CN107369168A - The method of purification of registration point under a kind of big pollution background - Google Patents

Abstract

The invention discloses a kind of method of purification of registration point under big pollution background, belong to data purification field, be related to image actual registration point to the registration point method of purification in the case of being heavily polluted.Including：2 registration points pair of arbitrary extracting in some registration points pair；Two sample points in two images are connected respectively, seek the ratio of two end points pixel gray level sum to midpoint institute line section；By GTG in two images in step 2 and the ratio between difference build core collection；Merge core and concentrate repeated and redundant point pair；Using RANSAC methods, concentrate and be sampled in core, interior point detection is carried out in complete or collected works, extracts actual registration point pair, and obtain the registration model of two images.The present invention is by constructing core collection, concentrate and be sampled in core, to improve sampling efficiency, solve different scenes, different visual angles, target shape itself, under the conditions of outward appearance changes target registration identification, and then solve the problems, such as the purification of foul pollution registration point pair.

Description

The method of purification of registration point under a kind of big pollution background

Technical field

The invention belongs to data purification technical field, is related to image registration direction, and in particular under a kind of big pollution background The method of purification of registration point is image actual registration point to the registration point method of purification in the case of being heavily polluted.

Background technology

Classical image registration realizes that the registration point pointed by usual local feature description's is dirty to existing by describing son Dye, namely：The registering existing correctly registration of description, there is also the registration of mistake.Therefore, in image registration field, through retouching The registration point for stating sub- registration purifies to needs.Generally RANSAC (Matthew Brown, David G.Lowe, Automatic panoramic image stitching using invariant features,International journal of computer vision,2007,74(1):59-73.) it is used for the purification of description, but when description In the case of being heavily polluted, directly carrying out purification using RANSAC, speed to be present slow, and forms actual no solution.

On description son pollution, on the one hand describe subalgorithm and determine in itself.Description is typically found at image local Region, and image registration needs to carry out on global image, the registration point of local association is not to ensuring that under global sense still So association.On the other hand due to the complexity of registration task, determining pollution, sometimes performance is serious.If same object is not With the registration in scene.Even same object, the scene presented in the air when different also can be different.So as to pollute all the time Along with registration, and sometimes pollute also very serious.The purification of the description son pair seriously polluted is one and unavoidable asked Topic.

The problem of on RANSAC purification efficiencies, existing some achievements (Chum,Matas,Optimal randomized RANSAC,IEEE Transactions on Pattern Analysis and Machine Intelligence,2008,30(8):1472-1482；Tom Botterill,Steven Mills,Richard Green, Fast RANSAC hypothesis generation for essential matrix estimation,Digital Image Computing Techniques and Applications(DICTA),2011 International Conference on.IEEE,2011:561-566；Anders Hast,JohanAndrea Marchetti, Optimal ransac-towards a repeatable algorithm for finding the optimal set, 2013,21:21-30；Hongxia Gao,Jianhe Xie,Yueming Hu,Ze Yang,Hough-RANSAC:A Fast and Robust Method for Rejecting Mismatches,Chinese Conference on Pattern Recognition.Springer Berlin Heidelberg,2014:363-370.).Wherein, Tom Botterill etc. will 25% is reduced the time required to RANSAC purifications；The optimization RANSAC methods of the propositions such as Chum are faster 2- than classical RANSAC 10 times；The it is proposeds such as Hongxia Gao when only 20% data are True Data, Hough-RANSAC methods can be shown in Effect；When contamination data reaches 95%, the method for the proposition such as Anders Hast is still effective.These methods all improve weight dirt The RANSAC efficiency under data is contaminated, but when true registration point is sparse to extreme, when particularly pollution reaches 99%, these sides The ability of point still seems insufficient in method detection.

The content of the invention

The present invention is directed to image registration, it is proposed that the method for purification of registration point, extraction are extremely dirty under a kind of big pollution background The registration model of dye, pollution exterior point is filtered, extracts point in true, be a technology in image registration field.

