CN104156696A - Bi-directional-image-based construction method for quick local changeless feature descriptor - Google Patents

Abstract

The invention provides a bi-directional-image-based construction method for a quick local changeless feature descriptor, and is applied to the technical field of image processing. The bi-directional-image-based construction method comprises the following steps: preprocessing, dividing subregions, selecting sampling points, selecting neighborhood points of sampling points, pre-performing interpolation, obtaining pixel values of the neighborhood points, calculating the principal direction interval and the secondary direction interval of each sampling point, constructing FRDOH sub descriptors, and constructing FRDOH descriptors. The bi-directional-image-based construction method can enable the FRDOH descriptor to have rotation invariability, can improve operation speed of an FRDOH descriptor, and can improve stability and the distinguishable property of the FRDOH descriptor at the same time.

Description

The building method of the quick local invariant feature descriptor based on twocouese figure

Technical field

The present invention relates to technical field of image processing, more particularly, relate to a kind of building method of image processing descriptor.

Background technology

Along with the development in intelligent epoch, the automatic identification technology based on machine vision is widely used and pays close attention to.But computing machine is far away from the height intellectuality of human brain, allow computing machine from a large amount of visual information, extract useful information and accurately identify target just become a problem that is worth research, various local invariant feature extraction algorithms provide the effective means that solves this difficult problem.Therefore, local invariant feature has obtained application in many aspects, comprises target identification, 3D reconstruction, scene classification and panorama splicing etc.Conventionally the research of local invariant feature mainly being comprised to three aspects, is respectively that feature detection, feature are described and characteristic matching.

At present existing multiple detection operator aspect feature detection, for example Harris detects operator and Hessian detects operator respectively based on Harris matrix and Hessian matrix detection angle point, apply on this basis the automatic scale selection theory that Lindeberg proposes, can obtain detecting son and Hessian-Affine detection with the Harris-Affine of affine unchangeability.Timor Kadir and Michael Brady have proposed Kadir-Brady conspicuousness in calendar year 2001 and have detected son, and extract the significant characteristics of image with this.Detect on sub basis in Kadir-Brady conspicuousness, obtained two improvement versions with affine unchangeability and robustness by follow-up continuous research.It is to find point of interest by comparing neighborhood information that FAST (Features from the Accelerated Segment Test) detects son, Matas has proposed a kind of method of surveyed area in the literature, be called maximum stable extremal region (MSERs), the method is to obtain the local feature with affine unchangeability by the regional area that in detected image, gray scale is the most stable.

Aspect feature describes, Feature Descriptor is that a kind of quantitative data of Local Structure of Image feature is described, and a desirable Feature Descriptor should possess high robustness, the property distinguished and speed, and the speed here comprises that descriptor builds speed and matching speed.Therefore the core that designs a descriptor is to construct one at the local invariant feature descriptor that has outstanding representation aspect robustness, the property distinguished and speed.

After the descriptor of acquisition feature, need suitable characteristic matching to complete object identification.Be generally first to ask for two distances (as Euclidean distance) between descriptor, then recently calculate corresponding coupling mark with bee-line or the shortest/time short distance, finally coupling mark is decided to recognition result compared with default threshold value.

What in local invariant feature descriptor, impact was the widest is SIFT descriptor.SIFT descriptor is the statistics of carrying out histogram of gradients in definite principal direction, thereby there is good stability and accuracy, especially have the occasion of change of scale, rotational transform, view transformation and affined transformation at some, the accuracy of SIFT descriptor will be much better than other descriptors at that time.But, because choosing of its principal direction is more time-consuming, be vulnerable to the impact of statistical error, and the computation process of gradient vector is more numerous and diverse, therefore SIFT descriptor is under some influence in actual applications.But because the ability of SIFT descriptor reply general pattern conversion is more comprehensive, many researchists are willing to be intended to carry out various optimization and improvement on the basis of SIFT.Such as PCA-SIFT descriptor is that principal component analysis (PCA) has been carried out comprehensively, greatly reducing the dimension of SIFT with SIFT descriptor, also improve the accuracy rate of descriptor; GLOH descriptor is to replace grid to carry out the statistics of gradient orientation histogram with polar coordinate representation, experimental results show that this method for expressing can effectively improve the robustness of descriptor; SURF descriptor has been introduced haar wavelet transform, on the basis of part advantage that ensures SIFT, realizes dimensionality reduction and has greatly improved speed; DAISY descriptor is to have utilized gradient direction and these two factors of location of pixels, has good effect in the time of affine variation and light intensity linear transformation, and still, due to the calculating of its complexity, DAISY descriptor shows generally aspect speed.

Because SIFT descriptor is to realize rotational invariance by the method for searching principal direction, and this method is easily subject to the impact of statistical error, is not invariable rotary truly, and therefore the stability of SIFT descriptor is easily affected.Other derivative descriptors of SIFT have been continued to use the principal direction finding method of SIFT descriptor, therefore also have this defect.

In addition also have two other newer descriptor, MROGH and RATMIC.MROGH descriptor has used the method that realizes rotational invariance in the time of the gradient of calculating sampling point, and making does not need the specially principal direction in calculated characteristics region in description process, and descriptor can be more stable in multiple situation like this; MROGH descriptor has increased the property distinguished of descriptor and the adaptive faculty to non-linear illumination variation simultaneously.But what the method for many support areas of having introduced MROGH descriptor improved entirety can separating capacity, the entirety of MROGH descriptor is consuming time has increased several times, so the performance of speed aspect is unsatisfactory.

