CN110197184A - A kind of rapid image SIFT extracting method based on Fourier transformation - Google Patents
A kind of rapid image SIFT extracting method based on Fourier transformation Download PDFInfo
- Publication number
- CN110197184A CN110197184A CN201910316073.4A CN201910316073A CN110197184A CN 110197184 A CN110197184 A CN 110197184A CN 201910316073 A CN201910316073 A CN 201910316073A CN 110197184 A CN110197184 A CN 110197184A
- Authority
- CN
- China
- Prior art keywords
- point
- characteristic point
- fourier transformation
- scale
- image
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/14—Fourier, Walsh or analogous domain transformations, e.g. Laplace, Hilbert, Karhunen-Loeve, transforms
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
Abstract
The invention discloses a kind of methods that quick SIFT feature may be implemented and extract.When establishing difference of Gaussian image pyramid, according to the convolution theorem of Fourier transformation, the former methodical convolution for directly asking gaussian kernel function Yu image subject to registration is replaced with the inverse Fourier transform of the Fourier transformation of Gaussian convolution kernel function and the product of the Fourier transformation of image subject to registration, the advantage for relying on Fast Fourier Transform (FFT) greatly reduces the calculation amount for extracting SIFT feature descriptor and calculates the time.The detection of subsequent extreme point and the descriptor generation method of the mechanism of rejecting and characteristic point are consistent with tradition SIFT, while to realize the advantages that keeping original SIFT method stability and noise immunity, realize that quick, real-time, stable, reliable SIFT feature is extracted.
Description
Technical field
The invention belongs to computer vision fields, more particularly to the characteristic point rapidly extracting of image, registration and identification etc.
Field.
Background technique
SIFT (Scale invariant features transform) algorithm is the generally acknowledged strongest image matching method of comprehensive matching ability, tool
The invariance such as standby scaling, translation, scale, stability is strong with noise immunity, is usually used in handling between two images in the presence of translation, rotation
Turn, the registration under the conditions of affine transformation etc..But this method has shortcoming again, that is, due to the structure of Gauss scale pyramid
It builds and the processes operands such as the descriptor computation of characteristic point is big, time-consuming.Existing improvement is mostly to be directed to how to improve SIFT
Feature point extraction speed, while the promotion of comprehensive performance is realized as far as possible.
Summary of the invention
The purpose of the present invention is to provide a kind of reliable methods that quick SIFT feature may be implemented and extract, due to improving
SIFT descriptor can more or less reduce the performance that SIFT feature carries out images match, therefore the present invention is from Gauss scale gold
The building of tower is started with.The convolution theorem of Fourier transformation in basis signal treatment theory, by calculating Gaussian convolution core and figure
The product of the Fourier transformation of piece, then the method for transformation of inverting replace original is methodical to do convolution in airspace, greatly reduce meter
Calculation amount and calculating time, to realize the building of quick difference of Gaussian image pyramid.The present invention does not change descriptor simultaneously
Generation method, to accelerate the speed that this method carries out feature point extraction on the premise of ensuring performance, realize quickly, it is real
When, stabilization, reliable SIFT feature extract.
It realizes the technical solution of the present invention is as follows: a kind of rapid image SIFT extracting method based on Fourier transformation, including
Following steps:
Step 1: the convolution theorem based on Fourier transformation constructs difference of Gaussian (DOG) scale space.
Step 2: in DOG space search extreme point, characteristic point position and place scale are primarily determined.
Step 3: the accurate positioning of characteristic point, while eliminating low contrast and eliminating skirt response, obtain stable key
Accurate location, the place dimensional information of point.
Step 4: being calculated feature vertex neighborhood histogram of gradients as neighborhood window using scale where 1.5 times of characteristic points,
It is characterized a distribution principal direction θ.
Step 5: it is calculated according to 16*16 neighborhood around key point and generates SIFT feature descriptor.
In step 1, the convolution theorem based on Fourier transformation, building difference of Gaussian (DOG) scale space is specifically done
Method are as follows:
Define the scale-space representation function of image:
L (x, y, σ)=G (x, y, σ) * I (x, y)
Wherein * represents convolution, I (x, y) representing input images, and G (x, y, σ) is the gaussian kernel function that scale is σ:
And the process for generating L (x, y, σ) can be accelerated by the convolution theorem of Fourier transformation, implementation method are as follows:
L (x, y, σ)=F-1{F{G(x,y,σ)}·F{I(x,y)}}
Wherein, F { } is Fourier transformation;F-1{ } is inverse Fourier transform.
