CN106023230B - A kind of dense matching method of suitable deformation pattern - Google Patents

Abstract

A dense matching method suitable for deformed images, belonging to the field of image matching technology, the method includes step 1: manually extract matching feature point pairs in the left image and right image respectively; step 2: use quadratic polynomial to correct the relative deformation, obtain the corrected right image; Step 3: carry out dense matching to the left image and the corrected right image; Step 4: determine the pixel coordinates of the matching pixel of the pixel in the left image in the original right image; the present invention will The polynomial correction and coordinate correspondence preservation mechanism is used for dense matching of deformed images, which can be used for image matching in various situations, especially in the case of large deformations of images, and can also obtain better matching results. The dense matching provides a new solution. In the corrected image, the relative parallax is used to constrain the search area, which narrows the search range of matching points, improves the matching efficiency, and increases the reliability of the matching results.

Description

A Dense Matching Method Suitable for Deformed Images

technical field

The invention belongs to the technical field of image matching, and in particular relates to a dense matching method suitable for deformed images.

Background technique

For the matching between deformed image pairs, domestic scholars have carried out more research on feature matching. In 2010, Yang Heng et al. proposed a new local invariant feature detection and description algorithm. Harris corner points are extracted from above, and then the extreme points of the three-dimensional scale space are searched in the fixed-size search window centered on the Harris corner point as the position and feature scale of the local feature points, and finally the main direction is calculated for each feature point, and Using gradient distance and direction histograms to describe local features and complete feature matching; Chen Mengting et al. proposed a high-resolution remote sensing image matching algorithm based on Harris corners and SIFT descriptors in 2012. First, the Harris corners were extracted, and then the The obtained points are described by SIFT, and finally the matching is completed using the nearest neighbor/sub-nearest neighbor ratio method; in 2012, Zhao Xi'an et al. proposed an automatic stereo image matching method with scale and rotation invariance, firstly constructing a three-scale image based on directional wavelet transform The feature point operator performs two-scale matching to ensure its scale invariance, and then constructs a 64-dimensional description vector of feature points to solve the rotation invariance of image matching, and finally completes feature matching; Ye Yuanxin et al. proposed a combination of SIFT in 2013 The multi-source remote sensing image matching method with edge information, firstly detects feature points in the Gaussian difference scale space, and uses phase consistency to extract reliable edge information, then combines the improved SIFT and shape context to describe the feature points, and finally the Ou In 2014, Zhang Zhengpeng et al. proposed a vehicle-mounted panorama sequence image matching method based on optical flow feature clustering, using non-parametric mean shift feature clustering ideas, and using SIFT multi-scale The position quantity and optical flow vector of the feature matching points construct the spatial domain and value domain of the image feature space, and use the corresponding salient image optical flow features in the feature space as the clustering conditions to realize the matching of panoramic sequence images, and adopt epipolar geometry In 2015, Xu Qiuhui proposed an improved DCCD and SIFT descriptor image matching method, using the improved DCCD to quickly detect key points on the image, and then determine the main direction of the key points to generate feature points , use the SIFT descriptor to describe the feature points, and after obtaining the feature matching results, use the BBF algorithm and the random sampling consensus algorithm (RANSAC) to perform rough matching of feature points and elimination of mismatched feature points; Zhang Zhengpeng et al. proposed an adaptive motion structure in 2015 The vehicle-mounted panorama sequence image matching method based on characteristics firstly defines an adaptive bandwidth matrix based on the sampling points in the spatial domain and the local spatial structure of the optical flow domain, and uses the distance weighting method of the local optical flow feature vector to describe the structural characteristics of the motion similarity in the optical flow domain. Relax the diffusion process, and then give the expression form of the adaptive multivariable kernel density function, and give the solution of the mean shift vector, the termination condition and the selection method of the seed point, and finally integrate the SIFT description feature and the motion structure feature to establish a unified panorama Image matching framework; Shaw In 2015, Xiongwu et al. proposed a fast matching method for oblique images with affine invariance. First, the initial affine matrix was calculated by estimating the camera axis orientation parameters of the image, and the corrected image was obtained through inverse affine transformation. SIFT matching, and improve the matching accuracy and reliability through multiple constraints; in 2015, Yan Li et al. proposed a slice-corrected spherical panoramic image matching method. First, the equidistant projection panoramic image is divided into several sub-image blocks according to Each sub-block is projected and transformed to obtain a distortion-corrected image, and SIFT feature description and extraction are performed on it, and all the results are combined to obtain the feature set of the entire panoramic image, and then the feature matching between multiple panoramic images is completed; Zhao Baowei proposed an improved Hough transform multispectral image matching method with disparity constraints in 2015. This method uses a pyramid image matching strategy. The top pyramid image is matched through a scale-invariant feature conversion operator to provide the initial disparity constraints. Other layers The left and right images of the pyramid are matched using the improved Hough transform image method. During the matching process, the upper layer matching results are used to provide disparity constraints for the current layer image matching, and the feature matching is finally completed.

