CN106991431B - Post-verification method for local feature point matching pairs - Google Patents

Abstract

The invention discloses a post-verification method for matching pairs of local feature points. The present invention first extracts local feature points in the image, and obtains candidate local feature point matching pairs through visual vocabulary; then, extracts attribute change values: main direction change value and orientation change value for the candidate local feature point matching pairs; The attribute change value and threshold value of , verify whether the two matching pairs are consistent; finally, the voting method is used to confirm whether the candidate local feature point matching pair is a correct matching pair according to the number of positive votes. This post-verification method can adapt to the effects of image cropping, rotation, scaling and other transformations, and can be used in applications such as image retrieval and classification based on visual vocabulary to improve the accuracy of retrieval and recognition. The invention has a very good verification effect for the feature point matching pairs in those non-perspective transformed images, and can greatly improve the accuracy and recall rate of copy retrieval in the application of image copy retrieval based on visual vocabulary.

Description

A Post-Verification Method for Local Feature Point Matching Pairs

technical field

The invention belongs to the field of computer image processing and image retrieval, and relates to a post-verification method for matching pairs of local feature points in two images.

Background technique

With the extensive research and application of local feature points in images, image analysis, recognition and retrieval based on local feature points has become an important method in the current image processing field. Borrowing from the bag-of-words model in document processing, an image can be represented as a collection of local features, thereby eliminating some redundant information in the image. In recent years, researchers have quantified the descriptors of local feature points into visual words, thus proposing the bag of visual words model. This model has become an important method of current image recognition and retrieval. The visual vocabulary bag-of-words model combined with inverted index is the most effective content-based image retrieval method, and it has very good robustness. In the application of image retrieval, it can deal with various editing operations and transformations of images; and the inverted index structure improves the retrieval efficiency and can realize real-time query in large-scale image databases. However, the visual vocabulary obtained by the feature vector quantization of local feature points has no clear meaning relative to the vocabulary in natural language. Its discriminative ability is weak and cannot fully represent the content of the local image. In order to ensure the distinguishing ability of visual vocabulary: the more visual vocabulary in the dictionary, the better; however, the more visual vocabulary, the weaker its anti-noise ability, and it takes more calculation to quantify the feature vector into visual vocabulary quantity. In addition, reducing the number of visual words in the dictionary in order to eliminate the influence of noise also leads to a decrease in the distinguishing ability of visual words, resulting in a high false matching rate of visual words. The mismatch of visual vocabulary brings difficulties to the subsequent image similarity calculation.

Researchers have proposed many constructive methods for the problem of mismatching visual vocabulary. These methods can be mainly divided into two categories: one is to add an additional descriptor to the visual vocabulary to improve the discriminative ability of the visual vocabulary, and the other is to perform spatial consistency based on matching pairs of candidate local feature points in the two images. Validation, thereby filtering out mismatched pairs. The existing additional descriptor methods include: the method of embedding color information into local feature points proposed by Liang Zheng (Liang Zheng, Shengjin Wang, Qi Tian, Coupled Binary Embedding for Large-Scale Image Retrieval, IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL.23, NO.8, 2014), Yao proposed the context information of visual vocabulary as an additional descriptor, and a Hamming embedding method proposed by H.Jégou. These methods need to add additional descriptors to the index library, which increases the storage consumption of the system. In addition, the robustness of additional descriptors is also a concern.