Consider the demand of speed aspect, RATMIC descriptor is revised on the basis of MROGH descriptor.Because many support areas above-mentioned are more consuming time, RATMIC descriptor has been abandoned the method that this raising descriptor can separating capacity, but the method for having used another to represent sampling point value replaces the method for compute gradient value used in MROGH descriptor.This method is 4 neighborhood points of sampled point periphery to be compared and encoded become binary string, and to the representation dimension of each point, 8 dimensions from MROGH descriptor have become 15 dimensions.Thereby improve the property distinguished of descriptor.See on the whole, it is far away from the contour descriptor consuming time of MROGH that RATMIC descriptor is wanted aspect speed, but performance aspect the property distinguished is slightly worse than MROGH descriptor.In view of the foregoing, need to design a kind of speed fast, can distinguish the descriptor that performance is good.

Summary of the invention

The object of the invention is to overcome shortcoming of the prior art with not enough, a kind of building method of the quick local invariant feature descriptor based on twocouese figure is provided; This building method can make FRDOH descriptor have rotational invariance, improves the arithmetic speed of FRDOH descriptor, can improve the stability of FRDOH descriptor and the property distinguished simultaneously.

In order to achieve the above object, the present invention is achieved by following technical proposals: a kind of building method of the quick local invariant feature descriptor based on twocouese figure, it is characterized in that, and comprise the following steps:

The first step, carries out pre-service, and pre-service comprises obtains unique point P _j, j ∈ 1 ..., and KP}, wherein KP is unique point P _jsum; Obtain unique point P _jto detect realization by detection with affine unchangeability;

Second step, set up the look-up table of sub-pix pixel value in characteristic area by pre-interpolation method: characteristic area is carried out to sub-pix segmentation and form the pre-interpolation area with sub-pix point, and adopt bilinear interpolation mode to obtain the pixel value of sub-pix point, form sub-pix pixel value look-up table;

The 3rd step, all pixels in characteristic area are sorted and form pixel sequence according to gray-scale value size, then pixel sequence average is divided into k interval, different interval corresponding different subregions respectively, the pixel in each interval is mapped to corresponding subregion, thereby characteristic area is divided into k sub regions; And set the subregion mark S of each sub regions, S ∈ 0 ..., k-1};

The 4th step, sampled point C refers to and in characteristic area, removes unique point P _jany one outer point, C _i(i=1 ..., 12) be the neighborhood point of choosing centered by sampled point C, sampled point C is carried out to the sub-descriptor construction of function of FRDPH; The sub-descriptor construction of function of FRDPH of sampled point C comprises the following steps:

A walks, and chooses the neighborhood point C of sampled point C _i(i=1 ..., 12): centered by sampled point C, ray for diagonal, 2L are that catercorner length is determined square one, four summits of square one and four limit mid points are set as to neighborhood point C ₁～neighborhood point C ₈; Then determining square two using four summits of square one as four limit mid points of square two, is neighborhood point C by four vertex of square two ₉～neighborhood point C ₁₂;

B step, by the neighborhood point C on characteristic area ₁～neighborhood point C ₁₂be mapped on pre-interpolation area; Obtain successively and neighborhood point C ₁～neighborhood point C ₁₂the sub-pix point of arest neighbors, neighborhood point C ₁～neighborhood point C ₁₂pixel value equal respectively the pixel value of the sub-pix point of arest neighbors; The pixel value of the sub-pix point of arest neighbors obtains by sub-pix pixel value look-up table;

C step, sets up x-y coordinate system one along the diagonal line of square one, and neighborhood is put to C ₁～neighborhood point C ₈pixel value project in x-y coordinate system one, be m level by principal direction interval division, the interval O of calculating sampling point C principal direction ₁, O ₁∈ 0 ..., m-1};

D step, sets up x-y coordinate system two along the diagonal line of square two, and neighborhood is put to C ₉～neighborhood point C ₁₂pixel value project in x-y coordinate system two, be n level by inferior direction interval division, C Direction interval O of calculating sampling point ₂, O ₂∈ 0 ..., n-1};

E walks, and obtains the subregion mark S of sampled point C;

F walks, and sets up the sub-descriptor of FRDPH of sampled point C:

The sub-descriptor function of FRDPH of sampled point C is

$\left\{\begin{array}{c}\mathrm{\γ}\left(C\right)=O_1+O_2\×m+S\×m\×n\\ \mathrm{FRDOH}\left(C\right)=\mathrm{\φ}\left(\mathrm{\γ}\left(C\right)\right)=(0,...,0,1_{\mathrm{\γ}\left(C\right)},0,...,0)\end{array}\right.$

Wherein, γ (C) is that sampled point C is mapped to the position in the sub-descriptor of FRDPH; Length=m × n × s of FRDOH (C), and the value of γ (C) position is 1, the value of other positions is 0;

The 5th step, judge sampled point C whether travel through in characteristic area Region (j) institute a little: if so, skip to the 6th step; Otherwise, repeat the 4th step;

The 6th step, structure FRDOH descriptor:

FRDOH descriptor function is $\mathrm{subdes}\left(j\right)=\underset{C\∈\mathrm{Region}\left(j\right)}{\mathrm{\Σ}}\mathrm{FRDOH}\left(C\right),j\∈\{1,...,\mathrm{KP}\}$

Wherein, KP is unique point P _jsum, Region (j) is arbitrary unique point P _jcharacteristic area.

