CN104834931A - Improved SIFT algorithm based on wavelet transformation - Google Patents

Abstract

In order to solve the problems of long operating time and low matching ratio of an SIFT (Scale Invariant Feature Transform) algorithm, the present invention puts forward an improved SIFT algorithm. On the basis of the original classic SIFT algorithm, a two-dimensional Mallat fast wavelet transformation algorithm is introduced, a low frequency component of an image is reconstructed, then the group number of Gauss pyramids is adjusted, downsampling frequency is reduced, and finally mismatching points are rejected through an optimized RANSAC algorithm. The improved SIFT algorithm of the present invention not only reduces the consumed time of matching, but also improves the matching rate, thus the improved SIFT algorithm is better than the original SIFT algorithm.

Description

A kind of scale invariant feature matching algorithm of the improvement based on wavelet transformation

Technical field

The present invention relates to the images match field in computer vision, specifically refer to a kind of scale invariant feature matching algorithm of the improvement based on wavelet transformation.

Background technology

Images match is being the focus studied of people and difficult point in recent decades always, it is in transformation space, find one or more conversion, make from two width of the Same Scene of different time, different sensors or different visual angles or multiple image spatially consistent, be applied to many fields at present, wherein most widely used is image registration field and moving target recognition and tracking field.Therefore image matching technology is in occupation of vital status.

Due to the change of shooting time, shooting angle, physical environment, the image taken is made to be subject to the impact of various noise.Under these conditions, how matching algorithm reaches that precision is high, high, the speed of coupling accuracy is fast, strong robustness and Parallel Implementation become the target that people pursue.For this problem, a lot of Chinese scholars conducts extensive research.Image matching algorithm is divided into four kinds by the ultimate principle based on coupling: coupling, feature-based matching, the coupling based on model and the coupling based on transform domain of being correlated with based on gray scale.At present, research is feature-based matching the most widely, wherein the most classical algorithm improved SIFT (scale invariant feature coupling) algorithm of summing up by D.G.Lowe in 2004, and this algorithm all can keep good unchangeability to translation, rotation, scaling, brightness change.Afterwards, a lot of scholar proposed different innovatory algorithm according to SIFT algorithm.

Chen Shu Rong etc. proposes Contourlet-SIFT Feature Correspondence Algorithm, mate, but calculated amount is bigger than normal, not requirement of real time to rotating under metric space after invariant features carries out contourlet transformation again.Cao Juan etc. propose based on D ²the improvement SIFT feature matching algorithm of OG feature point detection operator, is applicable to image information and enriches and the occasion higher to requirement of real-time, but the match point logarithm that algorithm extracts is relatively few, limits the image type of this algorithm process.Yang Xingfang proposes and replaces Euclidean distance as the similarity measurement between feature descriptor using city block distance, reduces the time complexity of similarity measurement formula, but does not improve robustness.Yu Lili proposes a kind of SIFT feature matching algorithm of the improvement based on image Radon change, reduce the dimension of SIFT feature vector, improve characteristic matching efficiency, but performance has much room for improvement when actual scene uses.Kong Jun proposes the binocular vision coupling of Multi resolution feature extraction, although matching rate is improved, mates consuming time longer, ageing poor.In order to improve correct matching rate further, strengthen robustness and real-time, the present invention proposes the scale invariant feature matching algorithm of the improvement based on wavelet transformation.

Summary of the invention

The object of the invention is to for there is the problems such as correct matching rate is not high, robustness is not strong, Riming time of algorithm is long in existing method, on the basis of original classical scale invariant feature matching algorithm, propose the scale invariant feature matching algorithm technology of the improvement based on wavelet transformation, develop a kind of strong robustness, correct matching rate is high, is applicable to the image matching method in the high scene of requirement of real-time.