The spatial verification method based on matching pairs of local feature points is a post-verification method. It is to calculate the spatial consistency of local feature points matching between the query image and the candidate image based on the candidate image or target already found during retrieval or classification. Because image editing and slight perspective transformation do not change the spatial relative relationship of local feature points on a certain target in the image, the spatial consistency between local feature point matching pairs is widely used for post-validation of filtering mismatched pairs method. The original method is to use the RANSAC method to obtain the transformation parameters of the image on the set of local feature point matching pairs, and the matching pairs that do not conform to the transformation model are considered to be mismatches. Due to the low efficiency of the RANSAC method, some researchers have proposed a weak geometric consistency method, which determines the transformation parameter of the image according to the difference between the scale and the main direction of the local feature points, and uses this parameter to filter the mismatch. In addition, Zhou proposed a spatial coding (Spatial Coding) method, which identifies the correct match by the consistency of the spatial position (relative positional relationship of x and y coordinates) of matching pairs (Zhou WG, Li HQ, Lu Y et al. .Encoding spatial context for large-scale partial-duplicate Web image retrieval. JOURNAL OF COMPUTER SCIENCE AND TECHNOLOGY 29(5):837-848 Sept. 2014). Lingyang Chu established a graphical model of principal orientation and position consistency. The graph model considers strongly connected matching pairs as correct matching pairs (Lingyang Chu, ShuqiangJiang, et al. Robust Spatial Consistency Graph Model for Partial Duplicate ImageRetrieval, IEEE TRANSACTIONS ON MULTIMEDIA, VOL.15, NO.8, PP.1982- 1986, DECEMBER 2013). Wu combines visual vocabulary into bundles through the maximum stable limit region, then indexes images based on bundles, and achieves similarity measurement through matching of visual vocabulary in bundles.

Aiming at the problem that the matching accuracy is reduced due to the problem that the distinguishing ability is weakened after the local features are quantified into visual words, the method of the present invention proposes to use the attribute change consistency between the matching pairs of local feature points and a voting method to confirm the correct matching. right. This method extends the existing post-verification method of spatial consistency of local feature points; it has high processing speed, can cope with various image editing operations, and can be applied to related applications such as image recognition and retrieval.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a post-verification method for matching pairs of local feature points in the current visual vocabulary-based image content retrieval application, which can be used to confirm correct matching pairs of local feature points and filter wrong matching pairs, thereby improving retrieval. 's accuracy.

The concrete steps of the method of the present invention are:

Step (1) obtaining matching pairs of local feature points in the two images according to the visual vocabulary corresponding to the local feature points;

The visual vocabulary is the vocabulary ID obtained by the quantization of the feature vector of the local feature points in the image;

The local feature points are obtained by a local feature point detection method (such as: SIFT), which has related attributes in the image: spatial position, scale, main direction and feature vector;

The matching pair of local feature points is a pair of local feature points with the same visual vocabulary in the two images. Suppose two images: Img and Img', where the local feature points are represented by _Vi and V _m 'respectively; if the visual vocabulary obtained by quantization of _Vi and V _m ' is the same, then (V _i , V _m ') is A matching pair of local feature points.

Step (2) Calculate the attribute change value of the matching pair of local feature points.

The attribute change value is the difference between the attributes of the two matched local feature points, which reflects the change of a local feature point in the original image after image transformation to the corresponding feature point in the result image. The method of the present invention proposes two attribute change values: the main direction change value and the azimuth change value. Suppose there are two matching pairs M1: (V _i , V _m ') and M2: (V _j , V _n ') between the images: Img and Img'. The main direction of V _i is expressed as θ _i , and the position attribute is expressed as (Px _i , Py _i ); the main direction of V _m ′ is expressed as θ _m ′, and the position attribute is expressed as (Px _m ′, Py _m ′). The main direction change value (Diff_Ori) is defined as the difference of the main directions of the local feature points in the matching pair, as shown in formula (1). The acquisition of the orientation change value requires two matching pairs. First, the orientations of the two local feature points of the two matching pairs in the same image are obtained. In the image Img, the orientation values of the local features V _i and V _j are Direction(i, j ), as shown in formula (2), where the arctan2 function is used to obtain the arc tangent value, the angle is: the origin points between the vector of the coordinates (Py _j -Py _i , Px _j -Px _i ) and the positive direction of the x-axis, The angle in the counterclockwise direction; the same azimuth values of V _m ' and V _n ' can be obtained by formula (3).

