CN104156696B - Bi-directional-image-based construction method for quick local changeless feature descriptor - Google Patents
Bi-directional-image-based construction method for quick local changeless feature descriptor Download PDFInfo
- Publication number
- CN104156696B CN104156696B CN201410353900.4A CN201410353900A CN104156696B CN 104156696 B CN104156696 B CN 104156696B CN 201410353900 A CN201410353900 A CN 201410353900A CN 104156696 B CN104156696 B CN 104156696B
- Authority
- CN
- China
- Prior art keywords
- point
- frdoh
- neighborhood
- sub
- neighborhood point
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Landscapes
- Image Analysis (AREA)
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
Technical field
The present invention relates to technical field of image processing, more particularly, it relates to a kind of construction side of image procossing description
Method.
Background technology
With the development of intellectualization times, it is widely used and closes based on the automatic identification technology of machine vision
Note.But computer is intelligent far away from the height of human brain, allows computer that useful information is extracted from a large amount of visual informations simultaneously
Accurately identify target and become a problem for being worth research, and various local invariant feature extraction algorithms are then provided and solved
The effective means of this difficult problem.Therefore, local invariant feature is applied in many aspects, including target recognition, 3D reconstruction,
Scene classification and panoramic mosaic etc..Generally three aspects are mainly included to the research of local invariant feature, be respectively feature detection,
Feature description and characteristic matching.
Have various detective operators at present in terms of feature detection, such as Harris detective operators and Hessian detections are calculated
Son is based respectively on Harris matrixes and Hessian matrixes detection angle point, the automatic chi for proposing using Lindeberg on this basis
Degree theory of selection, can obtain the detections of the Harris-Affine with affine-invariant features and Hessian-Affine detection.
Timor Kadir and Michael Brady proposes Kadir-Brady significances detection in calendar year 2001, and is extracted with this
The significant characteristics of image.On the basis of Kadir-Brady significances detection, had by follow-up constantly research
There are two modified versions of affine-invariant features and robustness.FAST(Features from the Accelerated Segment
Test) detect that son is then that point of interest is found by comparing neighborhood information, and Matas then proposes in the literature a kind of detection zone
The method in domain, referred to as maximum stable extremal region (MSERs), the method is by the most stable of partial zones of gray scale in detection image
Domain is obtaining the local feature with affine-invariant features.
In terms of feature description, Feature Descriptor is that a kind of quantitative data of Local Structure of Image feature is described, one
Preferably Feature Descriptor should possess high robustness, ga s safety degree and speed, and speed here includes that description builds speed
Degree and matching speed.Therefore the core of one description of design is construction one in terms of robustness, ga s safety degree and speed
There is local invariant feature description of outstanding representation.
Appropriate characteristic matching is then needed to complete object identification after description for obtaining feature.Usually first ask for two
The distance between description (such as Euclidean distance), then short with beeline or most/time short distance ratio is matched accordingly to calculate
Fraction, finally by matching fraction compared with default threshold value determining recognition result.
It is SIFT description to affect in local invariant feature description most wide.SIFT description son be it is determined that principal direction
On carry out the statistics of histogram of gradients, thus with good stability and accuracy, especially at some there is yardstick to become
Change, rotation transformation, the occasion of view transformation and affine transformation, the accuracy of SIFT description will be much better than other descriptions at that time
Son.But, because the selection of its principal direction is relatively time-consuming, it is vulnerable to the impact of statistical error, and the calculating process of gradient vector
Comparison is numerous and diverse, therefore SIFT description are under some influence in actual applications.But due to SIFT description son reply general patterns
The ability of conversion is more comprehensive, and many research worker are all ready to carry out various optimizations and improvement on the basis of SIFT.Such as
It is that PCA is described into son with SIFT to have carried out synthesis that PCA-SIFT describes son, greatly reduces the dimension of SIFT,
Improve the accuracy rate of description；GLOH description are then to replace grid with polar coordinate representation to carry out gradient orientation histogram
Statistics, experiment proves that this method for expressing can effectively improve the robustness of description；SURF describes son and introduces Haar wavelet transform change
Change, realize dimensionality reduction on the basis of the part advantage for ensureing SIFT and greatly improve speed；DAISY descriptions is then to utilize
Gradient direction and location of pixels the two factors, in affine change and light intensity linear transformation with good effect, but
Due to the calculating of its complexity, DAISY description then show general in terms of speed.
It is to realize rotational invariance by finding the method for principal direction because SIFT describes son, and this method is easily subject to
The impact of statistical error, the stability of not invariable rotary truly, therefore SIFT description is easily affected.
Other derivative description of SIFT have continued to use the principal direction finding method of SIFT description, therefore there is also this defect.
In addition with two other newer description, MROGH and RATMIC.MROGH descriptions is calculating sampled point
Gradient when used the method for realizing rotational invariance so that the main formula of characteristic area need not be specially calculated during description
To so describing son can more stablize in several cases；Simultaneously MROGH descriptions increased the ga s safety degree of description and right
The adaptability of non-linear illumination variation.But MROGH description introduce the method for many support areas improve it is overall can
Separating capacity, the entirety of MROGH description is time-consuming to increased several times, so the performance in terms of speed is unsatisfactory.