Although the research results of the above-mentioned scholars have achieved good matching results, only low-precision 3D surface reconstruction can be completed through feature matching, that is, the frame of the 3D surface can be obtained, and 3D reconstruction with high precision requirements cannot be completed. Further research is still needed. to complete the dense matching. The general dense matching method usually uses a certain matching measure as the basis for measuring whether two pixel points are matching points, such as: sum of absolute difference, sum of squared difference, truncated absolute difference, normalized cross-correlation, de-mean normalization Cross-correlation and other methods are sensitive to noise and can only complete dense matching for ideal images. For images with large deformations, when using the matching window for matching measurement calculations, the pixels that were originally matching points are deformed. This leads to a deviation in the calculation of the matching measure, which affects the final matching result.

Chen Wang et al. proposed a dense matching algorithm based on region boundary constraints and graph cut optimization in 2009. This method uses the constraints between region boundaries and boundary pixels to construct an energy function, so that the function can not only obtain the global optimal solution, but also The final reconstruction surface is relatively smooth, and finally a good dense matching result is obtained by using the classic image in the field of computer vision. When the image is scaled, the window-based region matching will have deviations in the implementation process of this method. In 2010, Ge Liang et al. proposed an improved dense matching algorithm for stereo pairs. This method first uses the region growing technique to find the single texture region in the image, and then uses the entire region as a matching primitive to obtain the dense disparity of the single texture region. Finally, the experiment was carried out with the international standard test image, which proved the feasibility and accuracy of the method. This method focuses on the single-textured area, but when the image is deformed, the single-textured area will also be deformed, resulting in the algorithm's failure. Reduced efficiency. In 2012, Li Haibin et al. proposed a 3D scene reconstruction method based on dense matching of candidate points, establishing grid nodes representing depth information in space, and reasonably planning the distribution of nodes in the depth direction to improve the timeliness of matching. The trinocular vision system replaces the binocular and improves the reliability of matching by making a second judgment. This method is mainly for the classic binocular stereo image processing. When the external parameter matrix is unknown and the relative rotation between the images occurs, the image cannot be completed. The epipolar line correction, so that subsequent dense matching cannot be completed. In 2013, Hu Chunhai et al. proposed a multi-baseline dense matching combining disparity growing and tensor, using the SIFT operator for preliminary feature matching, using the feature matching result as the root point of dense matching for disparity growing, and using three views Matching constraints improve the accuracy of dense matching, but using three views as a constraint for dense matching, when there are only two images, the constraints proposed by the algorithm cannot be fulfilled, thus affecting the reliability of the final matching result. Zhang Jielin et al. proposed a hierarchical matching method in 2014. Firstly, the thermonuclear signal function is used to detect feature points, and the local fusion strategy and the farthest point sampling method are used to optimize the feature matching results. The thermonuclear signal descriptor, and use the entropy to sort the feature points according to the significance for the initial layer matching, and finally realize the dense matching from coarse to fine through the local matching of the neighborhood of each layer of feature points. The problem of confusing topological relationships.