Diff_Ori(i,m)=θ _m ′-θ _i (1)

Direction(i,j)=arctan2(Py _j -Py _i ,Px _j -Px _i ) (2)

Direction′(m,n)=arctan2(Py′ _n -Py′ _m ,Px′ _n -Px′ _m ) (3)

Therefore, the change value of the orientation attribute Diff_Dir(i,m) is:

Diff_Dir(i,m)=Direction(i,j)-Direction'(m,n) (4)

Step (3) Confirm whether the matching pairs of local feature points are consistent according to the consistency of attribute change values between the matching pairs of local feature points.

The consistency of the attribute change value is judged through a threshold. All attribute change values meet the threshold requirements, then the two matching pairs of local feature points are considered to be consistent. The consistency of the main direction change values of the two matching pairs M1 and M2 is obtained by formulas (5) and (6), where TH_ORI is the threshold value of the main direction change consistency.

Diff_Ori_M(M1,M2)=|Diff_Ori(i,m)-Diff_Ori(j,n)| (5)

Among them, Diff_Ori_M(M1,M2) represents the difference between the main direction change values of the two matching pairs M1 and M2; Diff_Ori(i,m) represents the main direction change value of the matching pair M1; Diff_Ori(j,n) represents the matching pair M2 The main direction change value; Ori_Cons(M1, M2) represents the result value of the consistency of the two matching pairs M1 and M2.

The method of the invention eliminates the influence of image rotation on the orientation change value by subtracting the main direction change value. Specifically, the consistency judgment of the azimuth change values of the two matching pairs M1 and M2 is realized by formula (7)(8)(9); formula (8) is used to eliminate the influence of image rotation, where TH_Dir is the azimuth change consistency threshold.

Diff_Dir_M(M1,M2)=|Diff_Dir(i,m)-Diff_Dir(j,n)| (7)

Diff_Dir(M1,M2)=|Diff_Dir_M(M1,M2)-Diff_Ori_M(M1,M2)| (8)

Among them, Diff_Dir_M(M1,M2) represents the difference between the azimuth change values of the two matching pairs M1 and M2; Diff_Dir(i,m) represents the azimuth change value of the matching pair M1; Diff_Dir(j,n) represents the azimuth change value of the matching pair M2 value; Diff_Dir(M1, M2) represents the difference between the main direction change value of the two matching pairs M1 and M2 and the difference between the azimuth change value; Dir_Cons(M1, M2) represents the two matching pairs M1 and M2 The resulting value for orientation consistency.

Step (4) adopts voting method to determine whether a certain matching pair is the correct matching pair. That is: for a certain matching pair M _i , verify the consistency of attribute change values with other matching pairs, and cast a positive vote if they are consistent. If the ratio of the number of positive votes to the number of candidate matching pairs of the two images is greater than a given threshold (Th_Votes), the matching pair M _i is considered to be a correct match.

The present invention has the following beneficial effects with respect to the prior art:

The method of the present invention is different from the weak geometric constraint method, which is to separate and independently process the consistent features. The accuracy of the consistency verification between them.

The main direction change value and the orientation change value proposed by the method of the present invention have a very high matching pair consistency discrimination rate; assuming that the orientations between the feature points are randomly distributed, then the orientation change value is used. is TH_Dir/π; when TH_Dir is 0.087 (5 degrees), its probability is 2.8%. Similarly, when TH_ORI is 0.087 (5 degrees), the probability of misjudgment by the main direction change value is also 2.8%. The probability of misjudgment of consistency between two feature joints to discriminate matching pairs is: 0.00077. Therefore, the matching pair consistency verification method proposed by the method of the present invention has a very high accuracy rate.

The method of the invention regards the feature point matching problem as the matching problem of part of the image content in the two images, and the change of the orientation and main direction of the matching pair of feature points between the same image content should be consistent. In addition, since the consistency verification of feature point matching pairs has a very high accuracy, the voting method can effectively deal with the problem of partial image content matching. Therefore, the method of the present invention has very good robustness to complex editing operations such as image cropping and object addition.