Demand in terms of in view of speed, RATMIC descriptions is modified on the basis of MROGH description.Due to
Many support areas above-mentioned than relatively time-consuming, RATMIC description abandon it is this improve description can separating capacity side
Method, but represent the method for sampling point value come the side of calculating Grad used in replacing MROGH descriptions sub with another
Method.This method is that 4 neighborhood points of sampled point periphery are compared and are encoded to become binary string, i.e., to table that each is put
Show that 8 dimensions that dimension is described in son from MROGH become 15 dimensions.So as to improve the ga s safety degree of description.See on the whole, RATMIC
Description will be slightly worse than in terms of speed far faster than contour time-consuming description of MROGH, but the performance in terms of ga s safety degree
MROGH description.In view of the foregoing, need to design description that a kind of speed is fast, ga s safety degree energy is good.
The content of the invention
It is an object of the invention to overcome shortcoming of the prior art with deficiency, there is provided a kind of based on the quick of twocouese figure
The building method of local invariant feature description；The building method can make FRDOH description have rotational invariance, improve
The arithmetic speed of FRDOH description, while the stability and ga s safety degree of FRDOH description can be improved.
In order to achieve the above object, the technical scheme is that：It is a kind of based on the fast of twocouese figure
The building method of fast local invariant feature description, it is characterised in that comprise the following steps：
The first step, carries out pretreatment, and pretreatment includes obtaining characteristic point Pj, j ∈ { 1 ..., KP }, wherein KP is characterized a little
PjSum；Obtain characteristic point PjIt is that detection realization is carried out by detection with affine-invariant features；
Second step, by pre- interpolation method the look-up table of sub-pix pixel value in characteristic area is set up：Characteristic area is entered
Row sub-pix segments to form the pre- interpolation area with sub-pix point, and obtains the picture of sub-pix point using bilinear interpolation mode
Element value, forms sub-pix pixel value look-up table；
3rd step, all pixels point in characteristic area is ranked up to form pixel point sequence according to gray value size, so
Afterwards by pixel sequence average be divided into k it is interval, different intervals correspond to respectively different subregions, and each interval pixel reflects
Corresponding subregion is mapped to, so as to characteristic area is divided into k sub-regions；And set the sub-district field mark S of each sub-regions,
S∈{0,...,k-1}；
4th step, sampled point C is referred to and remove in characteristic area characteristic point PjAny one outer point, Ci(i=1 ..., 12)
It is then the neighborhood point chosen centered on sampled point C, FRDOH description subfunction constructions is carried out to sampled point C；Sampled point C's
FRDOH description subfunction constructions are comprised the following steps：
A is walked, and chooses the neighborhood point C of sampled point Ci(i=1 ..., 12)：Centered on sampled point C, rayFor diagonal
Line direction, 2L are that catercorner length determines square one, and four summits and four side midpoints of square one are set as into neighborhood point
C1～neighborhood point C8；Then using square one four summits as square two four side midpoints come determine square two, will just
Square two four vertex are neighborhood point C9～neighborhood point C12；
B is walked, by the neighborhood point C on characteristic area1～neighborhood point C12It is mapped on pre- interpolation area；Obtain successively and neighborhood
Point C1～neighborhood point C12The sub-pix point of arest neighbors, neighborhood point C1～neighborhood point C12Pixel value be respectively equal to the Asia picture of arest neighbors
The pixel value of vegetarian refreshments；The pixel value of the sub-pix point of arest neighbors is obtained by sub-pix pixel value look-up table；
C is walked, and the diagonal along square one sets up x-y coordinate system one, by neighborhood point C1～neighborhood point C8Pixel value projection
It is m levels by principal direction interval division in x-y coordinate system one, calculates sampled point C principal directions interval O1, O1∈{0,...,m-1}；
D is walked, and the diagonal along square two sets up x-y coordinate system two, by neighborhood point C9～neighborhood point C12Pixel value throw
Secondary direction interval division is n levels in x-y coordinate system two by shadow, calculates C Direction interval O of sampled point2, O2∈{0,...,n-
1}；
E is walked, and obtains the sub-district field mark S of sampled point C；
F is walked, and sets up FRDOH description of sampled point C：
FRDOH of sampled point C describe subfunction and are
Wherein, γ (C) is the position that sampled point C is mapped in FRDOH description；Length=m × the n of FRDOH (C) ×
Value at s, and γ (C) position is 1, and the value of other positions is 0；
5th step, judges whether sampled point C travels through the institute in characteristic area Region (j) a little：If so, the 6th is then skipped to
Step；Otherwise, the 4th step is repeated；
6th step, construction FRDOH description：
FRDOH describes subfunction
Wherein, KP is characterized point PjSum, Region (j) be any feature point PjCharacteristic area.
Building method of the present invention is advantageous in that：First, using twocouese figure, i.e., dual direction histogram, by different adjacent
The gradient direction of domain point set representations sampled point C, can improve the ga s safety degree of FRDOH description, have FRDOH description good
Description performance；2nd, the ga s safety degree to improve FRDOH descriptions, FRDOH description have chosen 12 neighbours in sampled point periphery
Domain point is calculated, but neighborhood point is general not to be located at pixel, needs to obtain pixel value by interpolation；Present invention construction
Method proposes a kind of method of pre- interpolation, and sub-pix subdivision is carried out to characteristic area initially with bilinear interpolation mode, it
Afterwards the pixel value of neighborhood point takes the pixel value of the sub-pix point of arest neighbors, in the case where ensureing that performance is unaffected, greatly drops
It is low time-consuming；3rd, sub-zone dividing is carried out using gray value sequence, as long as pixel keeps pixel value size to close in characteristic area
It is constant, subregion can keep constant, therefore FRDOH description have rotational invariance.
