, C is all C _iAverage.

Significance in the step 3) is designated as Sig (X _i), the matching capacity of the more little expression key point of its value is strong more, and its computing formula is:

$\mathrm{Sig}\left(X_i\right)=\mathrm{\α}\×{\hat{C}}_i^{\′}+(1-\mathrm{\α})\×{\hat{R}}_i$

Weighting coefficient α ∈ [0 wherein; 1], the weight of expression

.

In practical implementation, optimal effectiveness can be obtained usually in α=0.6.

Step 5) generates descriptor according to image local zone, key point place, and each SIFT characteristic is one 128 dimensional vector.

In practical implementation, can adopt original SIFT key point descriptor, as:

Coordinate axis is rotated to be the direction of key point, to guarantee rotational invariance; To each key point use 4 * 4 totally 16 seed points describe, just can produce 128 data for a key point like this, promptly finally form the 128 SIFT proper vectors tieed up; The influence that this moment, the SIFT proper vector was removed geometry deformation factors such as dimensional variation, rotation continues the length normalization method with proper vector again, then can further remove the influence of illumination variation.

Image index establishment step in the step 6) is following:

1) sets up the LSH index with the SIFT unit of being characterized as, and be stored in the disk;

2) according to the relation of image and SIFT characteristic, set up inverted index from the SIFT characteristic to image.

In practical implementation, can select different LSH technology to realize according to application demand:

1) if the high accuracy of application requirements can be selected E2LSH;

2) if low storage space of application requirements and low memory consumption can be selected MP-LSH.

Nearly multiimage matching inquiry step in the step 6) is following:

1) use the SIFT characteristic to reduce the SIFT characteristic that algorithm extracts user's input picture;

2) through the LSH search index SIFT feature set close with each SIFT characteristic of input picture, and they merging;

3), find out step 2 according to inverted index) the pairing image of SIFT characteristic in the result set, thus obtain the set of diagrams picture;

4) in the filter result image, with the SIFT number of features of user's input picture coupling image less than assign thresholds;

5), use the random sampling unification algorism further to verify to the filter result of step 4).

Claims (5)

1. SIFT characteristic cutting method towards nearly multiimage coupling is characterized in that the step of this method is following:

1) each width of cloth image in the image library is carried out the gaussian kernel process of convolution, use the difference of gaussian operator to detect extreme point, be called key point at the image multiscale space that obtains;

2) the key point contrast and the key point principal curvatures ratio that image are extracted carry out Gaussian normalization respectively;

3) linear weighted function of contrast after the employing Gaussian normalization and principal curvatures ratio is weighed the matching capacity of key point, is called significance;

4) set of keypoints that step 3) is obtained sorts according to the significance of key point from small to large, chooses the key point that the user specifies number, and realizes reducing;

5) to the key point after the step 4) reduction, generate descriptor, obtain the SIFT characteristic according to its position, yardstick and directional information;

6) all the SIFT characteristic sets through reducing that the image library all images extracted use the local sensitivity salted hash Salted to set up the image library index, and nearly multiimage matching inquiry function is provided.

2. a kind of SIFT characteristic cutting method towards nearly multiimage coupling according to claim 1, it is characterized in that: the local feature of key point presentation video in the said step 1), piece image comprises a plurality of key points.

3. a kind of SIFT characteristic cutting method towards nearly multiimage coupling according to claim 1 is characterized in that: said step 5) generates descriptor according to image local zone, key point place, and each SIFT characteristic is one 128 dimensional vector.

4. a kind of SIFT characteristic cutting method according to claim 1 towards nearly multiimage coupling, it is characterized in that: the index establishment step in the said step 6) is following:

1) sets up the LSH index with the SIFT unit of being characterized as, and be stored in the disk;

2) according to the relation of image and SIFT characteristic, set up inverted index from the SIFT characteristic to image.

5. a kind of SIFT characteristic cutting method according to claim 1 towards nearly multiimage coupling, it is characterized in that: the nearly multiimage matching inquiry functional steps in the said step 6) is following:

1) use the SIFT characteristic to reduce the SIFT characteristic that algorithm extracts user's input picture;

2) through the LSH search index SIFT feature set close with each SIFT characteristic of input picture, and they merging;

3), find out step 2 according to inverted index) the pairing image of SIFT characteristic in the result set, thus obtain the set of diagrams picture;

4) in the filter result image, with the SIFT number of features of user's input picture coupling image less than assign thresholds;

5), use the random sampling unification algorism further to verify to the filter result of step 4).