The method of the invention eliminates the influence of the image rotation on the difference of the azimuth change value by subtracting the difference value of the main direction change value. The main direction of the local feature point itself has rotation invariance. Therefore, the method of the present invention is rotationally robust.

The method of the invention has a very good verification effect on the feature point matching pairs in those non-perspective transformed images, and can greatly improve the accuracy and recall rate of copy retrieval in the application of image copy retrieval based on visual vocabulary.

Description of drawings

Fig. 1 represents the flow chart of the method of the present invention;

Fig. 2 Matching results of local feature points based on visual vocabulary;

Figure 3 is a schematic diagram of calculating the change value of the orientation attribute;

Fig. 4 The number of votes of matching pairs obtained based on the voting method;

Fig. 5 method of the present invention confirms the effect of correct matching pair;

FIG. 6 is a schematic diagram of the effectiveness comparison result of the post-verification method of the present invention.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings. It should be noted that the described embodiments are only for the understanding of the present invention, but do not have any limiting effect on it.

The concrete steps of the method of the present invention are:

Step (1) obtain the matching pairs of local feature points in the two images according to the visual vocabulary corresponding to the local feature points; there are currently many methods for the extraction of local feature points, and the method of the present invention adopts the currently widely used scale-invariant descriptor ( SIFT), which is robust to rotation, scale transformation, etc. After the SIFT descriptor is extracted, the image is represented as a set of local feature points {S _{i }} , and Si is a local feature point descriptor in the image, which has the following related properties in the image: feature vector (F _i ), Principal direction (θ _i ), scale (σ _i ), and spatial position (Px _i , Py _i ). The matching of local feature points can quantify feature vectors into visual vocabulary IDs, and then obtain matching pairs of local feature points based on the consistency of visual vocabulary IDs. In order to quantify the feature vectors of local feature points into visual vocabulary IDs, in this embodiment, a product quantization method is adopted, and K-means clustering is used to construct a visual vocabulary dictionary. This quantization method has very high quantization efficiency. In this embodiment, a 32-dimensional SIFT local feature descriptor is used, and the product quantization method divides the 32-dimensional feature vector into 4 groups, each of which has 8 dimensions. A set of 8-dimensional feature vectors is quantified into 32 word roots according to the sample base; a dictionary of 220 visual words can be constructed according to ^the combination of word roots. According to the above steps, suppose two images: Img and Img', where the local feature points are represented by _Vi and _Vm 'respectively; if the visual vocabulary IDs obtained by the quantization of the feature vectors of _Vi and _Vm ' are the same, then ( V _i , V _m ′) is a matching pair of local feature points. FIG. 2 is an example of the matching result in this embodiment. The line segment across the two images represents the matching pair of local feature points, the two endpoints are the spatial positions of the local feature points in the image, and the arrow line represents the scale and main direction of the feature points. In FIG. 2 , the white line segments represent correct matching pairs of local feature points, and the corresponding image contents are consistent; while the black line segments represent wrong matching pairs of local feature points. It can also be seen from the figure that the local contents of the two local feature points in the wrong matching pair have a certain similarity (both are edge points with a little radian), but from the whole image, their contents are inconsistent. The goal of this method is to identify and filter these false matching pairs.