Further scheme is：In the A steps of the 4th step, described four summits and four side midpoints by square one
It is set as neighborhood point C1～neighborhood point C8Refer to, positioned at rayThe summit of square one in positive direction is neighborhood point C1, along inverse
The summit and four side midpoints of square one are set as neighborhood point C by clockwise1～neighborhood point C8；It is described by square two
Four vertex are neighborhood point C9～neighborhood point C12Refer to, by the summit of square two from neighborhood point C1Left side started with the inverse time
The direction setting of pin is neighborhood point C9～neighborhood point C12.Building method of the present invention is using twocouese drawing method to FRDOH description
Constructed, by neighborhood point C1～neighborhood point C8With neighborhood point C9～neighborhood point C12The two different neighborhood point sets are representing
The gradient direction of sampled point C, to increase the feature ga s safety degree of FRDOH description.
In the C steps of the 4th step, the diagonal of described edge square one is set up x-y coordinate system one and is referred to, with CC1C7
Set up x-y coordinate system one, wherein rayFor x-axis positive direction, rayFor y-axis positive direction；
Described calculating sampled point C principal directions interval O1Refer to following steps：
C1 is walked, the pixel aggregate-value D in coordinates computed system one in x-axis and y-axisx1And D (C)y1(C)：
Wherein I (Ci) (i=1 ..., it is 8) neighborhood point Ci(i=1 ..., pixel value 8)；
C2 is walked, and obtains the main gradient direction angle θ of sampled point C1,
C3 is walked, according to main gradient direction angle θ1Determine corresponding principal direction interval O1,O1∈{0,...,m-1}。
In the D steps of the 4th step, the diagonal of described edge square two is set up x-y coordinate system two and is referred to, with CC9C12
Set up x-y coordinate system two, wherein rayFor x-axis positive direction, rayFor y-axis positive direction；
Described C Direction interval O of calculating sampled point2Refer to following steps：
D1 is walked, the pixel aggregate-value D in coordinates computed system two in x-axis and y-axisx2And D (C)y2(C)：
Wherein I (Ci) (i=9 ..., it is 12) neighborhood point Ci(i=9 ..., pixel value 12)；
D2 is walked, and obtains the subgradient deflection θ of sampled point C2,
D3 is walked, according to subgradient deflection θ2Determine corresponding Direction interval O2,O2∈{0,...,n-1}。
Further scheme is：Also include the 7th step after the 6th step：Hailin lattice distance is processed；The Hailin
Lattice distance is processed and refers to that FRDOH is described into each element vectorial in subfunction subdes (i) carries out evolution, forms new
FRDOH describes subfunction subdes (j) '；
New FRDOH describes subfunction subdes (j) '：Subdes (j) '=subdes (j)1/2.FRDOH describes sub- letter
Number subdes (i) forms rectangular histogram, and histogrammic transverse axis represents interval, and the longitudinal axis represents the number of sampled point C in correspondence interval.This
Invention building method introduces Hailin lattice distance and processes, and can adjust sampled point C numbers in rectangular histogram more less with sampled point C numbers
Interval characteristic similarity calculating in weight, further improve FRDOH description son ga s safety degree energy.
Pretreatment in the first step also includes that Gaussian smoothing filter is processed and characteristic area normalized.Carry out Gauss
The disposal of gentle filter, can remove noise, improves FRDOH and describes sub- performance.
Preferred scheme is：Division number k of subregion is 4 in 3rd step；Principal direction in the C steps of the 4th step
Interval division number m is 8；Division number n of time Direction interval is 4 in the D steps of the 4th step.Due to sampled point C time
Direction interval is determined by four neighborhood points, because this direction interval division is 4, can improve the stability of FRDOH description.Son
Region division is 4, and principal direction interval division is 8, and FRDOH descriptions can be made to have good stability simultaneously and can distinguish
Property.
Compared with prior art, the invention has the advantages that and beneficial effect：
1st, building method of the present invention describes method using twocouese figure, can improve the ga s safety degree of FRDOH description, makes
FRDOH description have good description performance；
2nd, building method of the present invention proposes pre- interpolation method, in the case where ensureing that performance is unaffected, greatly reduces
It is time-consuming；
3rd, building method of the present invention carries out sub-zone dividing using gray value sequence, makes FRDOH description have rotation not
Degeneration；
4th, building method of the present invention introduces Hailin lattice distance and processes, and sampled point C numbers are more in adjustable rectangular histogram and adopt
Weight of the less interval of sampling point C numbers in characteristic similarity calculating, further improves the ga s safety degree of FRDOH description
Energy.
Description of the drawings
Fig. 1 is the flow chart of the sub- building method of description of the invention；
Fig. 2 (a)～Fig. 2 (e) be the sub- building method of description of the invention the 3rd step in divide subregion example；
Fig. 3 is to choose neighborhood point C during the A of the 4th step of the sub- building method of description of the invention is walkediSchematic diagram；
Fig. 4 (a)～Fig. 4 (c) is to obtain neighborhood point C during the B of the 4th step of the sub- building method of description of the invention is walkediPixel
The example of value；
Fig. 5 is to describe the trial curve that subparameter is selected；
Fig. 6 (a)～Fig. 6 (c) is that the empirical curve after jpeg format compression process is carried out to image；
Fig. 7 (a)～Fig. 7 (c) is that the empirical curve after Fuzzy Processing is carried out to image；
Fig. 8 (a)～Fig. 8 (c) be image is rotated and change of scale after empirical curve；
Fig. 9 (a)～Fig. 9 (c) is that the empirical curve after view transformation is carried out to image；
Figure 10 (a)～Figure 10 (c) is that the empirical curve after light change process is carried out to image.
Specific embodiment
Below in conjunction with the accompanying drawings the present invention is described in further detail with specific embodiment.