The present invention is based on following consideration: long in order to solve former SIFT Riming time of algorithm, and the problem that matching rate is not high, proposes a kind of SIFT algorithm of improvement.On the basis of original classical SIFT algorithm, introduce two-dimentional Mallat fast wavelet transform algorithm, rebuild the low-frequency component of image, then gaussian pyramid group number is adjusted, reduce down-sampled number of times, reject Mismatching point finally by the RANSAC algorithm optimized.It is consuming time that algorithm after improvement not only reduces coupling, and matching rate have also been obtained raising.

The technical scheme that the present invention is based on the scale invariant feature matching algorithm of the improvement of wavelet transformation is as follows:

(1) carry out two-dimentional Mallat wavelet transformation to two width images subject to registration to decompose, obtain the low-frequency component of image and level and vertical high frequency composition, give up the radio-frequency component after wavelet transformation decomposes, the low-frequency component of image is reconstructed, obtains new image;

(2) two width new images tectonic scale spaces (DoG) are utilized, gaussian pyramid is generated by the difference of Gaussian convolution kernel of image and different scale, because new images is through wavelet transformation, give up partial information, when building difference of Gaussian pyramid, reduce down-sampled number of times, one deck fewer than the original generation number of plies, reduce the time in tectonic scale space;

(3) find key point in DoG space, the check point of centre and it with 8 consecutive point of yardstick and neighbouring yardstick corresponding 18 totally 26 points compare, guarantee finally can obtain abundant key point.

(4) due to DoG value to noise and edge more responsive, detect in DoG metric space that above Local Extremum also accurately will could orientate unique point as through further inspection.Position and the yardstick of key point is accurately determined by the three-dimensional quadratic function of matching.

The calculation expression in its Grad m (x, y) and direction θ (x, y) is

{[L(x+1,y)-L(x-1,y)] ²+[L(x,y+1)-L(x,y-1)] ²} ^1/2

$\mathrm{\θ}(x,y)={\tan }^{-1}\left[\frac{L(x,y+1)-L(x,y-1)}{L(x+1,y)-L(x-1,y)}\right]$

(5) unique point is described through and carries out image block to key point peripheral region, calculates the histogram of gradients in each piece, generates unique vector descriptor.In order to strengthen the robustness of coupling, to each key point use 4 × 4 totally 16 Seed Points describe, just can produce 128 data for a key point like this, the final SIFT feature forming 128 dimensions are vectorial.

(6) after the SIFT feature vector of two width images generates, we adopt the Euclidean distance of key point proper vector to be used as the similarity determination tolerance of key point in two width images for next step.Sampling eigen point and the nearest unique point of sample characteristics point Euclidean distance, in these two unique points, if nearest distance is less than certain proportion threshold value except distance near in proper order, then accept this pair match point.Reduce this proportion threshold value, SIFT match point number can reduce, but more stable.When calculating the Euclidean distance between unique point, have employed BBF algorithm to process the proper vector of 128 dimensions.

(7) after all unique points are slightly mated, use RANSAC algorithm estimate two images between identity transformation matrix and it can be used as geometrical constraint, and then remove some Mismatching points, complete the exact matching between image, improve matching efficiency.

Beneficial effect of the present invention: the present invention adopts two-dimentional Mallat wavelet transformation to carry out decomposition and reconstruction to image, gives up the radio-frequency component comprising much noise and a small amount of useful information of image, is only reconstructed the low-frequency component of image.Image subject to registration after two-dimensional wavelet transformation process, can reduce each pixel participating in coupling on the one hand, improve matching speed, decrease weak match point on the other hand, thus error hiding rate is declined.The present invention is when tectonic scale space, because the pyramidal last one deck of difference of Gaussian contains little point of interest, too much influence is not had to final matching result, and the information comprised after image wavelet transform process also reduces a part, so down-sampled number of times can be reduced, make the pyramidal number of plies of difference of Gaussian reduce one deck, reduce the time in tectonic scale space, part Mismatching point can also be removed simultaneously, improve matching efficiency.The present invention use after all unique points of image being carried out to thick coupling RANSAC algorithm estimate two images between identity transformation matrix and it can be used as geometrical constraint, and then remove some Mismatching points, complete the exact matching between image, improve matching efficiency.