The benefit of building method of the present invention is: one, adopt twocouese figure, i.e. dual direction histogram, by the gradient direction of different neighborhood point set representations sampled point C, can improve the property distinguished of FRDOH descriptor, makes FRDOH descriptor have good description performance; Two, for improving the property distinguished of FRDOH descriptor, FRDOH descriptor has been chosen 12 neighborhood points at sampled point periphery and has been calculated, but neighborhood point is not generally positioned at pixel place, need to obtain pixel value by interpolation; Building method of the present invention has proposed a kind of method of pre-interpolation, first adopt bilinear interpolation mode to carry out sub-pix segmentation to characteristic area, the pixel value of neighborhood point is got the pixel value of the sub-pix point of arest neighbors afterwards, in the impregnable situation of guaranteed performance, has greatly reduced consuming time; Three, adopt gray-scale value sequence to carry out subregion division, as long as pixel keeps pixel value magnitude relationship constant in characteristic area, subregion just can remain unchanged, and therefore FRDOH descriptor has rotational invariance.

Further scheme is: in the A step of described the 4th step, described four summits of square one and four limit mid points are set as to neighborhood point C ₁～neighborhood point C ₈refer to, be positioned at ray square one summit in positive dirction is neighborhood point C ₁, in the counterclockwise direction summit and the four limit mid points of square one are set as to neighborhood point C ₁～neighborhood point C ₈; Described four vertex by square two are neighborhood point C ₉～neighborhood point C ₁₂refer to, C is put to from neighborhood in the summit of square two ₁left side starts taking counterclockwise direction setting as neighborhood point C ₉～neighborhood point C ₁₂.Building method of the present invention adopts twocouese drawing method to construct FRDOH descriptor, puts C by neighborhood ₁～neighborhood point C ₈with neighborhood point C ₉～neighborhood point C ₁₂these two different neighborhood point sets represent the gradient direction of sampled point C, to increase the feature property distinguished of FRDOH descriptor.

In the C step of described the 4th step, the described diagonal line along square one is set up x-y coordinate system one and is referred to, with CC ₁c ₇set up x-y coordinate system one, wherein ray for x axle positive dirction, ray for y axle positive dirction;

The interval O of described calculating sampling point C principal direction ₁refer to and comprise the following steps:

C1 step, coordinates computed is the pixel aggregate-value D on x axle and y axle in _x1and D (C) _y1(C):

$\left\{\begin{array}{c}{D}_{y1}\left(C\right)=I\left(C_1\right)-I\left(C_5\right)+\frac{\sqrt{2}}{2}[I\left(C_2\right)-I\left(C_4\right)-I\left(C_6\right)+I\left(C_8\right)]\\ D_{x1}\left(C\right)=I\left(C_7\right)-I\left(C_3\right)+\frac{\sqrt{2}}{2}[I\left(C_8\right)+I\left(C_6\right)-I\left(C_4\right)-I\left(C_2\right)]\end{array}\right.$

Wherein I (C _i) (i=1 ..., 8) be neighborhood point C _i(i=1 ..., 8) pixel value;

C2 walks, and obtains the main gradient direction angle θ of sampled point C ₁,

C3 step, according to main gradient direction angle θ ₁determine the interval O of corresponding principal direction ₁, O ₁∈ 0 ..., m}.

In the D step of described the 4th step, the described diagonal line along square two is set up x-y coordinate system two and is referred to, with CC ₉c ₁₂set up x-y coordinate system two, wherein ray for x axle positive dirction, ray for y axle positive dirction;

Described C Direction interval O of calculating sampling point ₂refer to and comprise the following steps:

D1 step, coordinates computed is the pixel aggregate-value D on x axle and y axle in two _x2and D (C) _y2(C):

$\left\{\begin{array}{c}{D}_{y2}\left(C\right)=I\left(C_9\right)-I\left(C_{11}\right)\\ D_{x2}\left(C\right)=I\left(C_{12}\right)-I\left(C_{10}\right)\end{array}\right.$

Wherein I (C _i) (i=9 ..., 12) be neighborhood point C _i(i=9 ..., 12) pixel value;

D2 walks, and obtains the subgradient deflection θ of sampled point C ₂,

D3 step, according to subgradient deflection θ ₂determine corresponding inferior Direction interval O ₂, O ₂∈ 0 ..., n}.

Further scheme is: after described the 6th step, also comprise the 7th step: Hailin lattice are apart from processing; Described Hailin lattice refer to each element of vector in FRDOH descriptor function subdes (i) are carried out to evolution apart from processing, form new FRDOH descriptor function subdes (j) ';

New FRDOH descriptor function subdes (j) ' is: subdes (j) '=subdes (j) ^1/2.FRDOH descriptor function subdes (i) forms histogram, and histogrammic transverse axis represents interval, and the longitudinal axis represents the number of sampled point C in corresponding interval.Building method of the present invention is introduced Hailin lattice apart from processing, and the weight of the interval that in capable of regulating histogram, sampled point C number is more and sampled point C number is less in characteristic similarity calculates further promoted the performance distinguished of FRDOH descriptor.

In the described first step, pre-service also comprises Gaussian smoothing filtering processing and characteristic area normalized.Carry out Gaussian smoothing filtering processing, can remove noise, improve FRDOH descriptor performance.

Preferred scheme is: in described the 3rd step, the division number k of subregion is 4; In the C step of described the 4th step, the division number m in principal direction interval is 8; In the D step of described the 4th step, the division number n of time Direction interval is 4.Because the inferior Direction interval of sampled point C is determined by four neighborhood points, direction interval division is 4 because of this time, can improve the stability of FRDOH descriptor.Subregion is divided into 4, and principal direction interval division is 8, can make FRDOH descriptor have good stability and the property distinguished simultaneously.