Embodiment one
The building method of quick local invariant feature description of the present embodiment based on twocouese figure, its flow process such as Fig. 1 institutes
Show, comprise the following steps：
The first step, carries out pretreatment, and pretreatment includes obtaining characteristic point Pj, j ∈ { 1 ..., KP }, wherein KP is characterized a little
PjSum；Obtain characteristic point PjIt is that detection realization is carried out by detection with affine-invariant features；
Second step, by pre- interpolation method the look-up table of sub-pix pixel value in characteristic area is set up：Characteristic area is entered
Row sub-pix segments to form the pre- interpolation area with sub-pix point, and obtains the picture of sub-pix point using bilinear interpolation mode
Element value, forms sub-pix pixel value look-up table；
3rd step, all pixels point in characteristic area is ranked up to form pixel point sequence according to gray value size, so
Afterwards by pixel sequence average be divided into k it is interval, different intervals correspond to respectively different subregions, and each interval pixel reflects
Corresponding subregion is mapped to, so as to characteristic area is divided into k sub-regions；And set the sub-district field mark S of each sub-regions,
S∈{0,...,k-1}；For example, the characteristic area of Fig. 2 (a) is divided, subregion number k be 4, obtain as Fig. 2 (b)～
Subregion shown in Fig. 2 (e)；
4th step, sampled point C is referred to and remove in characteristic area characteristic point PjAny one outer point, Ci(i=1 ..., 12)
It is then the neighborhood point chosen centered on sampled point C, FRDOH description subfunction constructions is carried out to sampled point C；Sampled point C's
FRDOH description subfunction constructions are comprised the following steps：
A is walked, and chooses the neighborhood point C of sampled point Ci(i=1 ..., 12)：Centered on sampled point C, rayFor diagonal
Line direction, 2L are that catercorner length determines square one, positioned at rayThe summit of square one in positive direction is neighborhood point
C1, in the counterclockwise direction the summit and four side midpoints of square one are set as into neighborhood point C1～neighborhood point C8；As shown in Figure 3；So
Afterwards with four summit C of square one1、C3、C5、C7Square two is determined as four side midpoints of square two, by square
Two summit is from neighborhood point C1Left side starts with direction setting counterclockwise as neighborhood point C9～neighborhood point C12；
B is walked, by the neighborhood point C on characteristic area1～neighborhood point C12It is mapped on pre- interpolation area；Obtain successively and neighborhood
Point C1～neighborhood point C12The sub-pix point of arest neighbors, neighborhood point C1～neighborhood point C12Pixel value be respectively equal to the Asia picture of arest neighbors
The pixel value of vegetarian refreshments；The pixel value of the sub-pix point of arest neighbors is obtained by sub-pix pixel value look-up table；
Fig. 4 (a) is characteristic area, wherein Zij(i=1 ..., 5；J=1 ..., 5) pixel in region is characterized, and
CiIt is any one neighborhood point in characteristic area.By neighborhood point CiIt is mapped on pre- interpolation area, such as shown in Fig. 4 (b).In order to carry
At high speed, for neighborhood point C in pre- interpolation areai, directly more accurate pixel value is obtained by arest neighbors interpolation method；As schemed
Shown in 4 (c), Z '11, Z '12, Z'21, Z'22Be in pre- interpolation area with CiFour adjacent sub-pix points, and CiPixel value then
By the sub-pix point Z' of its arest neighbors21Pixel value determine；
C is walked, and the diagonal along square one sets up x-y coordinate system one, specifically with CC1C7X-y coordinate system one is set up, its
Middle rayFor x-axis positive direction, rayFor y-axis positive direction；By neighborhood point C1～neighborhood point C8Pixel value project to x-
It is m levels by principal direction interval division in y coordinate system one, calculates sampled point C principal directions interval O1, O1∈{0,...,m-1}；
Calculate sampled point C principal directions interval O1Refer to following steps：
C1 is walked, the pixel aggregate-value D in coordinates computed system one in x-axis and y-axisx1And D (C)y1(C)：
Wherein I (Ci) (i=1 ..., it is 8) neighborhood point Ci(i=1 ..., pixel value 8)；
C2 is walked, and obtains the main gradient direction angle θ of sampled point C1,
C3 is walked, according to main gradient direction angle θ1Determine corresponding principal direction interval O1,O1∈{0,...,m-1}；
D is walked, and the diagonal along square two sets up x-y coordinate system two, specifically with CC9C12X-y coordinate system two is set up, its
Middle rayFor x-axis positive direction, rayFor y-axis positive direction；By neighborhood point C9～neighborhood point C12Pixel value project to
It is n levels by secondary direction interval division in x-y coordinate system two, calculates C Direction interval O of sampled point2, O2∈{0,...,n-1}；
Calculate C Direction interval O of sampled point2Refer to following steps：
D1 is walked, the pixel aggregate-value D in coordinates computed system two in x-axis and y-axisx2And D (C)y2(C)：
Wherein I (Ci) (i=9 ..., it is 12) neighborhood point Ci(i=9 ..., pixel value 12)；
D2 is walked, and obtains the subgradient deflection θ of sampled point C2,
D3 is walked, according to subgradient deflection θ2Determine corresponding Direction interval O2,O2∈{0,...,n-1}；
E is walked, and obtains the sub-district field mark S of sampled point C；
F is walked, and sets up FRDOH description of sampled point C：
FRDOH of sampled point C describe subfunction and are
Wherein, γ (C) is the position that sampled point C is mapped in FRDOH description；Length=m × the n of FRDOH (C) ×
Value at s, and γ (C) position is 1, and the value of other positions is 0；