Accompanying drawing explanation

Fig. 1 is SIFT algorithm steps.

Embodiment

For making the object, technical solutions and advantages of the present invention clearly understand, below in conjunction with instantiation, and with reference to accompanying drawing, elaborate to the specific embodiment of the present invention, the present invention is including but not limited to example.

Concrete steps of the present invention are as follows:

1, two-dimentional Mallat wavelet transformation

Mallat is subject to the inspiration of tower algorithm in Multi-resolution analysis of wavelet transform theory with the applied research of image procossing, proposes the fast algorithm of the tower multiscale analysis and restructure of signal, is horse traction spy (Mallat) algorithm.

Use the product structure 2-d wavelet base of two identical one dimension wavelet functions and unidimensional scale function, the scaling function of generation and three wavelet functions are respectively:

$\left\{\begin{array}{c}\mathrm{\φ}(x,y)=\mathrm{\φ}\left(x\right)\mathrm{\φ}\left(y\right)\\ {\mathrm{\ψ}}^1(x,y)=\mathrm{\φ}\left(x\right)\mathrm{\ψ}\left(y\right)\\ {\mathrm{\ψ}}^2(x,y)=\mathrm{\ψ}\left(x\right)\mathrm{\φ}\left(y\right)\\ {\mathrm{\ψ}}^3(x,y)=\mathrm{\ψ}\left(x\right)\mathrm{\ψ}\left(y\right)\end{array}\right.$

If for picture signal to be analyzed, its two dimension is approached image and is

$A_{i}f=A_{j+1}f+D_{j+1}^{1}f+D_{j+1}^{2}f+D_{j+1}^{3}f$

In formula

$\left\{\begin{array}{c}{A}_{j+1}f=\underset{m=-\∞}{\overset{\∞}{\mathrm{\Σ}}}\underset{n=-\∞}{\overset{\∞}{\mathrm{\Σ}}}{c}_{j+1}(m,n){\mathrm{\φ}}_{j+1}(m,n)\\ D_{j+1}^{i}f=\underset{m=-\∞}{\overset{\∞}{\mathrm{\Σ}}}\underset{n=-\∞}{\overset{\∞}{\mathrm{\Σ}}}{d}_{j+1}^i(m,n){\mathrm{\φ}}_{j+1}(m,n)\end{array}\right.(i=\mathrm{1,2,3})$

Utilize the orthogonality of scaling function and wavelet function, can obtain:

$c_{j+1}(m,n)=\underset{k=-\∞}{\overset{\∞}{\mathrm{\Σ}}}\underset{l=-\∞}{\overset{\∞}{\mathrm{\Σ}}}h(k-2m)h(l-2n)c_j(k,l)$

Above formula illustrates, the scale coefficient c of j+1 metric space _j+1(m, n) can by the scale coefficient c of j metric space _j(k, l) is weighted summation through two dimensional filter coefficient and obtains.

Again

$\left\{\begin{array}{c}{d}_{j+1}^1=\underset{k=-\∞}{\overset{\∞}{\mathrm{\Σ}}}\underset{l=-\∞}{\overset{\∞}{\mathrm{\Σ}}}h(k-2m)g(l-2n)c_j(k,l)\\ d_{j+1}^2=\underset{k=-\∞}{\overset{\∞}{\mathrm{\Σ}}}\underset{l=-\∞}{\overset{\∞}{\mathrm{\Σ}}}g(k-2m)h(l-2n)c_j(k,l)\\ d_{j+1}^3=\underset{k=-\∞}{\overset{\∞}{\mathrm{\Σ}}}\underset{l=-\∞}{\overset{\∞}{\mathrm{\Σ}}}g(k-2m)g(l-2n)c_j(k,l)\end{array}\right.$

Introduce matrix operator, make H _rand H _crepresent respectively with scaling filter coefficient pair array the operator of row and column effect, G _rand G _cthe expression operator of wavelet filter coefficient to row and column effect respectively, then two-dimentional Mallat decomposition algorithm is