Compared with prior art, tool of the present invention has the following advantages and beneficial effect:

1, building method of the present invention adopts twocouese figure describing method, can improve the property distinguished of FRDOH descriptor, makes FRDOH descriptor have good description performance;

2, building method of the present invention has proposed pre-interpolation method, in the impregnable situation of guaranteed performance, has greatly reduced consuming time;

3, building method of the present invention adopts gray-scale value sequence to carry out subregion division, makes FRDOH descriptor have rotational invariance;

4, building method of the present invention is introduced Hailin lattice apart from processing, and the weight of the interval that in capable of regulating histogram, sampled point C number is more and sampled point C number is less in characteristic similarity calculates further promoted the performance distinguished of FRDOH descriptor.

Brief description of the drawings

Fig. 1 is the process flow diagram of the sub-building method of description of the invention;

Fig. 2 (a)～Fig. 2 (e) is the example of dividing subregion in the 3rd step of the sub-building method of description of the invention;

Fig. 3 chooses neighborhood point C during the A of the 4th step of the sub-building method of description of the invention walks _ischematic diagram;

Fig. 4 (a)～Fig. 4 (c) obtains neighborhood point C during the B of the 4th step of the sub-building method of description of the invention walks _ithe example of pixel value;

Fig. 5 is the trial curve that descriptor parameter is selected;

Fig. 6 (a)～Fig. 6 (c) carries out jpeg format to image to compress empirical curve after treatment;

Fig. 7 (a)～Fig. 7 (c) carries out the empirical curve after Fuzzy Processing to image;

Fig. 8 (a)～Fig. 8 (c) be to image be rotated with change of scale after empirical curve;

Fig. 9 (a)～Fig. 9 (c) carries out the empirical curve after view transformation to image;

Figure 10 (a)～Figure 10 (c) carries out the empirical curve after illumination conversion process to image.

Embodiment

Below in conjunction with accompanying drawing and embodiment, the present invention is described in further detail.

Embodiment mono-

The building method of the quick local invariant feature descriptor of the present embodiment based on twocouese figure, its flow process as shown in Figure 1, comprises the following steps:

The first step, carries out pre-service, and pre-service comprises obtains unique point P _j, j ∈ 1 ..., and KP}, wherein KP is unique point P _jsum; Obtain unique point P _jto detect realization by detection with affine unchangeability;

Second step, set up the look-up table of sub-pix pixel value in characteristic area by pre-interpolation method: characteristic area is carried out to sub-pix segmentation and form the pre-interpolation area with sub-pix point, and adopt bilinear interpolation mode to obtain the pixel value of sub-pix point, form sub-pix pixel value look-up table;

The 3rd step, all pixels in characteristic area are sorted and form pixel sequence according to gray-scale value size, then pixel sequence average is divided into k interval, different interval corresponding different subregions respectively, the pixel in each interval is mapped to corresponding subregion, thereby characteristic area is divided into k sub regions; And set the subregion mark S of each sub regions, S ∈ 0 ..., k-1}; For example, the characteristic area of Fig. 2 (a) is divided, subregion number k is 4, obtains the subregion as shown in Fig. 2 (b)～Fig. 2 (e);

The 4th step, sampled point C refers to and in characteristic area, removes unique point P _jany one outer point, C _i(i=1 ..., 12) be the neighborhood point of choosing centered by sampled point C, sampled point C is carried out to the sub-descriptor construction of function of FRDPH; The sub-descriptor construction of function of FRDPH of sampled point C comprises the following steps:

A walks, and chooses the neighborhood point C of sampled point C _i(i=1 ..., 12): centered by sampled point C, ray for diagonal, 2L are that catercorner length is determined square one, be positioned at ray square one summit in positive dirction is neighborhood point C ₁, in the counterclockwise direction summit and the four limit mid points of square one are set as to neighborhood point C ₁～neighborhood point C ₈; As shown in Figure 3; Then with square four summit C of one ₁, C ₃, C ₅, C ₇determine square two as four limit mid points of square two, C is put to from neighborhood in the summit of square two ₁left side starts taking counterclockwise direction setting as neighborhood point C ₉～neighborhood point C ₁₂;

B step, by the neighborhood point C on characteristic area ₁～neighborhood point C ₁₂be mapped on pre-interpolation area; Obtain successively and neighborhood point C ₁～neighborhood point C ₁₂the sub-pix point of arest neighbors, neighborhood point C ₁～neighborhood point C ₁₂pixel value equal respectively the pixel value of the sub-pix point of arest neighbors; The pixel value of the sub-pix point of arest neighbors obtains by sub-pix pixel value look-up table;

Fig. 4 (a) is characteristic area, wherein Z _ij(i=1 ..., 5; J=1 ..., 5) be the pixel in characteristic area, and C _iit is any one neighborhood point in characteristic area.Neighborhood is put to C _ibe mapped on pre-interpolation area, as shown in Fig. 4 (b).For raising speed, for neighborhood point C in pre-interpolation area _i, directly obtain more accurate pixel value by arest neighbors method of interpolation; As shown in Fig. 4 (c), Z' ₁₁, Z' ₁₂, Z' ₂₁, Z' ₂₂be in pre-interpolation area with C _ifour adjacent sub-pix points, and C _ipixel value by the sub-pix point Z' of its arest neighbors ₂₁pixel value determine;

C step, sets up x-y coordinate system one along the diagonal line of square one, specifically with CC ₁c ₇set up x-y coordinate system one, wherein ray for x axle positive dirction, ray for y axle positive dirction; Neighborhood is put to C ₁～neighborhood point C ₈pixel value project in x-y coordinate system one, be m level by principal direction interval division, the interval O of calculating sampling point C principal direction ₁, O ₁∈ 0 ..., m-1};