5th step, judges whether sampled point C travels through the institute in characteristic area Region (j) a little：If so, the 6th is then skipped to
Step；Otherwise, the 4th step is repeated；
6th step, construction FRDOH description：
FRDOH describes subfunction
Wherein, KP is characterized point PjSum, Region (j) be any feature point PjCharacteristic area；
7th step, Hailin lattice distance is processed；Hailin lattice distance process is referred to and FRDOH is described in subfunction subdes (i)
Each element of vector carries out evolution, forms new FRDOH and describes subfunction subdes (j) '；
New FRDOH describes subfunction subdes (j) '：Subdes (j) '=subdes (j)1/2。
The present embodiment building method is advantageous in that：First, describe son to FRDOH using twocouese drawing method to construct,
By neighborhood point C1～neighborhood point C8With neighborhood point C9～neighborhood point C12The two different neighborhood point sets are representing sampled point C's
Gradient direction, can improve the ga s safety degree of FRDOH description, make FRDOH description have good description performance；2nd, to carry
The ga s safety degree of high FRDOH descriptions, FRDOH description have chosen 12 neighborhood points and calculated in sampled point periphery, but
Neighborhood point is general not to be located at pixel, needs to obtain pixel value by interpolation；Building method of the present invention proposes a kind of pre-
The method of interpolation, sub-pix subdivision is carried out initially with bilinear interpolation mode to characteristic area, afterwards the pixel value of neighborhood point
The pixel value of the sub-pix point of arest neighbors is taken, in the case where ensureing that performance is unaffected, is greatly reduced time-consuming；3rd, adopt
Gray value sequence carries out sub-zone dividing, as long as pixel holding pixel value magnitude relationship is constant in characteristic area, subregion is just
Can keep constant, therefore FRDOH description have rotational invariance.
FRDOH describes subfunction subdes (i) and forms rectangular histogram, and histogrammic transverse axis represents interval, and the longitudinal axis represents correspondence
The number of sampled point C in interval.Building method of the present invention introduces Hailin lattice distance and processes, and can adjust sampled point C in rectangular histogram
The weight that number is more and the less interval of sampled point C numbers is in characteristic similarity calculating, further improves FRDOH description
Ga s safety degree energy.
Division number k of subregion is 4；Interval division number m of principal direction is 8；Division number n of secondary Direction interval is
4.Because the secondary Direction interval of sampled point C is determined by four neighborhood points, because this direction interval division is 4, FRDOH can be improved
The stability of description.Sub-zone dividing is 4, and principal direction interval division is 8, FRDOH descriptions can be made to have simultaneously good
Stability well and ga s safety degree.Fig. 5 is from interval division number m of division number k and principal direction of many seed regions
Trial curve.Can obtain from Fig. 5, during division number k=4 of subregion top performance can be obtained；Principal direction interval division
The numerical value of number m is bigger, and FRDOH describes that sub- performance is better, but considers the dimension and performance of description, in order to principal direction it is interval
Division number m preferably adopt 8.Dimension=m × n × the s=128 of FRDOH description.
Pretreatment in the first step also includes that Gaussian smoothing filter is processed and characteristic area normalized.Carry out Gaussian smoothing
Filtering Processing, can remove noise, improves FRDOH and describes sub- performance.
In order to completely examine the performance of FRDOH description, by FRDOH descriptions and RATMIC description, MROGH descriptions
Son, DAISY description, SIFT description, SURF descriptions and ORB describe son and carry out contrast test.Various description are right respectively
Image on Oxford data bases carries out multiple feature description, afterwards characterization results is carried out into process and draws various description
Repetitive rate and accuracy, finally repetitive rate and accuracy are contrasted.Fig. 6 (a)～Fig. 6 (c) is to carry out difference to image
Empirical curve after the jpeg format compression process of degree.Fig. 7 (a)～Fig. 7 (c) is that different degrees of fuzzy place is carried out to image
Empirical curve after reason.Fig. 8 (a)～Fig. 8 (c) is different degrees of rotation to be carried out to image and the experiment after change of scale is bent
Line.Fig. 9 (a)～Fig. 9 (c) is that the empirical curve after different degrees of view transformation is carried out to image.Figure 10 (a)～Figure 10 (c)
It is that the empirical curve after different degrees of light change is processed is carried out to image.Can be obtained by test, the performance of FRDOH description is excellent
In most of description, this illustrates that building method of the present invention can effectively improve the accuracy rate of description.FRDOH description son with
MROGH describes son and compares, and the performance in most of image slightly has deficiency.This is because MROGH describes submethod in description
Increase many support areas in construction, so as to improve the ga s safety degree of MROGH description.But increase the mode institute of many support areas
Performance boost is brought to be with very big time loss as cost.FRDOH descriptions is not using the side for increasing many support areas
Method；It is that FRDOH describes one of subsolution technical problem certainly to reduce time loss.For synthesis, FRDOH describes sub- arithmetic speed
Hurry up, time loss is few, while having good stability and ga s safety degree.