Step (2) Calculate the attribute change value of the matching pair of local feature points. The attribute change value is the difference between the attributes of the two matched local feature points, which reflects the change of a local feature point in the original image after image transformation to the corresponding feature point in the result image. The method of the present invention proposes two attribute change values: the main direction change value and the azimuth change value. Suppose there are two matching pairs M1: (V _i , V _m ') and M2: (V _j , V _n ') between the images: Img and Img'. The principal direction and position properties of V _i and V _m ′ are expressed as: (θ _i , Px _i , Py _i ) and (θ _m ′, Px _m ′, Py _m ′), respectively. The main direction change value (Diff_Ori) is defined as the difference of the main directions of the local feature points in the matching pair, as shown in formula (10). The acquisition of the orientation change value requires two matching pairs. First, the orientations of the two local feature points of the two matching pairs in the same image are obtained. In the image Img, the orientation values of the local features V _i and V _j are Direction(i, j ), as shown in formula (11), where the arctan2 function is used to obtain the arc tangent value, the angle is: the origin points between the vector of the coordinates (Py _j -Py _i , Px _j -Px _i ) and the positive direction of the x-axis, The angle in the counterclockwise direction; the same azimuth values of V _m ' and V _n ' can be obtained by formula (12).

Diff_Ori(i,m)=θ _m ′-θ _i (10)

Direction(i,j)=arctan2(Py _j -Py _i ,Px _j -Px _i ) (11)

Direction'( _m , _n )=arctan2( _Py'n -Py'm ,Px'n _-Px'm ) (12)

Therefore, the change value of the orientation attribute Diff_Dir(i,m) is:

Diff_Dir(i,m)=Direction(i,j)-Direction'(m,n) (13)

The change value of the orientation attribute can be visually represented by Fig. 3; the orientation attribute in the image is represented by the angle between the line connecting the two feature points and the x-axis. If the image is not subject to a rotation attack, the orientation properties of two identical matching pairs should be the same. The orientation attribute change value expresses the stability of orientation of two feature points in the image. Likewise, the main direction property change value is also robust to most operations except rotation.

Step (3) Confirm whether the matching pairs of local feature points are consistent according to the consistency of attribute change values between the matching pairs of local feature points. The consistency of the attribute change value is judged through a threshold. All attribute change values meet the threshold requirements, then the two matching pairs of local feature points are considered to be consistent. The two matching pairs M1: (V _i , V _m ′) and M2: (V _j , V _n ′) have the same main direction change values through formulas (14) and (15), where TH_ORI is consistent with the main direction changes Sex threshold. This method uses the comparison between the two matching pairs to confirm; if the image is rotated, the corresponding matching feature points in the image will be rotated accordingly; thus the main direction change value between the two matching pairs will not be affected. The effect of image rotation. In this embodiment, we have conducted relevant experiments on relevant test libraries, and finally set TH_ORI to 0.087 (5 degrees).

Diff_Ori_M(M1,M2)=|Diff_Ori(i,m)-Diff_Ori(j,n)| (14)

Among them, Diff_Ori_M(M1,M2) represents the difference between the main direction change values of the two matching pairs M1 and M2; Diff_Ori(i,m) represents the main direction change value of the matching pair M1; Diff_Ori(j,n) represents the matching pair M2 The main direction change value; Ori_Cons(M1, M2) represents the result value of the consistency of the two matching pairs M1 and M2.

The method of the invention eliminates the influence of image rotation on the orientation change value by subtracting the main direction change value. Specifically, the consistency judgment of the azimuth change values of the two matching pairs M1 and M2 is realized by formulas (16) (17) (18); formula (17) is used to eliminate the influence of image rotation, where TH_Dir is the azimuth change consistency threshold.

Diff_Dir_M(M1,M2)=|Diff_Dir(i,m)-Diff_Dir(j,n)| (16)

Diff_Dir(M1,M2)=|Diff_Dir_M(M1,M2)-Diff_Ori_M(M1,M2)| (17)

Among them, Diff_Dir_M(M1,M2) represents the difference between the azimuth change values of the two matching pairs M1 and M2; Diff_Dir(i,m) represents the azimuth change value of the matching pair M1; Diff_Dir(j,n) represents the azimuth change value of the matching pair M2 value; Diff_Dir(M1, M2) represents the difference between the main direction change value of the two matching pairs M1 and M2 and the difference between the azimuth change value; Dir_Cons(M1, M2) represents the two matching pairs M1 and M2 The resulting value for orientation consistency.