Contents of the invention

Aiming at the deficiencies in the above-mentioned prior art, the present invention provides a dense matching method suitable for deformed images.

Technical scheme of the present invention is as follows:

A dense matching method suitable for deformed images, comprising the following steps:

Step 1: Manually extract matching feature point pairs in the left image and the right image respectively;

The left image and the right image are a pair of relatively deformed original images to be matched; the matching feature point pairs are greater than 5 pairs; the matching feature point pairs constitute matching feature points, which can cover the overlapping area of the left image and the original right image ;

Step 2: Use a quadratic polynomial to correct the relative deformation of the original right image to obtain the corrected right image;

Step 2-1: Convert the pixel coordinates of the matching feature points in the left image to the image plane Cartesian coordinates, and convert the pixel coordinates of all pixels in the original right image to the image plane Cartesian coordinates;

Step 2-2: substituting the Cartesian coordinates of the image plane of the matching feature point into the quadratic polynomial shown in formula (1), and solving the coefficient of the quadratic polynomial;

Among them, (X', Y') is the Cartesian coordinates of the image plane of the matching feature points in the original right image; (x', y') is the Cartesian coordinates of the image plane of the matching feature points in the left image, a ₀ , a ₁ , a ₂ , a ₃ , a ₄ , a ₅ , b ₀ , b ₁ , b ₂ , b ₃ , b ₄ , b ₅ are quadratic polynomial coefficients;

Step 2-3: Perform the following operations on all the pixels in the original right image to obtain the corrected right image: First, the Cartesian coordinates of the image plane (X , Y) into the quadratic polynomial shown in formula (2) to obtain the corrected image plane Cartesian coordinates (X _new , Y _new ) of the pixel, and then transform the image plane Cartesian coordinates (X _new , Y _new ) is the pixel coordinate (I _new , J _new );

Among them, (X, Y) is the Cartesian coordinates of the image plane of the pixels in the original right image; (X _new , Y _new ) is the Cartesian coordinates of the image plane of the pixels in the corrected right image; a ₀ , a ₁ , a ₂ , a ₃ , a ₄ , a ₅ , b ₀ , b ₁ , b ₂ , b ₃ , b ₄ , b ₅ are polynomial coefficients solved by step 2-2;

Step 2-4: According to the corresponding relationship between the pixel coordinates (I, J) of the pixels in the original right image and the pixel coordinates (I _new , J _new ) of the pixels in the corrected right image, divide the pixels in the corrected right image According to the category of the pixel in the corrected right image, the gray value of the pixel coordinate (I _new , J _new ) of the pixel in the corrected right image is determined. Correspondence between pixel coordinates (I, J) and pixel coordinates (I _new , J _new ) of pixels in the corrected right image, if pixel coordinates (I, J) of pixels in the original right image and pixels in the corrected right image The pixel coordinates (I _new , J _new ) of the pixel coordinates (I new , J new ) are one-to-one correspondence, then record the corrected right image pixel as a type pixel, and record the corrected pixel coordinates (I _new , J _new ) of the right image pixel Correspondence with the pixel coordinates (I, J) of the pixel in the original right image, and assign the gray value of the pixel coordinate (I, J) of the pixel in the original right image to the pixel of the pixel in the right image Coordinates (I _new , J _new ) position; if the pixel coordinates (I, J) of multiple pixel points in the original right image correspond to the pixel coordinates (I _new , J _new ) of the same pixel in the corrected right image, Then the right image pixel after the correction is recorded as a b-type pixel, and the pixel coordinates (I _new , J _new ) of the right image pixel after the correction and the pixel coordinates (I, J new ) of a plurality of pixels in the original right image are recorded. ), and after taking the average of the gray values of the pixel coordinates (I, J) of the pixels in the original right image corresponding to the pixel coordinates (I _new , J _new ) of the corrected right image pixels Assign the pixel coordinate (I _new , J _new ) position of the right image pixel point after correction; if there is no pixel point in the original right image corresponding to the pixel coordinate (I _new , J _new ) of the right image pixel point after correction, then the After class correction, the right image pixel point is recorded as the c class pixel point, and the gray value of the pixel coordinate (I _new , J _new ) position of the right image pixel point after correction is calculated using bilinear interpolation;