Embodiment two
The building method of quick local invariant feature description of the present embodiment based on twocouese figure, comprises the following steps：
The first step, carries out pretreatment, and pretreatment includes obtaining characteristic point Pj, j ∈ { 1 ..., KP }, wherein KP is characterized a little
PjSum；Obtain characteristic point PjIt is that detection realization is carried out by detection with affine-invariant features；
Second step, by pre- interpolation method the look-up table of sub-pix pixel value in characteristic area is set up：Characteristic area is entered
Row sub-pix segments to form the pre- interpolation area with sub-pix point, and obtains the picture of sub-pix point using bilinear interpolation mode
Element value, forms sub-pix pixel value look-up table；
3rd step, all pixels point in characteristic area is ranked up to form pixel point sequence according to gray value size, so
Afterwards by pixel sequence average be divided into k it is interval, different intervals correspond to respectively different subregions, and each interval pixel reflects
Corresponding subregion is mapped to, so as to characteristic area is divided into k sub-regions；And set the sub-district field mark S of each sub-regions,
S∈{0,...,k-1}；
4th step, will remove characteristic point P in characteristic areajOuter all pixels point is set as sampled point C, respectively to each
Individual sampled point C carries out FRDOH description subfunction constructions；Each sampled point C FRDOH description subfunction construction include with
Lower step：
A is walked, and chooses the neighborhood point of sampled point C：Centered on sampled point C, rayIt is diagonal for diagonal, 2L
Line length determines square one, and four summits and four side midpoints of square one are set as into neighborhood point C1～neighborhood point C8；Then
Square two is determined using four summits of square one as four side midpoints of square two, by four summits of square two
It is set as neighborhood point C9～neighborhood point C12；
B is walked, by the neighborhood point C on characteristic area1～neighborhood point C12It is mapped on pre- interpolation area；Obtain successively and neighborhood
Point C1～neighborhood point C12The sub-pix point of arest neighbors, neighborhood point C1～neighborhood point C12Pixel value be respectively equal to the Asia picture of arest neighbors
The pixel value of vegetarian refreshments；The pixel value of the sub-pix point of arest neighbors is obtained by sub-pix pixel value look-up table；
C is walked, and the diagonal along square one sets up x-y coordinate system one, by neighborhood point C1～neighborhood point C8Pixel value projection
It is m levels by principal direction interval division in x-y coordinate system one, calculates sampled point C principal directions interval O1, O1∈{0,...,m-1}；
D is walked, and the diagonal along square two sets up x-y coordinate system two, by neighborhood point C9～neighborhood point C12Pixel value throw
Secondary direction interval division is n levels in x-y coordinate system two by shadow, calculates C Direction interval O of sampled point2, O2∈{0,...,n-
1}；
E is walked, and obtains the sub-district field mark S of sampled point C；
F is walked, and sets up FRDOH description of sampled point C：
FRDOH of sampled point C describe subfunction and are
Wherein, γ (C) is the position that sampled point C is mapped in FRDOH description；Length=m × the n of FRDOH (C) ×
Value at s, and γ (C) position is 1, and the value of other positions is 0；
5th step, judges whether sampled point C travels through the institute in characteristic area Region (j) a little：If so, the 6th is then skipped to
Step；Otherwise, the 4th step is repeated；
6th step, construction FRDOH description：
FRDOH describes subfunction
Wherein, KP is characterized point PjSum, Region (j) be any feature point PjCharacteristic area.
FRDOH of the present invention describes sub- building method and is advantageous in that：First, using twocouese figure, i.e., dual direction histogram,
By the gradient direction of different neighborhood point set representations sampled point C, the ga s safety degree of FRDOH description can be improved, describe FRDOH
Son has good description performance；2nd, the ga s safety degree to improve FRDOH descriptions, description have chosen in sampled point periphery
12 neighborhood points are calculated, but neighborhood point is general not to be located at pixel, needs to obtain pixel value by interpolation；This
Bright building method proposes a kind of method of pre- interpolation, carries out sub-pix to characteristic area initially with bilinear interpolation mode thin
Point, afterwards the pixel value of neighborhood point takes the pixel value of the sub-pix point of arest neighbors, in the case where ensureing that performance is unaffected, pole
Reduce greatly time-consuming；3rd, sub-zone dividing is carried out using gray value sequence, as long as pixel keeps pixel value big in characteristic area
Little relation is constant, and subregion can keep constant, therefore FRDOH description have rotational invariance.
Above-described embodiment is the present invention preferably embodiment, but embodiments of the present invention not by above-described embodiment
Limit, other any spirit without departing from the present invention and the change, modification, replacement made under principle, combine, simplification,
Equivalent substitute mode is should be, is included within protection scope of the present invention.