Difference of Gaussian function D (x, y, σ) is finally generated by the adjacent scale that invariant k keeps apart according to two:
D (x, y, σ)=L (x, y, k σ)-L (x, y, σ)
In step 2, in DOG space search extreme point, characteristic point position and place scale, search strategy are primarily determined
Are as follows:
In the space DOG, eight neighborhood and its institute of each pixel where with it in same same image of scale
Totally 26 pixels are compared for nine neighborhoods of scale adjacent up and down, and only the pixel is that maximum or minimum Shi Caihui is chosen as
Candidate extreme point.
In step 3, whether D (x)≤0.03 is met by the absolute size of extreme point in the detection space DOG, if full
It is sufficient then think that the contrast is lower, then cast out the extreme point, otherwise then retains the extreme point and carry out subsequent detection.Meanwhile it going
Except the specific steps of skirt response are as follows:
Obtain the Hessian matrix of the difference image around characteristic point, it may be assumed that
With Tr (H)=Dxx+DyyIndicate the mark of the matrix, Det (H)=DxxDyy-(Dxy) 2 indicate the matrix determinant
Value, decision condition are as follows:
R=10 is usually taken, if above formula is set up, retains this feature point, otherwise just casts out this feature point.
In step 4, the principal direction of characteristic point is determined with direction histogram.Coordinate is the point of (x, y), mould and direction
Be expressed as follows:
M (x, y)={ [L (x+1, y)-L (x-1, y)]2+[L(x,y+1)-L(x,y-1)]2}1/2
It is sampled in the neighborhood window centered on characteristic point, with the gradient side of gradient orientation histogram statistics neighborhood territory pixel
To the histogram of gradients within the scope of 0 °~360 ° is averagely divided into 36 columns, every 10 degree of columns.The peak value pair of histogram of gradients
It answers angle to represent the principal direction of characteristic point, when the ratio between minor peaks and main peak value are more than 80%, this direction is considered the spy
Levy the auxiliary direction of point.One characteristic point may be designated with multiple directions, to enhance robustness.
In step 5, is calculated according to 16*16 neighborhood around key point and generates SIFT feature descriptor, specific practice are as follows:
First centered on characteristic point, reference axis is rotated to be to the principal direction of characteristic point, i.e., in characteristic point neighbors around
By reference axis rotate θ angle (θ is characterized principal direction), after rotation in neighborhood pixel new coordinate are as follows:
Later centered on characteristic point, the pixel of 16 × 16 neighborhoods around it is evenly dividing as 4 × 4 fritters, every
The histogram of gradients that eight directions are drawn on a fritter, forms seed point.Therefore the descriptor of each characteristic point is by 16 seeds
Point is formed, and each seed point has the information in 8 directions, to obtain the feature point description symbol of 4 × 4 × 8=128 dimension.
Detailed description of the invention:
Fig. 1 is flow chart of the invention.
Fig. 2 is the schematic diagram of step 5.
Claims (2)
1. a kind of rapid image SIFT extracting method based on Fourier transformation, which comprises the following steps:
Step 1: the convolution theorem based on Fourier transformation constructs difference of Gaussian (DOG) scale space.
Step 2: in DOG space search extreme point, characteristic point position and place scale are primarily determined.
Step 3: the accurate positioning of characteristic point, while eliminating low contrast and eliminating skirt response, obtain stable key point
Accurate location, place dimensional information.
Step 4: feature vertex neighborhood histogram of gradients is calculated as neighborhood window using scale where 1.5 times of characteristic points, for spy
Sign point distribution principal direction θ.
Step 5: it is calculated according to 16*16 neighborhood around key point and generates SIFT feature descriptor.