The interval O of calculating sampling point C principal direction ₁refer to and comprise the following steps:

C1 step, coordinates computed is the pixel aggregate-value D on x axle and y axle in _x1and D (C) _y1(C):

$\left\{\begin{array}{c}{D}_{y1}\left(C\right)=I\left(C_1\right)-I\left(C_5\right)+\frac{\sqrt{2}}{2}[I\left(C_2\right)-I\left(C_4\right)-I\left(C_6\right)+I\left(C_8\right)]\\ D_{x1}\left(C\right)=I\left(C_7\right)-I\left(C_3\right)+\frac{\sqrt{2}}{2}[I\left(C_8\right)+I\left(C_6\right)-I\left(C_4\right)-I\left(C_2\right)]\end{array}\right.$

Wherein I (C _i) (i=1 ..., 8) be neighborhood point C _i(i=1 ..., 8) pixel value;

C2 walks, and obtains the main gradient direction angle θ of sampled point C ₁,

C3 step, according to main gradient direction angle θ ₁determine the interval O of corresponding principal direction ₁, O ₁∈ 0 ..., m};

D step, sets up x-y coordinate system two along the diagonal line of square two, specifically with CC ₉c ₁₂set up x-y coordinate system two, wherein ray for x axle positive dirction, ray for y axle positive dirction; Neighborhood is put to C ₉～neighborhood point C ₁₂pixel value project in x-y coordinate system two, be n level by inferior direction interval division, C Direction interval O of calculating sampling point ₂, O ₂∈ 0 ..., n-1};

C Direction interval O of calculating sampling point ₂refer to and comprise the following steps:

D1 step, coordinates computed is the pixel aggregate-value D on x axle and y axle in two _x2and D (C) _y2(C):

$\left\{\begin{array}{c}{D}_{y2}\left(C\right)=I\left(C_9\right)-I\left(C_{11}\right)\\ D_{x2}\left(C\right)=I\left(C_{12}\right)-I\left(C_{10}\right)\end{array}\right.$

Wherein I (C _i) (i=9 ..., 12) be neighborhood point C _i(i=9 ..., 12) pixel value;

D2 walks, and obtains the subgradient deflection θ of sampled point C ₂,

D3 step, according to subgradient deflection θ ₂determine corresponding inferior Direction interval O ₂, O ₂∈ 0 ..., n};

E walks, and obtains the subregion mark S of sampled point C;

F walks, and sets up the sub-descriptor of FRDPH of sampled point C:

The sub-descriptor function of FRDPH of sampled point C is

$\left\{\begin{array}{c}\mathrm{\γ}\left(C\right)=O_1+O_2\×m+S\×m\×n\\ \mathrm{FRDOH}\left(C\right)=\mathrm{\φ}\left(\mathrm{\γ}\left(C\right)\right)=(0,...,0,1_{\mathrm{\γ}\left(C\right)},0,...,0)\end{array}\right.$

Wherein, γ (C) is that sampled point C is mapped to the position in the sub-descriptor of FRDPH; Length=m × n × s of FRDOH (C), and the value of γ (C) position is 1, the value of other positions is 0;

The 5th step, judge sampled point C whether travel through in characteristic area Region (j) institute a little: if so, skip to the 6th step; Otherwise, repeat the 4th step;

The 6th step, structure FRDOH descriptor:

FRDOH descriptor function is $\mathrm{subdes}\left(j\right)=\underset{C\∈\mathrm{Region}\left(j\right)}{\mathrm{\Σ}}\mathrm{FRDOH}\left(C\right),j\∈\{1,...,\mathrm{KP}\}$

Wherein, KP is unique point P _jsum, Region (j) is arbitrary unique point P _jcharacteristic area;

The 7th step, Hailin lattice are apart from processing; Hailin lattice refer to each element of vector in FRDOH descriptor function subdes (i) are carried out to evolution apart from processing, form new FRDOH descriptor function subdes (j) ';

New FRDOH descriptor function subdes (j) ' is: subdes (j) '=subdes (j) ^1/2.

The benefit of the present embodiment building method is: one, adopt twocouese drawing method to construct FRDOH descriptor, put C by neighborhood ₁～neighborhood point C ₈with neighborhood point C ₉～neighborhood point C ₁₂these two different neighborhood point sets represent the gradient direction of sampled point C, can improve the property distinguished of FRDOH descriptor, make FRDOH descriptor have good description performance; Two, for improving the property distinguished of FRDOH descriptor, FRDOH descriptor has been chosen 12 neighborhood points at sampled point periphery and has been calculated, but neighborhood point is not generally positioned at pixel place, need to obtain pixel value by interpolation; Building method of the present invention has proposed a kind of method of pre-interpolation, first adopt bilinear interpolation mode to carry out sub-pix segmentation to characteristic area, the pixel value of neighborhood point is got the pixel value of the sub-pix point of arest neighbors afterwards, in the impregnable situation of guaranteed performance, has greatly reduced consuming time; Three, adopt gray-scale value sequence to carry out subregion division, as long as pixel keeps pixel value magnitude relationship constant in characteristic area, subregion just can remain unchanged, and therefore FRDOH descriptor has rotational invariance.

FRDOH descriptor function subdes (i) forms histogram, and histogrammic transverse axis represents interval, and the longitudinal axis represents the number of sampled point C in corresponding interval.Building method of the present invention is introduced Hailin lattice apart from processing, and the weight of the interval that in capable of regulating histogram, sampled point C number is more and sampled point C number is less in characteristic similarity calculates further promoted the performance distinguished of FRDOH descriptor.