Claims (7)
1. a kind of quick local invariant feature based on twocouese figure describes the building method of son, it is characterised in that including following
Step：
The first step, carries out pretreatment, and pretreatment includes obtaining characteristic point Pj, j ∈ { 1 ..., KP }, wherein KP is characterized point Pj's
Sum；Obtain characteristic point PjIt is that detection realization is carried out by detection with affine-invariant features；
Second step, by pre- interpolation method the look-up table of sub-pix pixel value in characteristic area is set up：Asia is carried out to characteristic area
Pixel subdivision forms the pre- interpolation area with sub-pix point, and obtains the pixel of sub-pix point using bilinear interpolation mode
Value, forms the look-up table of sub-pix pixel value；
3rd step, all pixels point in characteristic area is ranked up to form pixel point sequence according to gray value size, then will
Pixel sequence average be divided into k it is interval, different intervals correspond to respectively different subregions, and each interval pixel is mapped to
Corresponding subregion, so as to characteristic area is divided into k sub-regions；And set the sub-district field mark S, S ∈ of each sub-regions
{0,...,k-1}；
4th step, sampled point C is referred to and remove in characteristic area characteristic point PjAny one outer point, Ci(i=1 ..., 12) be then
The neighborhood point chosen centered on sampled point C, FRDOH description subfunction constructions are carried out to sampled point C；The FRDOH of sampled point C
Son description subfunction construction is comprised the following steps：
A is walked, and chooses the neighborhood point C of sampled point Ci(i=1 ..., 12)：Centered on sampled point C, ray PjC is diagonal side
It is that catercorner length determines square one to, 2L, four summits and four side midpoints of square one is set as into neighborhood point C1~
Neighborhood point C8；Then using square one four summits as square two four side midpoints come determine square two, by pros
Four vertex of shape two are neighborhood point C9～neighborhood point C12；
B is walked, by the neighborhood point C on characteristic area1～neighborhood point C12It is mapped on pre- interpolation area；Obtain and neighborhood point C successively1
～neighborhood point C12The sub-pix point of arest neighbors, neighborhood point C1～neighborhood point C12Pixel value be respectively equal to the sub-pix point of arest neighbors
Pixel value；The pixel value of the sub-pix point of arest neighbors is obtained by sub-pix pixel value look-up table；
C is walked, and the diagonal along square one sets up x-y coordinate system one, by neighborhood point C1～neighborhood point C8Pixel value project to x-
It is m levels by principal direction interval division in y coordinate system one, calculates sampled point C principal directions interval O1, O1∈{0,...,m-1}；
D is walked, and the diagonal along square two sets up x-y coordinate system two, by neighborhood point C9～neighborhood point C12Pixel value project to
It is n levels by secondary direction interval division in x-y coordinate system two, calculates C Direction interval O of sampled point2, O2∈{0,...,n-1}；
E is walked, and obtains the sub-district field mark S of sampled point C；
F is walked, and sets up FRDOH description of sampled point C：
FRDOH of sampled point C describe subfunction and are
Wherein, γ (C) is the position that sampled point C is mapped in FRDOH description；Length=m × n × the s of FRDOH (C), and
Value at γ (C) position is 1, and the value of other positions is 0；
5th step, judges whether sampled point C travels through the institute in characteristic area Region (j) a little：If so, the 6th step is then skipped to；
Otherwise, the 4th step is repeated；
6th step, construction FRDOH description：
FRDOH describes subfunction
Wherein, KP is characterized point PjSum, Region (j) be any feature point PjCharacteristic area.
2. the quick local invariant feature based on twocouese figure according to claim 1 describes the building method of son, and it is special
Levy and be, in the A steps of the 4th step, described is set as neighborhood point C by four summits and four side midpoints of square one1~
Neighborhood point C8Refer to, positioned at rayThe summit of square one in positive direction is neighborhood point C1, in the counterclockwise direction by square
One summit and four side midpoints are set as neighborhood point C1～neighborhood point C8；Described four vertex by square two are neighbour
Domain point C9～neighborhood point C12Refer to, by the summit of square two from neighborhood point C1Left side starts with direction setting counterclockwise as neighbour
Domain point C9～neighborhood point C12。
3. the quick local invariant feature based on twocouese figure according to claim 2 describes the building method of son, and it is special
Levy and be, in the C steps of the 4th step, the diagonal of described edge square one is set up x-y coordinate system one and referred to, with CC1C7Build
Vertical x-y coordinate system one, wherein rayFor x-axis positive direction, rayFor y-axis positive direction；
Described calculating sampled point C principal directions interval O1Refer to following steps：
C1 is walked, the pixel aggregate-value D in coordinates computed system one in x-axis and y-axisx1And D (C)y1(C)：
Wherein I (Ci) (i=1 ..., it is 8) neighborhood point Ci(i=1 ..., pixel value 8)；
C2 is walked, and obtains the main gradient direction angle θ of sampled point C1,
C3 is walked, according to main gradient direction angle θ1Determine corresponding principal direction interval O1,O1∈{0,...,m-1}。
4. the quick local invariant feature based on twocouese figure according to claim 2 describes the building method of son, and it is special
Levy and be, in the D steps of the 4th step, the diagonal of described edge square two is set up x-y coordinate system two and referred to, with CC9C12
Set up x-y coordinate system two, wherein rayFor x-axis positive direction, rayFor y-axis positive direction；
Described C Direction interval O of calculating sampled point2Refer to following steps：
D1 is walked, the pixel aggregate-value D in coordinates computed system two in x-axis and y-axisx2And D (C)y2(C)：
Wherein I (Ci) (i=9 ..., it is 12) neighborhood point Ci(i=9 ..., pixel value 12)；
D2 is walked, and obtains the subgradient deflection θ of sampled point C2,
D3 is walked, according to subgradient deflection θ2Determine corresponding Direction interval O2,O2∈{0,...,n-1}。
5. the quick local invariant feature based on twocouese figure according to any one of claim 1 to 4 describes the structure of son
Make method, it is characterised in that also include the 7th step after the 6th step：Hailin lattice distance is processed；The Hailin lattice distance
Process is referred to carries out evolution by each element that FRDOH describes vector in subfunction subdes (i), forms new FRDOH and retouches
State subfunction subdes (j) '；
New FRDOH describes subfunction subdes (j) '：Subdes (j) '=subdes (j)1/2。
6. the quick local invariant feature based on twocouese figure according to any one of claim 1 to 4 describes the structure of son
Make method, it is characterised in that pretreatment in the first step is also included at Gaussian smoothing filter process and characteristic area normalization
Reason.