$\left\{\begin{array}{c}{C}_{j+1}=H_r{H}_c{C}_j\\ D_{j+1}^1=H_r{H}_c{C}_j\\ D_{j+1}^2=G_r{H}_c{C}_j\\ D_{j+1}^3=G_r{G}_c{C}_j\end{array}\right.(j=\mathrm{0,1},....,J)$

Two dimension Mallat restructing algorithm is

$C_j=H_r^{*}{H}_c^{*}{C}_{j+1}+H_r^{*}{G}_c^{*}{D}_{j+1}^{*}+G_r^{*}{H}_c^{*}{D}_{j+1}^2+G_r^{*}{G}_c^{*}{D}_{j+1}^3$

Utilize above-mentioned two-dimentional Mallat Wavelet Transformation Algorithm to carry out decomposition and reconstruction to image, obtain new image.

2, the SIFT algorithm improved

SIFT algorithm searches key point on different metric spaces, and calculate the direction of key point.The realization of SIFT algorithm comprises two aspects, and one is the generation of image characteristic point, and another is the coupling of SIFT feature point between different images.SIFT feature is extracted and be can be analyzed to four steps, as shown in Figure 1.

2.1 tectonic scale spaces

The metric space L (x, y, σ) of one width two-dimensional image I (x, y) is defined as a change Gaussian function for yardstick and the convolution of original image, namely

L(x,y,σ)＝G(x,y,σ)*I(x,y)

In formula, σ is the space scale factor of metric space, and G (x, y, σ) is gaussian kernel function, and it is defined as

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

In order to stable key point can be detected in metric space, difference of Gaussian (DoG) operator is used to be similar to Laplce---Gauss (LoG) operator of dimension normalization.Generated by the difference of Gaussian convolution kernel of image and different scale:

D(x,y,σ)＝L(x,y,kσ)-L(x,y,σ)

Gaussian pyramid is O group, S layer altogether, is generally taken as 4 groups, 5 layers, often organizes ground floor and obtains by upper one group of last one deck is down-sampled.Before tectonic scale space, first utilize two-dimensional wavelet transformation to treat registering images herein and carry out pre-service, given up radio-frequency component, only remained low-frequency component and it is rebuild, so without the need to doing the down-sampled number of times the same with during former figure size.Therefore, the group number that the present invention is directed to gaussian pyramid adjusts, and reduces down-sampled number of times, makes the pyramidal number of plies of difference of Gaussian reduce one deck.

2.2 extreme points detect

Key point is made up of the Local Extremum in DoG space.In order to find DoG Function Extreme Value point, the check point of centre and it with 8 consecutive point of yardstick and neighbouring yardstick corresponding 18 totally 26 points compare, finally guarantee can both extreme point be detected at metric space and two dimensional image space.

2.3 key point location

Due to DoG value to noise and edge more responsive, therefore, in DoG metric space, detect that above Local Extremum also accurately will could orientate unique point as through further inspection.In order to strengthen stability, the raising noise resisting ability of coupling, accurately determined position and the yardstick of key point by the three-dimensional quadratic function of matching, by removing the key point of low contrast and unstable skirt response point.

Come for key point gives direction by the gradient solving each extreme point, adopt histogram of gradients statistic law to determine the direction of key point, the calculation expression in its Grad m (x, y) and direction θ (x, y) is

{[L(x+1,y)-L(x-1,y)] ²+[L(x,y+1)-L(x,y-1)] ²} ^1/2

$\mathrm{\θ}(x,y)={\tan }^{-1}\left[\frac{L(x,y+1)-L(x,y-1)}{L(x+1,y)-L(x-1,y)}\right]$

2.4 Feature Descriptors generate

Unique point is described through carries out image block to key point peripheral region, calculates the histogram of gradients in each piece, generates unique vector descriptor.In order to strengthen the robustness of coupling, to each key point use 4 × 4 totally 16 Seed Points describe, just can produce 128 data for a key point like this, the final SIFT feature forming 128 dimensions are vectorial.