In this embodiment, we have carried out relevant experiments on relevant test libraries, and finally set TH_Dir to also be 0.087 (5 degrees).

Because the calculation of the orientation consistency between the candidate matching pairs in the two images requires a large amount of calculation. In this embodiment, in order to improve the verification efficiency, the method adopts the following two strategies. Strategy 1 is: prioritize the judgment of the consistency of the main direction; if it is confirmed that the two matching pairs are inconsistent, there is no need to judge the consistency of the subsequent direction, thereby reducing the amount of calculation. The second strategy is: when a matching pair has been verified by other matching pairs or has obtained enough affirmative votes, no subsequent verification is required, and the correct matching pair is directly confirmed, thereby improving efficiency. In addition, we can also filter candidate matching pairs by setting a high-frequency vocabulary, thereby improving the efficiency of verification.

Step (4) adopts voting method to determine whether a certain matching pair is the correct matching pair. That is: for a certain matching pair M _i , verify the consistency of attribute change values with other matching pairs, and cast a positive vote if they are consistent. If the number of positive votes divided by the number of candidate matching pairs of the two images is greater than a given threshold (Th_Votes), the matching pair M _i is considered to be a correct match. In this embodiment, Th_Votes is set to 0.2 according to the experiment. According to the above steps, the algorithm flow for calculating the correct matching logarithm between two images is as follows:

Figure 4 is an example of the result of the candidate matching pair verification, in which the numbers on the side of each matching pair represent the number of positive votes obtained, and the number of positive votes obtained for those mismatched pairs is very small, most of which are 0. Therefore, the method can effectively identify those mismatched pairs.

In order to prove the effectiveness of the method, in this example, we conducted relevant tests. We adopt copy image retrieval based on local feature point matching as the application scenario of this method. The standard test library Holidays and Web image library for copy image detection are used as test libraries. The Holidays test library includes 157 original images, as well as jpeg images of various compression ratios obtained on them, and various cropped copies of image sets. The web test library is a collection of various copy images we collected from the Internet, with an average of 32 copy images in a set. This test uses the mAP value in the field of image retrieval as the evaluation index. The test results are shown in Figure 6. The Spatial Coding method is currently the best post-verification method for visual vocabulary, and the Baseline method is a copy image retrieval effect that is directly based on the number of candidate matching pairs without post-verification. It can be confirmed from Figure 6 that this method has certain advantages on both test libraries.

Figure 5 shows the application effect of the method in the retrieval of copied images. The gray and black lines represent correct matching pairs, while the white lines represent false matching pairs identified by this method.

Claims (3)