2. a kind of rapid image SIFT extracting method based on Fourier transformation according to claim 1, which is characterized in that
When generating difference of Gaussian image pyramid, it is based on Fourier transformation convolution theorem, uses the Fourier of Gaussian convolution kernel function
The inverse Fourier transform of the product of the Fourier transformation of transformation and image subject to registration replaces original is methodical directly to seek Gaussian kernel letter
Several convolution with image subject to registration, rely on the advantage of Fast Fourier Transform (FFT), greatly reduce and extract SIFT feature descriptor
Calculation amount and calculating time.Remaining step is consistent with traditional SIFT feature extraction algorithm, thus guarantee SIFT stability,
While reliability, the quick detection and description of image characteristic point are realized.Whole process technical requirements are as follows:
Define the scale-space representation function of image:
L (x, y, σ)=G (x, y, σ) * I (x, y)
Wherein*Convolution, I (x, y) representing input images are represented, G (x, y, σ) is for scaleσGaussian kernel function:
And the process for generating L (x, y, σ) can be accelerated by the convolution theorem of Fourier transformation, implementation method are as follows:
L (x, y, σ)=F-1{F{G(x,y,σ)}·F{I(x,y)}}
Wherein, F { } is Fourier transformation;F-1{ } is inverse Fourier transform.
Difference of Gaussian function D (x, y, σ) is finally generated by the adjacent scale that invariant k keeps apart according to two:
D (x, y, σ)=L (x, y, k σ)-L (x, y, σ)
Later, in the space DOG, eight neighborhood and its institute of each pixel where with it in same same image of scale
Totally 26 pixels are compared for nine neighborhoods of scale adjacent up and down, and only the pixel is that maximum or minimum Shi Caihui is chosen as
Candidate extreme point.
Then, whether D (x)≤0.03 is met by the absolute size of candidate extreme point in the detection space DOG, recognized if meeting
It is lower for the contrast, then cast out the extreme point, otherwise then retains the extreme point and carry out subsequent detection.Meanwhile removing edge
The specific steps of response are as follows:
Obtain the Hessian matrix of the difference image around characteristic point, it may be assumed that
With Tr (H)=Dxx+DyyIndicate the mark of the matrix, Det (H)=DxxDyy-(Dxy)2It indicates the value of the matrix determinant, determines
Condition are as follows:
R=10 is usually taken, if above formula is set up, retains this feature point, otherwise just casts out this feature point.
Then, the principal direction of characteristic point is determined with direction histogram.Coordinate is the point of (x, y), and the expression formula in mould and direction is such as
Under:
M (x, y)={ [L (x+1, y)-L (x-1, y)]2+[L(x,y+1)-L(x,y-1)]2}1/2
It is sampled in the neighborhood window centered on characteristic point, it, will with the gradient direction of gradient orientation histogram statistics neighborhood territory pixel
Histogram of gradients within the scope of 0 °~360 ° is averagely divided into 36 columns, every 10 degree of columns.The peak value corresponding angles of histogram of gradients
This direction is considered this feature point when the ratio between minor peaks and main peak value are more than 80% by the principal direction that degree represents characteristic point
Auxiliary direction.One characteristic point may be designated with multiple directions, to enhance robustness.