In order to solve the above-mentioned technical problem, the technical solution adopted by the present invention is：Registration point under a kind of big pollution background Method of purification, it is characterised in that methods described includes step：Step 1: for two image I subject to registration₁With I₂Some registrations Point arbitrarily extracts 2 registration points pair to P；Step 2: connection figure is as I₁In two sample points, obtain two sample points Middle point coordinates, ask two end points to midpoint institute line section pixel gray level and the ratio between r₁；Using same method connection figure as I₂ In two sample points, ask GTG on the line segment of midpoint both sides and the ratio between r₂；Ratio calculated r₁With r₂Between poor e₁₂；Step 3: By GTG in two images in step 2 and the ratio between difference build core collection；Step 4: merging core concentrates repeated and redundant point pair； Step 5: using RANSAC methods, concentrated in core and carry out traversal sampling, interior point detection is carried out in complete or collected works, extracts actual registration Point pair, and obtain the registration model of two images.

In the above method, in the step 2, GTG and the ratio between computational methods be：Two in connection piece image are taken out Sampling point, obtains the middle point coordinates of two sample points first, then ask respectively two end points to midpoint institute line section pixel gray level it With, so ask GTG on the line segment of midpoint both sides and the ratio between.In the step 3, the specific method of structure core collection is：A. travel through The sampling of two registration points pair, repeat step one, step 2, obtain all sample points pair GTG and the ratio between poor e_ij；B. pass through Given threshold value, select GTG and the ratio between poor e_ijLess sample point is to setting up core collection.The attribute that core collection uses is constructed in method For：Two actual registration points are to the line on respective image due to undergoing identical content, the ash between them on midpoint Rank ratio is closer to；Conversely, the content that false registration point is undergone to line is different, the GTG ratio between them on midpoint Value difference is away from larger.Traversal frequency in sampling in the step 5 isWherein n is detection registration collection P element Number.

Present invention has the advantages that:The present invention gives a kind of method of purification of registration point under big pollution background, the present invention There is a situation where seriously to pollute for registration point caused by method for registering images, explore the method for purification of actual registration point pair.It is main The target in image is identified to registration point to purifying, solves the problems, such as the registering identification of target in image, wrapped Include：Solve the registering identification of target under different scenes, solve the registration identification of target under different visual angles, this figure of solution target The registration identification of target under the conditions of shape, outward appearance change.

Brief description of the drawings

The content expressed by this specification accompanying drawing and the mark in figure are briefly described below：

Fig. 1 is the method flow schematic diagram of the embodiment of the present invention.

Fig. 2 is the sampling figure of the embodiment of the present invention.

Fig. 3 is the core collection and complete or collected works' schematic diagram of the embodiment of the present invention.

Embodiment

Below against accompanying drawing, by the description to embodiment, for example involved each component of embodiment of the invention Shape, construction, the mutual alignment and annexation between each several part, the effect of each several part and operation principle, manufacturing process and Operate with method etc., is described in further detail, to help those skilled in the art to inventive concept of the invention, technology Scheme has more complete, accurate and deep understanding.

The method of purification of registration point, utilizes the spatial domain attribute structure of image pixel under a kind of big pollution background provided by the invention Core collection is made, concentrates and is sampled in core, to improve sampling efficiency, and then solves the purification of foul pollution registration point pair.This method Solves the extraction of sparse registration model, if 2 images subject to registration are respectively I₁With I₂, its registration point is to by existing method for registering (such as SIFT, SURF etc.) is obtained, and is set to P={ (p_1i,p_2i)|p_1i∈I₁,p_2i∈I₂, i=1,2 ..., n }, wherein m (p_1i,p_2i) =1 represents point to (p_1i,p_2i) it is actual registration point pair, m (p_1i,p_2i)=0 represents point to (p_1i,p_2i) it is false registration point pair. Problem to be solved by this invention is all actual registration points for seeking P to collecting P_TSo that P_T={ (p_1i,p_2i)|m(p_1i,p_2i)= 1,(p_1i,p_2i) ∈ P, and the registration model between two images is sought, namely homography conversion matrix H so that