2.5 Feature Points Matching

After the SIFT feature vector of two width images generates, we adopt the Euclidean distance of key point proper vector to be used as the similarity determination tolerance of key point in two width images for next step.Sampling eigen point and the nearest unique point of sample characteristics point Euclidean distance, in these two unique points, if nearest distance is less than certain proportion threshold value except distance near in proper order, then accept this pair match point.Reduce this proportion threshold value, SIFT match point number can reduce, but more stable.When calculating the Euclidean distance between unique point, have employed BBF algorithm to process the proper vector of 128 dimensions.

After all unique points are slightly mated, use RANSAC algorithm estimate two images between identity transformation matrix and it can be used as geometrical constraint, and then remove some Mismatching points, complete the exact matching between image, improve matching efficiency.

Claims (5)

1. the scale invariant feature matching algorithm based on the improvement of wavelet transformation, it is characterized in that carrying out two-dimentional Mallat wavelet transformation to two width images subject to registration decomposes, obtain the low-frequency component of image and level and vertical high frequency composition, give up the radio-frequency component after wavelet transformation decomposes, the low-frequency component of image is reconstructed, obtains new image; Utilize two width new images tectonic scale spaces (DoG), gaussian pyramid is generated by the difference of Gaussian convolution kernel of image and different scale, because new images is through wavelet transformation, give up partial information, when building difference of Gaussian pyramid, reduce down-sampled number of times, one deck fewer than the original generation number of plies, reduce the time in tectonic scale space; Find key point in DoG space, the check point of centre and it with 8 consecutive point of yardstick and neighbouring yardstick corresponding 18 totally 26 points compare, guarantee finally can obtain abundant key point.After the SIFT feature vector of two width images generates, next step adopts the Euclidean distance of key point proper vector to be used as the similarity determination tolerance of key point in two width images.Sampling eigen point and the nearest unique point of sample characteristics point Euclidean distance, in these two unique points, if nearest distance is less than certain proportion threshold value except distance near in proper order, then accept this pair match point.Reduce this proportion threshold value, SIFT match point number can reduce, but more stable.When calculating the Euclidean distance between unique point, have employed BBF algorithm to process the proper vector of 128 dimensions.After all unique points are slightly mated, use RANSAC algorithm estimate two images between identity transformation matrix and it can be used as geometrical constraint, and then remove some Mismatching points, complete the exact matching between image, improve matching efficiency.

2. the scale invariant feature matching algorithm of the improvement based on wavelet transformation according to claim 1, it is characterized in that: adopt two-dimentional Mallat wavelet transformation to carry out decomposition and reconstruction to image, give up the radio-frequency component comprising much noise and a small amount of useful information of image, only the low-frequency component of image is reconstructed.When tectonic scale space, because the pyramidal last one deck of difference of Gaussian contains little point of interest, too much influence is not had to final matching result, and the information comprised after image wavelet transform process also reduces a part, so down-sampled number of times can be reduced, the pyramidal number of plies of difference of Gaussian is made to reduce one deck.Use after all unique points of image being carried out to thick coupling RANSAC algorithm estimate two images between identity transformation matrix and it can be used as geometrical constraint, and then remove some Mismatching points, complete the exact matching between image.

3. the scale invariant feature matching algorithm of the improvement based on wavelet transformation according to claim 1 and 2, it is characterized in that: described two-dimentional Mallat wavelet transformation decomposes image, give up the radio-frequency component of image, only the low-frequency component of image is reconstructed.

4. the scale invariant feature matching algorithm of the improvement based on wavelet transformation according to claim 1 and 2, it is characterized in that: when tectonic scale space, decrease down-sampled number of times, make the pyramidal number of plies of difference of Gaussian reduce one deck, reduce the redundant computation of algorithm, improve real-time.

5. the scale invariant feature matching algorithm of the improvement based on wavelet transformation according to claim 1 and 2, is characterized in that: use RANSAC algorithm to remove part Mismatching point, improve accuracy and the accuracy of coupling.