Step 3: Perform dense matching on the left image and the corrected right image;

Step 4: According to the category of the pixel in the left image that matches the pixel in the corrected right image, determine the pixel coordinate position of the pixel in the left image that matches the pixel in the original right image. The specific method is: for each pixel in the left image A pixel point is followed in sequence: if the matching pixel point of the pixel point in the left image in the corrected right image is a type a pixel point, then according to the coordinate correspondence relationship saved for the type a pixel point in steps 2-4, use The pixel coordinates in the original right image replace the pixel coordinates of the matching pixel in the corrected right image; if the matching pixel of the pixel in the left image in the corrected right image is a type b pixel, then the pixel in the left image Use the SIFT operator to generate feature descriptors, use the SIFT operator to generate feature descriptors for multiple pixels in the original right image corresponding to the matching pixels in the corrected right image, perform SIFT feature matching, and determine The pixel coordinates of the matching pixel of the pixel in the original right image; if the matching pixel of the pixel in the left image in the corrected right image is a class c pixel, the matching pixel in the corrected right image The pixel coordinates are converted to the rectangular coordinates of the image plane, and then according to the inverse transformation of the quadratic polynomial shown in formula (2), after solving the rectangular coordinates of the image plane of the matching pixel in the original right image, the pixels in the left image are respectively The corresponding pixel points in the original right image use the SIFT operator to generate feature descriptors, perform SIFT feature matching, and determine the pixel coordinates of the matching pixel points in the left image in the original right image.

Beneficial effects: the present invention proposes a dense matching method suitable for deformed images for the image matching problem of deformed images, which has the following advantages:

1. Compared with the existing methods for solving the matching problem of deformed images, the present invention can obtain dense matching results, provide conditions for completing fine three-dimensional surface reconstruction, and overcome the shortcomings of only using feature matching when solving the matching problem of deformed images in the past;

2. Compared with the existing dense matching method, the present invention can adapt to image matching in more situations, and can obtain better matching effects especially when the image is greatly deformed;

3. For the first time, the polynomial correction and coordinate correspondence preservation mechanism is used for the dense matching of deformed images, which provides a new solution for the dense matching of multi-source images;

4. Polynomials are used to correct the deformation between images, and the parallax-constrained search area is used to search for matching points in the corrected right image, which narrows the search range of matching points, reduces the time-consuming matching process, and improves matching efficiency;

5. The position after polynomial correction is used as the constraint condition for matching point search, which increases the reliability of the matching process, especially the reliability of the matching results in areas such as texture repetition.

Description of drawings

Fig. 1 is a flow chart of a dense matching method suitable for deformed images according to an embodiment of the present invention;

Fig. 2 (a) is a schematic diagram of the left image in the image pair to be matched according to an embodiment of the present invention;

Fig. 2 (b) is a schematic diagram of the right image in the image pair to be matched according to an embodiment of the present invention;

Fig. 3 is a schematic diagram of the corrected right image in an embodiment of the present invention;

Figure 4(a) is a schematic diagram of a 3×3 matching template in an embodiment of the present invention;

FIG. 4( b ) is a schematic diagram of a 3×3 search area in an embodiment of the present invention.

Detailed ways

An embodiment of the present invention will be described in detail below in conjunction with the accompanying drawings.