The method of purification of registration point, exists serious for registration point caused by method for registering images under a kind of big pollution background The situation of pollution, is the method for purification for exploring actual registration point pair, and method and step specifically includes：

Step 1: for 2 image I subject to registration₁With I₂In n registration point to P={ (p_1i,p_2i) | i=1,2 ..., N }, arbitrarily extract i-th₁And i₂Individual registration point pairWithWherein p_1i∈I₁, p_2i∈I₂(in such as Fig. 2 P₁₁、p₁₂It is I₁On 2 registration points).

Step 2: connection piece image I₁In two sample points, obtain the middle point coordinates of two sample points, ask two End points to pixel gray level in midpoint institute line section and the ratio between r₁；The second width image I is connected using same method₂In two take out Sampling point, ask GTG on the line segment of midpoint both sides and the ratio between r₂；Seek ratio r₁With r₂Between poor e₁₂.The GTG of midpoint and the ratio between Computational methods are：1. connect piece image I₁In two sample pointsWith(the p in such as Fig. 2₁₁、p₁₂), obtain first The middle point coordinates q of two sample points₁<i₁,i₂>(the q in such as Fig. 2₁), then ask respectively two end points to midpoint institute line section picture Plain GTG sum, so ask GTG on the line segment of midpoint both sides and the ratio between r₁.2. it is 1. similar with step, connect the second width image I₂ In two sample pointsWith(the p in such as Fig. 2₂₁、p₂₂), wherein point coordinates q is obtained first₂<i₁,i₂>(in Fig. 2 q₂), then ask respectively two end points to midpoint institute line section pixel gray level sum, and then ask GTG on the line segment of midpoint both sides and The ratio between r₂.In set P traversal sampling after, obtain all sample points pair GTG and the ratio between differenceIt will be sampled by dictionary Arrangement, it is：

R={ e₁₂,e₁₃,…,e_1n,e₂₃,e₂₄,…,e_2n,…,e_n-1,n}.

Such as：Seek image I₁Decline in line segmentWithUpper pixel gray level and the ratio between r₁<i₁,i₂>.It might as well set at pixel p GTG be g (p), then image I₁Middle conductorOn pixel gray level and be：Figure As I₁Middle conductor p_1i2q₁On pixel gray level and be：Line segmentWithUpper picture Plain GTG and the ratio between be：Similarly, image I is sought₂Decline in line segmentWith Upper pixel gray level and the ratio betweenObtain r₁<i₁,i₂>With r₂<i₁,i₂>DifferenceIt is designated as：

Step 3: by GTG in two images in step 2 and the ratio between difference build core collection.By constructing core collection, Core is concentrated and is sampled, and then improves sampling efficiency.Structure core collection method be：A. two registration points are traveled through to take out sampling, traversal Sample number isWherein n is detection registration collection P element number.Repeat step one, step 2, obtain all Sample point to step 1. with step 2. in GTG and the ratio between poor e_ij.B. by given threshold value, e is selected_ijLess sampling Point is to setting up core collection.Construct core attribute that collection uses for：Two actual registration points are to the line on respective image due to experience Identical content, the GTG ratio between them on midpoint are closer to；Conversely, in false registration point undergone to line Hold difference, so as to which the GTG ratio difference between them on midpoint is away from larger.Such as threshold value T₁With T₂If T₁＜ T₂, construction collection Close P core collection P_c,(if the black registration point in Fig. 3 is to institute Show).