7. the quick local invariant feature based on twocouese figure according to any one of claim 1 to 4 describes the structure of son
Make method, it is characterised in that division number k of subregion is 4 in the 3rd step；The C Bu Zhong principal directions area of the 4th step
Between division number m be 8；Division number n of time Direction interval is 4 in the D steps of the 4th step.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410353900.4A CN104156696B (en) | 2014-07-23 | 2014-07-23 | Bi-directional-image-based construction method for quick local changeless feature descriptor |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410353900.4A CN104156696B (en) | 2014-07-23 | 2014-07-23 | Bi-directional-image-based construction method for quick local changeless feature descriptor |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104156696A CN104156696A (en) | 2014-11-19 |
| CN104156696B true CN104156696B (en) | 2017-04-12 |
Family
ID=51882193
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410353900.4A Active CN104156696B (en) | 2014-07-23 | 2014-07-23 | Bi-directional-image-based construction method for quick local changeless feature descriptor |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104156696B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110008966A (en) * | 2019-04-08 | 2019-07-12 | 湖南博睿基电子科技有限公司 | One kind being based on polar quick SIFT feature point extracting method |
| CN113052116B (en) * | 2021-04-06 | 2022-02-22 | 深圳华声医疗技术股份有限公司 | Ultrasonic video data processing method and device, ultrasonic equipment and storage medium |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2011069023A2 (en) * | 2009-12-02 | 2011-06-09 | Qualcomm Incorporated | Fast subspace projection of descriptor patches for image recognition |
| CN103093226A (en) * | 2012-12-20 | 2013-05-08 | 华南理工大学 | Construction method of RATMIC descriptor for image feature processing |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9530073B2 (en) * | 2010-04-20 | 2016-12-27 | Qualcomm Incorporated | Efficient descriptor extraction over multiple levels of an image scale space |
-
2014
- 2014-07-23 CN CN201410353900.4A patent/CN104156696B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2011069023A2 (en) * | 2009-12-02 | 2011-06-09 | Qualcomm Incorporated | Fast subspace projection of descriptor patches for image recognition |
| CN103093226A (en) * | 2012-12-20 | 2013-05-08 | 华南理工大学 | Construction method of RATMIC descriptor for image feature processing |
Non-Patent Citations (1)
| Title |
|---|
| 一种对亮度变化鲁棒的局部特征描述子及其快速生成算法;田时髦等;《中国期刊全文数据库 小型微型计算机***》;20130831;第34卷(第8期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104156696A (en) | 2014-11-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109978839B (en) | Method for detecting wafer low-texture defects | |
| CN110443128B (en) | Finger vein identification method based on SURF feature point accurate matching | |
| CN110334762B (en) | Feature matching method based on quad tree combined with ORB and SIFT | |
| CN107092871B (en) | Remote sensing image building detection method based on multiple dimensioned multiple features fusion | |
| CN104217221A (en) | Method for detecting calligraphy and paintings based on textural features | |
| CN103971126A (en) | Method and device for identifying traffic signs | |
| CN104778701A (en) | Local image describing method based on RGB-D sensor | |
| CN105719306B (en) | A kind of building rapid extracting method in high-resolution remote sensing image | |
| CN106650580B (en) | Goods shelf quick counting method based on image processing | |
| CN113392856B (en) | Image forgery detection device and method | |
| CN104123554A (en) | SIFT image characteristic extraction method based on MMTD | |
| KR101182683B1 (en) | A Visual Shape Descriptor Generating Method Using Sectors and Shape Context of Contour Lines and the Recording Medium thereof | |
| CN114862855B (en) | Textile defect detection method and system based on template matching | |
| Femiani et al. | Shadow-based rooftop segmentation in visible band images | |
| CN109785283B (en) | Texture feature matching method and device for fabric segmentation | |
| CN103093226A (en) | Construction method of RATMIC descriptor for image feature processing | |
| CN104156696B (en) | Bi-directional-image-based construction method for quick local changeless feature descriptor | |
| CN110516731A (en) | A kind of visual odometry feature point detecting method and system based on deep learning | |
| CN104268550A (en) | Feature extraction method and device | |
| CN111968115B (en) | Method and system for detecting orthopedic consumables based on rasterization image processing method | |
| CN103336964A (en) | SIFT image matching method based on module value difference mirror image invariant property | |
| CN111047614B (en) | Feature extraction-based method for extracting target corner of complex scene image | |
| Corke et al. | Image Feature Extraction | |
| JP4285640B2 (en) | Object identification method, apparatus and program | |
| CN112101283A (en) | Intelligent identification method and system for traffic signs |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20181008 Address after: 101300 room five, five floor, 1 building, 6 Jin An West Road, Shunyi District, Beijing. Patentee after: YUNTIANHAN TECHNOLOGY DEVELOPMENT Co.,Ltd. Address before: 510640 No. five, 381 mountain road, Guangzhou, Guangdong, Tianhe District Patentee before: South China University of Technology |
|
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20210914 Address after: 102200 Room 201, block B, national power investment group science and Technology Research Institute Co., Ltd., South District of future science city, Changping District, Beijing Patentee after: YUNTIANHAN TECHNOLOGY DEVELOPMENT Co.,Ltd. Patentee after: STATE POWER INVESTMENT CORPORATION Research Institute Address before: 101300 room five, five floor, 1 building, 6 Jin An West Road, Shunyi District, Beijing. Patentee before: YUNTIANHAN TECHNOLOGY DEVELOPMENT Co.,Ltd. |
|
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20221216 Address after: 102200 Room 201, block B, national power investment group science and Technology Research Institute Co., Ltd., South District of future science city, Changping District, Beijing Patentee after: YUNTIANHAN TECHNOLOGY DEVELOPMENT Co.,Ltd. Address before: 102200 Room 201, block B, national power investment group science and Technology Research Institute Co., Ltd., South District of future science city, Changping District, Beijing Patentee before: YUNTIANHAN TECHNOLOGY DEVELOPMENT Co.,Ltd. Patentee before: STATE POWER INVESTMENT CORPORATION Research Institute |