As shown in Figure 1, the dense matching method suitable for deformed images in this embodiment includes the following steps:

Step 1: Manually extract 9 pairs of matching feature points from the left image and the right image respectively. The left image and the right image are a pair of relatively deformed original images to be matched; the matching feature point pairs constitute matching feature points, which can cover the left image and The overlapping area of the right image, as shown in Figure 2, where Figure 2(a) is the left image, Figure 2(b) is the right image, and the black border in Figure 2(b) is the overlapping area between the right image and the left image, The white circle represents the location of the distribution of feature matching points;

Step 2: Use a quadratic polynomial to correct the relative deformation of the original right image to obtain the corrected right image;

Step 2-1: Convert the pixel coordinates of the matching feature points in the left image and the pixel coordinates of all the pixels in the original right image to the rectangular coordinates of the image plane;

Set the pixel coordinates of the pixels in the left image or the right image to be (I, J), wherein the origin of the pixel coordinate system is located at the upper left corner of the image, the horizontal direction of the X-axis is positive to the right, and the vertical direction of the Y-axis is positive downwards, Combining the image pixel size and the coordinates of the principal point of the image, the pixel coordinates are converted into the Cartesian coordinates of the image plane (X, Y), where the origin of the Cartesian coordinate system of the image plane is located in the center of the image, the horizontal direction of the X-axis is positive to the right, and the vertical direction of the Y-axis Up is positive, and the conversion formula is as follows:

Where: W is the width of the image, H is the height of the image, P is the pixel size, x ₀ is the abscissa of the principal point of the image, and y ₀ is the ordinate of the principal point of the image;

Step 2-2: Substituting 9 pairs of Cartesian coordinates of the image plane of the matching feature points into the quadratic polynomial shown in formula (2), and solving the coefficient of the quadratic polynomial:

Among them, (X', Y') is the Cartesian coordinates of the image plane of the matching feature points in the original right image; (x', y') is the Cartesian coordinates of the image plane of the matching feature points in the left image, a ₀ , a ₁ , a ₂ , a ₃ , a ₄ , a ₅ , b ₀ , b ₁ , b ₂ , b ₃ , b ₄ , and b ₅ are quadratic polynomial coefficients; the method for solving the quadratic polynomial coefficients in this embodiment is as follows:

Step 2-2-1. In the actual solution process, the values on the left and right sides of formula (2) cannot be absolutely the same, therefore, formula (2) is rewritten into the form of error equation:

Among them, v _x' and v _y' are the solution errors of quadratic polynomials;

Step 2-2-2, rewrite the above error equation into the form of matrix and vector: v=AZ-l, the specific formula is as follows:

in:

Substituting the image plane Cartesian coordinates of 9 pairs of matching feature points, by the method of least squares: Z=(A ^T A) ^-1 A ^T l, calculate the various coefficients of the polynomial;

Step 2-3: Perform the following operations on all pixels in the original right image to obtain the corrected right image as shown in Figure 3: First, the pixel coordinates in the original right image are (I, J) corresponding to The Cartesian coordinates of the image plane (X, Y) are substituted into the quadratic polynomial shown in formula (2), and the Cartesian coordinates of the image plane (X _new , Y _new ) of the point after correction are obtained, and then the Cartesian coordinates of the image plane (X new ) are obtained. _new , Y _new ) to pixel coordinates (I _new , J _new )

Among them, (X, Y) is the rectangular coordinates of the image plane of the pixels of the original right image; (X _new , Y _new ) is the rectangular coordinates of the image plane after correction of the pixels whose image plane coordinates are (X, Y) in the original right image Coordinates; a ₀ , a ₁ , a ₂ , a ₃ , a ₄ , a ₅ , b ₀ , b ₁ , b ₂ , b ₃ , b ₄ , and b ₅ are the polynomial coefficients to be solved in step 2-2;

The rounding of the initial pixel coordinates is specifically adding 0.5 to the abscissa and ordinate of the initial pixel coordinates and rounding down;