The division number k of subregion is 4; The division number m in principal direction interval is 8; The division number n of inferior Direction interval is 4.Because the inferior Direction interval of sampled point C is determined by four neighborhood points, direction interval division is 4 because of this time, can improve the stability of FRDOH descriptor.Subregion is divided into 4, and principal direction interval division is 8, can make FRDOH descriptor have good stability and the property distinguished simultaneously.Fig. 5 is the trial curve of selecting the division number k of multiple subregion and the division number m in principal direction interval.From Fig. 5, can obtain, when the division number k=4 of subregion, can obtain top performance; The numerical value of the division number m in principal direction interval is larger, and FRDOH descriptor performance is better, but considers dimension and the performance of descriptor, for the division number m in principal direction interval preferably adopts 8.Dimension=m × n × the s=128 of FRDOH descriptor.

In the first step, pre-service also comprises Gaussian smoothing filtering processing and characteristic area normalized.Carry out Gaussian smoothing filtering processing, can remove noise, improve FRDOH descriptor performance.

In order to completely examine the performance of FRDOH descriptor, FRDOH descriptor and RATMIC descriptor, MROGH descriptor, DAISY descriptor, SIFT descriptor, SURF descriptor and ORB descriptor are carried out to contrast test.Various descriptors carry out repeatedly feature to the image on Oxford database respectively to be described, and afterwards feature is described to result and is processed the repetition rate and the accuracy that draw various descriptors, finally repetition rate and accuracy is contrasted.Fig. 6 (a)～Fig. 6 (c) is that the jpeg format that image is carried out in various degree compresses empirical curve after treatment.Fig. 7 (a)～Fig. 7 (c) carries out the empirical curve after Fuzzy Processing in various degree to image.Fig. 8 (a)～Fig. 8 (c) carries out the empirical curve after rotation and change of scale in various degree to image.Fig. 9 (a)～Fig. 9 (c) carries out the empirical curve after view transformation in various degree to image.Figure 10 (a)～Figure 10 (c) carries out the empirical curve after illumination conversion process in various degree to image.Can be obtained by test, the performance of FRDOH descriptor is better than most of descriptor, and this has illustrated that building method of the present invention can effectively improve the accuracy rate of descriptor.FRDOH descriptor is compared with MROGH descriptor, and in most of image, performance is slightly not enough.This is because MROGH descriptor method increases many support areas in the structure of descriptor, thereby has improved the property distinguished of MROGH descriptor.But increase the performance boost that mode is brought of many support areas taking very big time loss as cost.FRDOH descriptor does not adopt the method that increases many support areas; Reducing time loss is one of technical matters of FRDOH descriptor solution.Comprehensive, FRDOH descriptor fast operation, time loss is few, has good stability and the property distinguished simultaneously.

Embodiment bis-

The building method of the quick local invariant feature descriptor of the present embodiment based on twocouese figure, comprises the following steps:

The first step, carries out pre-service, and pre-service comprises obtains unique point P _j, j ∈ 1 ..., and KP}, wherein KP is unique point P _jsum; Obtain unique point P _jto detect realization by detection with affine unchangeability;

Second step, set up the look-up table of sub-pix pixel value in characteristic area by pre-interpolation method: characteristic area is carried out to sub-pix segmentation and form the pre-interpolation area with sub-pix point, and adopt bilinear interpolation mode to obtain the pixel value of sub-pix point, form sub-pix pixel value look-up table;

The 3rd step, all pixels in characteristic area are sorted and form pixel sequence according to gray-scale value size, then pixel sequence average is divided into k interval, different interval corresponding different subregions respectively, the pixel in each interval is mapped to corresponding subregion, thereby characteristic area is divided into k sub regions; And set the subregion mark S of each sub regions, S ∈ 0 ..., k-1};

The 4th step, will remove unique point P in characteristic area _jouter all pixels are set as sampled point C, respectively each sampled point C are carried out to the sub-descriptor construction of function of FRDPH; The sub-descriptor construction of function of FRDPH of each sampled point C comprises the following steps:

A step, chooses the neighborhood point of sampled point C: centered by sampled point C, ray for diagonal, 2L are that catercorner length is determined square one, four summits of square one and four limit mid points are set as to neighborhood point C ₁～neighborhood point C ₈; Then determining square two using four summits of square one as four limit mid points of square two, is neighborhood point C by four vertex of square two ₉～neighborhood point C ₁₂;

B step, by the neighborhood point C on characteristic area ₁～neighborhood point C ₁₂be mapped on pre-interpolation area; Obtain successively and neighborhood point C ₁～neighborhood point C ₁₂the sub-pix point of arest neighbors, neighborhood point C ₁～neighborhood point C ₁₂pixel value equal respectively the pixel value of the sub-pix point of arest neighbors; The pixel value of the sub-pix point of arest neighbors obtains by sub-pix pixel value look-up table;

C step, sets up x-y coordinate system one along the diagonal line of square one, and neighborhood is put to C ₁～neighborhood point C ₈pixel value project in x-y coordinate system one, be m level by principal direction interval division, the interval O of calculating sampling point C principal direction ₁, O ₁∈ 0 ..., m-1};

D step, sets up x-y coordinate system two along the diagonal line of square two, and neighborhood is put to C ₉～neighborhood point C ₁₂pixel value project in x-y coordinate system two, be n level by inferior direction interval division, C Direction interval O of calculating sampling point ₂, O ₂∈ 0 ..., n-1};