Step 4: merge core collection P_cMiddle repeated and redundant point pair, merge core and concentrate repeated and redundant point to being to seek sample point two-by-two To the union of collection.If two points drawing for the first time are to for { (A₁,A₂),(B₁,B₂), two points drawn for the second time to for {(A₁,A₂),(C₁,C₂), and twice sampling gained on the difference at image midpoint within threshold value allowed band, then will be Core, which is concentrated, repeats point twice to (A₁,A₂) merge into a point pair：

{(A₁,A₂),(B₁,B₂)}∪{(A₁,A₂),(C₁,C₂)={ (A₁,A₂),(B₁,B₂),(C₁,C₂)}。

Step 5: using RANSAC methods, P is concentrated in core_cIt is sampled, interior point detection is carried out in complete or collected works P (in Fig. 3 Dash-dotted gray line and solid black lines registration point to constituting detection complete or collected works P), extract actual registration point pair, obtain two images I₁ With I₂Actual registration point to P_T={ (p_1i,p_2i)|m(p_1i,p_2i)=1, (p_1i,p_2i)∈P}.Seek the registration between two images Model, namely homography conversion matrix H.H is asked using least square method so that：Core collection The methods of sampling is：Every time in core collection P_cIn take out that four registrations are right at random, the homography model matrix for asking this four registrations pair to determine H.The method of interior point detection is in complete or collected works：Registration point all in complete or collected works P is determined to substituting into above-mentioned random sampling one by one Model, for given difference ε₀If the registration point in P is right | H (p_1i)-p_2i| ＜ ε₀, then the registration point is to for interior point pair.It is interior Point detection is to detect as far as possible more interior points.

The present invention is exemplarily described above in conjunction with accompanying drawing, it is clear that present invention specific implementation is not by aforesaid way Limitation, as long as the improvement of the various unsubstantialities of inventive concept and technical scheme of the present invention progress is employed, or without changing Enter and the design of the present invention and technical scheme are directly applied into other occasions, within protection scope of the present invention.This hair Bright protection domain should be determined by the scope of protection defined in the claims.

Claims (5)

1. a purification method of registration points under a large pollution background, is characterized in that, described method comprises steps: Step 1. For several registration point pairs P of the two images _I1 and _I2 to be registered, randomly extract two registration point pairs; Step 2, connect the two sampling points in the image I ₁ , find out the midpoint coordinates of the two sampling points, and ask the ratio r ₁ of the pixel gray scale sum on the line segment connected from the two endpoints to the midpoint; use the same method Connect the two sampling points in the image I ₂ , find the ratio r ₂ of the gray scale sum on the line segments on both sides of the midpoint; calculate the difference e ₁₂ between the ratio r ₁ and r ₂ ; Step 3, constructing a core set by the difference between the grayscale sum ratio of the two images in step 2; Step 4, merging redundant point pairs in the core set; Step 5. Using the RANSAC method, traversal sampling is performed in the kernel set, interior point detection is performed in the full set, real registration point pairs are extracted, and a registration model of the two images is obtained. 2. the purification method of registration point under the heavy pollution background according to claim 1, is characterized in that, in described step 2, the calculation method of the ratio of gray scale sum is: connect two sampling points in an image, First find the coordinates of the midpoint of the two sampling points, and then calculate the sum of the pixel gray levels on the line segment connecting the two endpoints to the midpoint, and then calculate the ratio of the gray level sum on the line segments on both sides of the midpoint. 3. the purification method of registration point under the large pollution background according to claim 1, is characterized in that, in described step 3, the concrete method of constructing core set is: A. traverse the sampling of two registration points, repeat step 1 , Step 2, calculate the difference e _ij of the gray-scale sum ratio of all sampling point pairs; B. Through a given threshold, select the sampling point pair with a smaller gray-scale sum ratio difference e _ij to form a kernel set. 4. the purification method of registration point under the heavy pollution background according to claim 1, is characterized in that, the attribute that constructs core set adopts in the method is: two real registration points are connected on the respective images due to experiencing the same content, the gray scale ratios about the midpoint between them are relatively close; on the contrary, the content experienced by the false registration point pairs is different, and the gray scale ratio between them about the midpoint is quite different. 5. the purification method of registration point under the heavy pollution background according to claim 1, is characterized in that, the traversal sampling times in described step 5 is where n is the number of elements in the detection registration set P.