Step 2-4: For the right image shown in Figure 3, there are three correspondences between the original pixel coordinates (I, J) and the pixel coordinates (I _new , J _new ) due to the rounding operation. First, according to the pixel in the original right image The corresponding relationship between the pixel coordinates (I, J) of the point and the pixel coordinates (I _new , J _new ) of the pixel in the corrected right image is to divide the category of the pixel in the corrected right image, and then according to the pixel in the corrected right image The category of the point determines the gray value of the pixel coordinate (I _new , J _new ) of the pixel point in the right image after correction. The specific method is: judge the pixel coordinate (I, J ) of the pixel point in the original right image The corresponding relationship of the pixel coordinates (I _new , J _new ) of the image pixels, if the pixel coordinates (I, J) of the pixel points in the original right image and the pixel coordinates (I _new , J _new ) of the corrected right image pixels are One-to-one correspondence, then record the corrected right image pixel point as a type pixel point, and record the pixel coordinates (I _new , J _new ) of the corrected right image pixel point and the pixel coordinates of the pixel point in the original right image (I, J) corresponding relationship, assign the gray value of the pixel coordinates (I, J) position of the pixel point in the original right image to the pixel coordinates (I _new , J _new ) position of the right image pixel point after the positive; if If there are multiple pixel coordinates (I, J) of pixels in the original right image corresponding to the pixel coordinates (I _new , J _new ) of the same pixel in the corrected right image, then mark the corrected right image pixel as Be the b-type pixel, and record the correspondence between the pixel coordinates (I _new , J _new ) of the pixel of the right image after correction and the pixel coordinates (I, J) of multiple pixels in the original right image, and compare all the corrected The pixel coordinates (I _new , J _new ) of the right image pixel point correspond to the pixel coordinates (I, J) of the pixel point in the original right image. (I _new , J _new ) position; if there is no pixel coordinate (I, J) of the pixel point in the original right image corresponding to the pixel coordinate (I _new , J _new ) of the pixel point of the corrected right image, then the corrected The right image pixel is recorded as the c class pixel, and the gray value of the pixel coordinate (I _new , J _new ) position of the right image pixel after correction is calculated using bilinear interpolation;

Step 3: Perform dense matching on the left image and the corrected right image;

Step 3-1: In the left image, select an area with an odd number greater than 1 × an odd number greater than 1 as a matching template, and select the pixel at the center of the area as the pixel to be matched. In this embodiment, select the pixel coordinates in the left image The pixel point of (3, 3) is used as the pixel point to be matched, and a 3×3 size area is selected as the matching template, as shown in Figure 4(a) within the dark gray solid line frame;

Step 3-2: First, manually select a pair of matching points from the left image and the corrected right image, and calculate the left and right disparity p and the upper and lower disparity q of the left image and the corrected right image; then in the corrected right image Taking the pixel point with the pixel position (3+q, 3+p) as the center, select an odd number greater than 1 × an odd number greater than 1 as the search area, that is, the pixel coordinates include (2+q, 2+p), (2+q, 3+p), (2+q, 4+p), (3+q, 2+p), (3+q, 3+p), (3+q, 4+p), (4+q, 2+p), (4+q, 3+p), (4+q, 4+p) pixel position area, in this embodiment, p=566, q=370, the pixel of this embodiment The pixel at the position (373, 569) is used as the center of the search area, and a 3×3 area is selected as the search area, as shown in Figure 4(b) within the dotted box;

Step 3-3: Take the pixels in the search area in the corrected right image as the center pixels in turn, and use each center pixel and several surrounding pixels to form an area of the same size as the matching template as the target area, for example, take The pixel in the search area in the dotted line box shown in Figure 4(b) is located at the pixel point (2+q, 2+p) as the center, and the surrounding 8 pixel points are the same as those shown in Figure 4(a) The area of the same size as the 3×3 matching template shown in Figure 4(b) is used as the target area, as shown in Figure 4(b) within the gray solid line frame; use the formula (6) to calculate the pixel points in each target area and the pixel points in the matching template The correlation coefficient; then select the pixel with the largest correlation coefficient as the matching pixel in the corrected right image of the pixel to be matched;

The formula for calculating the correlation coefficient is as follows:

In the formula _: ρ is the correlation coefficient; (c, r) is the difference between the pixel coordinates of the central pixel point and the pixel coordinates of the pixel point to be matched; The gray value of the coordinate (i, j) position; g′ _{i, j} is the gray value of the pixel coordinate (i, j) position when the upper left corner of the corrected right image target area is the coordinate origin of the pixel coordinate system; m , n are the height and width of the matching template respectively; g _{i+r, j+c} are the grayscale of the pixel coordinate (i+r, j+c) position when the upper left corner of the left image matching template is taken as the coordinate origin of the pixel coordinate system value; g' _{i+r, j+c} is the grayscale value of the pixel coordinate (i+r, j+c) position when the upper left corner of the corrected right image target area is the coordinate origin of the pixel coordinate system; the height and width are in pixels, such as a 3×3 matching area, which means that the matching area is 3 pixels high and 3 pixels wide;

Step 4: According to the category of the pixel in the left image that matches the pixel in the corrected right image, determine the pixel coordinate position of the pixel in the left image that matches the pixel in the original right image. The specific method is: for each pixel in the left image Perform the following operations on a pixel point in turn: if the matching pixel point of the pixel point in the left image in the corrected right image is a type a pixel point, according to the coordinate correspondence saved for the type a pixel point in steps 2-4, use the original The pixel coordinates in the right image replace the pixel coordinates of the matching pixel in the corrected right image; if the pixel in the left image matches the pixel in the corrected right image as a b-type pixel, the pixel in the left image uses The SIFT operator generates feature descriptors, and uses the SIFT operator to generate feature descriptors for multiple pixels in the original right image corresponding to the matching pixels in the corrected right image, and performs SIFT feature matching to determine the pixels in the left image The pixel coordinates of the matching pixel in the original right image; if the matching pixel of the pixel in the left image in the corrected right image is a class c pixel, the pixel coordinate of the matching pixel in the corrected right image Convert to the Cartesian coordinates of the image plane, and then according to the inverse transformation of the quadratic polynomial shown in formula (5), after solving the Cartesian coordinates of the image plane of the matching pixel in the original right image, the pixels in the left image and the original Corresponding pixel in the right image adopts SIFT operator to generate feature descriptor, carries out SIFT feature matching, determines the pixel coordinates of the matching pixel of pixel in the original right image in the left image, and the corresponding pixel in the original right image , refers to the above-mentioned obtained after converting the rectangular coordinates of the image plane in the original right image into unrounded pixel coordinates, and then rounding up or down the abscissa and ordinate of the aforementioned unrounded pixel coordinates 4 pixels around unrounded pixel coordinates.

Claims (1)

1. a kind of dense matching method of suitable deformation pattern, it is characterised in that：The method includes：

Step 1：Matching characteristic point pair is manually extracted in left image and right image respectively；

The left image and right image are the original image to be matched of a pair of relative deformation；The matching characteristic point is to being more than 5 pairs； The matching characteristic point can cover the overlapping region of left image and right image to constituting matching characteristic point；

Step 2：The relative deformation that original right image is corrected using quadratic polynomial, the right image after being corrected；It specifically includes Following steps：

Step 2-1：The pixel coordinate of matching characteristic point in left image is converted into image plane rectangular co-ordinate, it will be in original right image The pixel coordinate of all pixels point is converted to image plane rectangular co-ordinate；

Step 2-2：The image plane rectangular co-ordinate of matching characteristic point is substituted into quadratic polynomial shown in formula (1), it is secondary to solve this Polynomial coefficient；

Wherein, (X ', Y ') is the image plane rectangular co-ordinate of matching characteristic point in original right image；(x ', y ') is in left image Image plane rectangular co-ordinate with characteristic point, a₀、a₁、a₂、a₃、a₄、a₅、b₀、b₁、b₂、b₃、b₄、b₅For quadratic polynomial coefficient；

Step 2-3：All pixels point in original right image is proceeded as follows, the right image after being corrected：It first, will be former Pixel coordinate is secondary shown in corresponding image plane rectangular co-ordinate (X, Y) the substitution formula (2) of pixel of (I, J) in beginning right image Multinomial, the image plane rectangular co-ordinate (X of the pixel after being corrected_new, Y_new), then again by the image plane rectangular co-ordinate (X_new, Y_new) be converted to pixel coordinate (I_new, J_new)；