E walks, and obtains the subregion mark S of sampled point C;

F walks, and sets up the sub-descriptor of FRDPH of sampled point C:

The sub-descriptor function of FRDPH of sampled point C is

$\left\{\begin{array}{c}\mathrm{\γ}\left(C\right)=O_1+O_2\×m+S\×m\×n\\ \mathrm{FRDOH}\left(C\right)=\mathrm{\φ}\left(\mathrm{\γ}\left(C\right)\right)=(0,...,0,1_{\mathrm{\γ}\left(C\right)},0,...,0)\end{array}\right.$

Wherein, γ (C) is that sampled point C is mapped to the position in the sub-descriptor of FRDPH; Length=m × n × s of FRDOH (C), and the value of γ (C) position is 1, the value of other positions is 0;

The 5th step, judge sampled point C whether travel through in characteristic area Region (j) institute a little: if so, skip to the 6th step; Otherwise, repeat the 4th step;

The 6th step, structure FRDOH descriptor:

FRDOH descriptor function is $\mathrm{subdes}\left(j\right)=\underset{C\∈\mathrm{Region}\left(j\right)}{\mathrm{\Σ}}\mathrm{FRDOH}\left(C\right),j\∈\{1,...,\mathrm{KP}\}$

Wherein, KP is unique point P _jsum, Region (j) is arbitrary unique point P _jcharacteristic area.

The benefit of FRDOH descriptor building method of the present invention is: one, adopt twocouese figure, it is dual direction histogram, by the gradient direction of different neighborhood point set representations sampled point C, can improve the property distinguished of FRDOH descriptor, make FRDOH descriptor there is good description performance; Two, for improving the property distinguished of FRDOH descriptor, descriptor has been chosen 12 neighborhood points at sampled point periphery and has been calculated, but neighborhood point is not generally positioned at pixel place, need to obtain pixel value by interpolation; Building method of the present invention has proposed a kind of method of pre-interpolation, first adopt bilinear interpolation mode to carry out sub-pix segmentation to characteristic area, the pixel value of neighborhood point is got the pixel value of the sub-pix point of arest neighbors afterwards, in the impregnable situation of guaranteed performance, has greatly reduced consuming time; Three, adopt gray-scale value sequence to carry out subregion division, as long as pixel keeps pixel value magnitude relationship constant in characteristic area, subregion just can remain unchanged, and therefore FRDOH descriptor has rotational invariance.

Above-described embodiment is preferably embodiment of the present invention; but embodiments of the present invention are not restricted to the described embodiments; other any do not deviate from change, the modification done under Spirit Essence of the present invention and principle, substitutes, combination, simplify; all should be equivalent substitute mode, within being included in protection scope of the present invention.

Claims (7)

1. a building method for the quick local invariant feature descriptor based on twocouese figure, is characterized in that, comprises the following steps:

The first step, carries out pre-service, and pre-service comprises obtains unique point P _j, j ∈ 1 ..., and KP}, wherein KP is unique point P _jsum; Obtain unique point P _jto detect realization by detection with affine unchangeability;

Second step, set up the look-up table of sub-pix pixel value in characteristic area by pre-interpolation method: characteristic area is carried out to sub-pix segmentation and form the pre-interpolation area with sub-pix point, and adopt bilinear interpolation mode to obtain the pixel value of sub-pix point, form the look-up table of sub-pix pixel value;

The 3rd step, all pixels in characteristic area are sorted and form pixel sequence according to gray-scale value size, then pixel sequence average is divided into k interval, different interval corresponding different subregions respectively, the pixel in each interval is mapped to corresponding subregion, thereby characteristic area is divided into k sub regions; And set the subregion mark S of each sub regions, S ∈ 0 ..., k-1};

The 4th step, sampled point C refers to and in characteristic area, removes unique point P _jany one outer point, C _i(i=1 ..., 12) be the neighborhood point of choosing centered by sampled point C, sampled point C is carried out to the sub-descriptor construction of function of FRDPH; The sub-descriptor construction of function of FRDPH of sampled point C comprises the following steps:

A walks, and chooses the neighborhood point C of sampled point C _i(i=1 ..., 12): centered by sampled point C, ray for diagonal, 2L are that catercorner length is determined square one, four summits of square one and four limit mid points are set as to neighborhood point C ₁～neighborhood point C ₈; Then determining square two using four summits of square one as four limit mid points of square two, is neighborhood point C by four vertex of square two ₉～neighborhood point C ₁₂;

B step, by the neighborhood point C on characteristic area ₁～neighborhood point C ₁₂be mapped on pre-interpolation area; Obtain successively and neighborhood point C ₁～neighborhood point C ₁₂the sub-pix point of arest neighbors, neighborhood point C ₁～neighborhood point C ₁₂pixel value equal respectively the pixel value of the sub-pix point of arest neighbors; The pixel value of the sub-pix point of arest neighbors obtains by sub-pix pixel value look-up table;

C step, sets up x-y coordinate system one along the diagonal line of square one, and neighborhood is put to C ₁～neighborhood point C ₈pixel value project in x-y coordinate system one, be m level by principal direction interval division, the interval O of calculating sampling point C principal direction ₁, O ₁∈ 0 ..., m-1};

D step, sets up x-y coordinate system two along the diagonal line of square two, and neighborhood is put to C ₉～neighborhood point C ₁₂pixel value project in x-y coordinate system two, be n level by inferior direction interval division, C Direction interval O of calculating sampling point ₂, O ₂∈ 0 ..., n-1};

E walks, and obtains the subregion mark S of sampled point C;