Finally, reference axis to be rotated to be to the principal direction of characteristic point centered on characteristic point, i.e., will be sat in characteristic point neighbors around
Parameter rotate θ angle (θ is characterized principal direction), after rotation in neighborhood pixel new coordinate are as follows:
Later centered on characteristic point, the pixel of 16 × 16 neighborhoods around it is evenly dividing as 4 × 4 fritters, each small
The histogram of gradients that eight directions are drawn on block, forms seed point.Therefore the descriptor of each characteristic point is by 16 seed dots
At each seed point has the information in 8 directions, to obtain the feature point description symbol of 4 × 4 × 8=128 dimension.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910316073.4A CN110197184A (en) | 2019-04-19 | 2019-04-19 | A kind of rapid image SIFT extracting method based on Fourier transformation |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910316073.4A CN110197184A (en) | 2019-04-19 | 2019-04-19 | A kind of rapid image SIFT extracting method based on Fourier transformation |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110197184A true CN110197184A (en) | 2019-09-03 |
Family
ID=67752044
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910316073.4A Pending CN110197184A (en) | 2019-04-19 | 2019-04-19 | A kind of rapid image SIFT extracting method based on Fourier transformation |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110197184A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111860272A (en) * | 2020-07-13 | 2020-10-30 | 敦泰电子(深圳)有限公司 | Image processing method, chip and electronic device |
CN112149728A (en) * | 2020-09-22 | 2020-12-29 | 成都智遥云图信息技术有限公司 | Rapid multi-modal image template matching method |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106558072A (en) * | 2016-11-22 | 2017-04-05 | 重庆信科设计有限公司 | A kind of method based on SIFT feature registration on remote sensing images is improved |
CN106778487A (en) * | 2016-11-19 | 2017-05-31 | 南宁市浩发科技有限公司 | A kind of 2DPCA face identification methods |
-
2019
- 2019-04-19 CN CN201910316073.4A patent/CN110197184A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106778487A (en) * | 2016-11-19 | 2017-05-31 | 南宁市浩发科技有限公司 | A kind of 2DPCA face identification methods |
CN106558072A (en) * | 2016-11-22 | 2017-04-05 | 重庆信科设计有限公司 | A kind of method based on SIFT feature registration on remote sensing images is improved |
Non-Patent Citations (2)
Title |
---|
刘波 等: "基于卷积定理的人脸验证CNN模型加速", 《北京工业大学学报》 * |
肖健: "SIFT特征匹配算法研究与改进", 《万方数据知识服务平台》 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111860272A (en) * | 2020-07-13 | 2020-10-30 | 敦泰电子(深圳)有限公司 | Image processing method, chip and electronic device |
CN111860272B (en) * | 2020-07-13 | 2023-10-20 | 敦泰电子(深圳)有限公司 | Image processing method, chip and electronic device |
CN112149728A (en) * | 2020-09-22 | 2020-12-29 | 成都智遥云图信息技术有限公司 | Rapid multi-modal image template matching method |
CN112149728B (en) * | 2020-09-22 | 2023-11-17 | 成都智遥云图信息技术有限公司 | Rapid multi-mode image template matching method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106960451B (en) | Method for increasing number of feature points of image weak texture area | |
CN108256394B (en) | Target tracking method based on contour gradient | |
CN104751142B (en) | A kind of natural scene Method for text detection based on stroke feature | |
CN110334762B (en) | Feature matching method based on quad tree combined with ORB and SIFT | |
CN104050681B (en) | A kind of road vanishing Point Detection Method method based on video image | |
CN105096317A (en) | Fully automatic calibration method for high performance camera under complicated background | |
CN107123130B (en) | Kernel correlation filtering target tracking method based on superpixel and hybrid hash | |
CN110222661B (en) | Feature extraction method for moving target identification and tracking | |
CN106408597A (en) | Neighborhood entropy and consistency detection-based SAR (synthetic aperture radar) image registration method | |
CN104123554A (en) | SIFT image characteristic extraction method based on MMTD | |
CN106023171A (en) | Image corner detection method based on turning radius | |
CN115272306B (en) | Solar cell panel grid line enhancement method utilizing gradient operation | |
CN108133211B (en) | Power distribution cabinet detection method based on mobile terminal visual image | |
CN115471682A (en) | Image matching method based on SIFT fusion ResNet50 | |
CN108319961B (en) | Image ROI rapid detection method based on local feature points | |
CN115457277A (en) | Intelligent pavement disease identification and detection method and system | |
CN110197184A (en) | A kind of rapid image SIFT extracting method based on Fourier transformation | |
CN108257153A (en) | A kind of method for tracking target based on direction gradient statistical nature | |
CN109766850B (en) | Fingerprint image matching method based on feature fusion | |
CN111199558A (en) | Image matching method based on deep learning | |
CN109711420B (en) | Multi-affine target detection and identification method based on human visual attention mechanism | |
Zhang et al. | Research on surface defect detection of rare-earth magnetic materials based on improved SSD | |
CN112102379B (en) | Unmanned aerial vehicle multispectral image registration method | |
CN111898589B (en) | Unmanned aerial vehicle image rapid registration method based on GPU+feature recognition | |
CN106934395B (en) | Rigid body target tracking method adopting combination of SURF (speeded Up robust features) and color features |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190903 |
|
WD01 | Invention patent application deemed withdrawn after publication |