Wherein, (X, Y) is the image plane rectangular co-ordinate of original right image slices vegetarian refreshments；(X_new, Y_new) in the right image after correction The image plane rectangular co-ordinate of pixel；a₀、a₁、a₂、a₃、a₄、a₅、b₀、b₁、b₂、b₃、b₄、b₅For the multinomial solved by step 2-2 Coefficient；

Step 2-4：It is sat according to the pixel coordinate (I, J) of pixel in original right image and pixel pixel in right image after correction Mark (I_new, J_new) correspondence, divide correct after right image pixel classification, according to pixel in the right image after correction The classification of point determines correct after in right image pixel pixel coordinate (I_new, J_new) position gray value；Specific method is：

Judge the pixel coordinate (I, J) of pixel in original right image and the pixel coordinate (I for correcting rear right image slices vegetarian refreshments_new, J_new) correspondence, if in original right image pixel pixel coordinate (I, J) with correct rear right image slices vegetarian refreshments pixel Coordinate (I_new, J_new) be one-to-one relationship, then such correction rear right image slices vegetarian refreshments is denoted as a class pixels, and record and entangle Pixel coordinate (the I of positive rear right image slices vegetarian refreshments_new, J_new) with original right image in pixel pixel coordinate (I, J) correspondence The gray value of position pixel coordinate (I, J) of pixel in original right image is assigned to correct rear right image slices vegetarian refreshments by relationship Pixel coordinate (I_new, J_new) position；If the pixel coordinate (I, J) for having multiple pixels in original right image and right figure after correction Pixel coordinate (the I of the same pixel as in_new, J_new) corresponding, then such correction rear right image slices vegetarian refreshments is denoted as b class pixels Point, and record the pixel coordinate (I for correcting rear right image slices vegetarian refreshments_new, J_new) with original right image in multiple pixels pixel The correspondence of coordinate (I, J), by all pixel coordinate (I with correction rear right image slices vegetarian refreshments_new, J_new) corresponding original right The gray value of position pixel coordinate (I, J) of pixel is assigned to correct the pixel of rear right image slices vegetarian refreshments after being averaged in image Coordinate (I_new, J_new) position；If there is no pixel in original right image and correcting the pixel coordinate of rear right image slices vegetarian refreshments (I_new, J_new) corresponding, then such correction rear right image slices vegetarian refreshments is denoted as c class pixels, is calculated using bilinear interpolation method Correct the pixel coordinate (I of rear right image slices vegetarian refreshments_new, J_new) position gray value；

Step 3：Right image to left image and after correcting carries out dense matching；

Step 4：Determine that the pixel coordinate of matched pixel point of the pixel in original right image in left image, specific method are：

Each pixel in left image is proceeded as follows successively：If right image of the pixel after correction in left image In matched pixel point be a class pixels, according to the coordinate correspondence relationship for being directed to a class pixels in step 2-4 and preserving, use is original Pixel coordinate in right image replaces the pixel coordinate of matched pixel point in the right image after correcting；If pixel in left image Matched pixel point in right image after correction is b class pixels, and spy is generated using SIFT operators to pixel in left image Descriptor is levied, according to the coordinate correspondence relationship for being directed to the preservation of b class pixels in step 2-4, in the right image after correction Feature descriptor is generated using SIFT operators with multiple pixels in the corresponding original right image of pixel, carries out SIFT feature Matching, determines matched pixel point pixel coordinate of the pixel in original right image in left image；If pixel in left image Matched pixel point in right image after correction is c class pixels, by the matched pixel point in the right image after the correction Pixel coordinate is converted to image plane rectangular co-ordinate, then according still further to the inverse transformation of quadratic polynomial shown in formula (2), solves this After image plane rectangular co-ordinate with pixel in original right image, respectively to phase in pixel in left image and original right image The pixel answered generates feature descriptor using SIFT operators, carries out SIFT feature matching, determines that pixel is in original in left image The pixel coordinate of matched pixel point in beginning right image.