F walks, and sets up the sub-descriptor of FRDPH of sampled point C:

The sub-descriptor function of FRDPH of sampled point C is

$\left\{\begin{array}{c}\mathrm{\γ}\left(C\right)=O_1+O_2\×m+S\×m\×n\\ \mathrm{FRDOH}\left(C\right)=\mathrm{\φ}\left(\mathrm{\γ}\left(C\right)\right)=(0,...,0,1_{\mathrm{\γ}\left(C\right)},0,...,0)\end{array}\right.$

Wherein, γ (C) is that sampled point C is mapped to the position in the sub-descriptor of FRDPH; Length=m × n × s of FRDOH (C), and the value of γ (C) position is 1, the value of other positions is 0;

The 5th step, judge sampled point C whether travel through in characteristic area Region (j) institute a little: if so, skip to the 6th step; Otherwise, repeat the 4th step;

The 6th step, structure FRDOH descriptor:

FRDOH descriptor function is $\mathrm{subdes}\left(j\right)=\underset{C\∈\mathrm{Region}\left(j\right)}{\mathrm{\Σ}}\mathrm{FRDOH}\left(C\right),j\∈\{1,...,\mathrm{KP}\}$

Wherein, KP is unique point P _jsum, Region (j) is arbitrary unique point P _jcharacteristic area.

2. the building method of the quick local invariant feature descriptor based on twocouese figure according to claim 1, is characterized in that, in the A of described the 4th step step, described four summits of square one and four limit mid points is set as to neighborhood point C ₁～neighborhood point C ₈refer to, be positioned at ray square one summit in positive dirction is neighborhood point C ₁, in the counterclockwise direction summit and the four limit mid points of square one are set as to neighborhood point C ₁～neighborhood point C ₈; Described four vertex by square two are neighborhood point C ₉～neighborhood point C ₁₂refer to, C is put to from neighborhood in the summit of square two ₁left side starts taking counterclockwise direction setting as neighborhood point C ₉～neighborhood point C ₁₂.

3. the building method of the quick local invariant feature descriptor based on twocouese figure according to claim 2, is characterized in that, in the C step of described the 4th step, the described diagonal line along square one is set up x-y coordinate system one and referred to, with CC ₁c ₇set up x-y coordinate system one, wherein ray for x axle positive dirction, ray for y axle positive dirction;

The interval O of described calculating sampling point C principal direction ₁refer to and comprise the following steps:

C1 step, coordinates computed is the pixel aggregate-value D on x axle and y axle in _x1and D (C) _y1(C):

$\left\{\begin{array}{c}{D}_{y1}\left(C\right)=I\left(C_1\right)-I\left(C_5\right)+\frac{\sqrt{2}}{2}[I\left(C_2\right)-I\left(C_4\right)-I\left(C_6\right)+I\left(C_8\right)]\\ D_{x1}\left(C\right)=I\left(C_7\right)-I\left(C_3\right)+\frac{\sqrt{2}}{2}[I\left(C_8\right)+I\left(C_6\right)-I\left(C_4\right)-I\left(C_2\right)]\end{array}\right.$

Wherein I (C _i) (i=1 ..., 8) be neighborhood point C _i(i=1 ..., 8) pixel value;

C2 walks, and obtains the main gradient direction angle θ of sampled point C ₁,

C3 step, according to main gradient direction angle θ ₁determine the interval O of corresponding principal direction ₁, O ₁∈ 0 ..., m}.

4. the building method of the quick local invariant feature descriptor based on twocouese figure according to claim 2, is characterized in that, in the D step of described the 4th step, the described diagonal line along square two is set up x-y coordinate system two and referred to, with CC ₉c ₁₂set up x-y coordinate system two, wherein ray for x axle positive dirction, ray for y axle positive dirction;

Described C Direction interval O of calculating sampling point ₂refer to and comprise the following steps:

D1 step, coordinates computed is the pixel aggregate-value D on x axle and y axle in two _x2and D (C) _y2(C):

$\left\{\begin{array}{c}{D}_{y2}\left(C\right)=I\left(C_9\right)-I\left(C_{11}\right)\\ D_{x2}\left(C\right)=I\left(C_{12}\right)-I\left(C_{10}\right)\end{array}\right.$

Wherein I (C _i) (i=9 ..., 12) be neighborhood point C _i(i=9 ..., 12) pixel value;

D2 walks, and obtains the subgradient deflection θ of sampled point C ₂,

D3 step, according to subgradient deflection θ ₂determine corresponding inferior Direction interval O ₂, O ₂∈ 0 ..., n}.

5. according to the building method of the quick local invariant feature descriptor based on twocouese figure described in any one in claim 1 to 4, it is characterized in that, after described the 6th step, also comprise the 7th step: Hailin lattice are apart from processing; Described Hailin lattice refer to each element of vector in FRDOH descriptor function subdes (i) are carried out to evolution apart from processing, form new FRDOH descriptor function subdes (j) ';

New FRDOH descriptor function subdes (j) ' is: subdes (j) '=subdes (j) ^1/2.

6. according to the building method of the quick local invariant feature descriptor based on twocouese figure described in any one in claim 1 to 4, it is characterized in that, in the described first step, pre-service also comprises Gaussian smoothing filtering processing and characteristic area normalized.

7. according to the building method of the quick local invariant feature descriptor based on twocouese figure described in any one in claim 1 to 4, it is characterized in that, in described the 3rd step, the division number k of subregion is 4; In the C step of described the 4th step, the division number m in principal direction interval is 8; In the D step of described the 4th step, the division number n of time Direction interval is 4.