1. a post-verification method for matching pairs of local feature points, is characterized in that specifically comprising the steps: Step (1) obtaining matching pairs of local feature points in the two images according to the visual vocabulary corresponding to the local feature points; The visual vocabulary is the vocabulary ID obtained by the quantization of the feature vector of the local feature points in the image; The local feature points are obtained by a local feature point detection method, and have relevant attributes in the image: spatial position, scale, main direction and feature vector; The described local feature point matching pair is a pair of local feature points with the same visual vocabulary in the two images; suppose two images: Img and Img', where the local feature points are represented by V _i and V _m 'respectively; if The visual vocabulary obtained by V _i and V _m ' is the same, then (V _i , V _m ') is a matching pair of local feature points; Step (2) calculate the attribute change value of the matching pair of local feature points; The attribute change value is the difference between the attributes of the two matched local feature points, which reflects the change of a local feature point in the original image after the image is transformed to the corresponding feature point in the result image. It is assumed that the image has two attributes. Change value: main direction change value and azimuth change value; Assume image: There are two matching pairs between Img and Img' M1: (V _i ,V _m ') and M2: (V _j ,V _n '); V _i The main direction of V m is expressed as θ _i , and the position attribute is expressed as (Px _i ,Py _i ); the main direction of V _m ′ is expressed as θ _m ′, and the position attribute is expressed as (Px _m ′,Py _m ′); the main direction change value (Diff_Ori) is defined as the difference between the main directions of the local feature points in the matching pair, as shown in formula (1); the acquisition of the orientation change value requires two matching pairs, first obtain the two local matching pairs in the same image. The orientation of the feature point, the orientation value of the local features V _i and V _j in the image Img is Direction(i, j), as shown in formula (2), where the arctan2 function is used to obtain the arc tangent value, the angle is: the origin The angle between the vector pointing to the coordinates (Py _j -Py _i , Px _j -Px _i ) and the positive direction of the x-axis, along the counterclockwise direction; the same orientation values of V _m ' and V _n ' can be obtained by formula (3) get; Diff_Ori(i,m)=θ _m ′-θ _i (1) Direction(i,j)=arctan2(Py _j -Py _i ,Px _j -Px _i ) (2) Direction′(m,n)=arctan2(Py′ _n -Py′ _m ,Px′ _n -Px′ _m ) (3) Therefore, the change value of the orientation attribute Diff_Dir(i,m) is: Diff_Dir(i,m)=Direction(i,j)-Direction'(m,n) (4) Step (3) confirming whether the matching pairs of local feature points are consistent according to the consistency of attribute change values between the matching pairs of local feature points; Step (4) adopts the voting method to determine whether a certain matching pair is a correct matching pair; that is, for a certain matching pair M _i , verify the consistency of the attribute change values with other matching pairs, and vote one if they are consistent. A positive vote; if the ratio of the positive votes to the number of candidate matching pairs of the two images is greater than a given threshold Th_Votes, the matching pair M _i is considered to be a correct match. 2. the back verification method of a kind of local feature point matching pair according to claim 1, is characterized in that step 3 is specifically realized as follows: The consistency of the attribute change values is judged by a threshold value, and when all the attribute change values meet the threshold value requirements, it is considered that the two local feature point matching pairs are consistent; The two matching pairs M1: (V _i , V _m ') and M2: (V _j , V _n ') have the same main direction change values through formulas (5) and (6), where TH_ORI is consistent in the main direction changes sex threshold; Diff_Ori_M(M1,M2)=|Diff_Ori(i,m)-Diff_Ori(j,n)| (5)

Among them, Diff_Ori_M(M1,M2) represents the difference between the main direction change values of the two matching pairs M1 and M2; Diff_Ori(i,m) represents the main direction change value of the matching pair M1; Diff_Ori(j,n) represents the matching pair M2 The main direction change value; Ori_Cons(M1, M2) represents the result value of the consistency of the two matching pairs M1 and M2. 3. the back verification method of a kind of local feature point matching pair according to claim 2, is characterized in that: The effect of image rotation on the orientation change is removed by subtracting the principal orientation change; specifically: The consistency judgment of the orientation change values of the two matching pairs M1 and M2 is realized by formula (7) (8) (9); formula (8) is used to eliminate the influence of image rotation, wherein TH_Dir is the orientation change consistency threshold; Diff_Dir_M(M1,M2)=|Diff_Dir(i,m)-Diff_Dir(j,n)| (7) Diff_Dir(M1,M2)=|Diff_Dir_M(M1,M2)-Diff_Ori_M(M1,M2)| (8)

Among them, Diff_Dir_M(M1,M2) represents the difference between the azimuth change values of the two matching pairs M1 and M2; Diff_Dir(i,m) represents the azimuth change value of the matching pair M1; Diff_Dir(j,n) represents the azimuth change value of the matching pair M2 value; Diff_Dir(M1, M2) represents the difference between the main direction change value of the two matching pairs M1 and M2 and the difference between the azimuth change value; Dir_Cons(M1, M2) represents the two matching pairs M1 and M2 The resulting value for orientation consistency.

Images