Noise robust corner detection method based on multi-scale local similarity

Abstract

A noise robust corner detection method based on multi-scale local similarity is composed of the steps of extracting noise bit planes and residual images, constructing left and right sub-window filters, determining directional derivative response vectors, extracting edge chain code sets, determining local similarity of response vectors, determining corner point measurement of multi-scale local similarity, and determining noise robust corner points. A comparative simulation experiment was conducted using the method of the present invention and a multi-scale differential ratio corner detection method. The experimental results show that under the same test image and evaluation criteria, the method of the present invention is compared with the comparative method. In the case of mixed noise, the formation of pseudo corner points is suppressed, the influence of edge burrs on corner point detection is avoided, and the corner point information in the image can be better extracted. The method of the present invention shows good performance in terms of correct detection rate and false detection rate, improves the accuracy of corner point positioning, and can be used for noise robust corner point detection of multi-scale local similarity.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a noise robust corner point detection method based on multi-scale local similarity. Background Art

[0002] Corner points are one of the most stable features in an image and are widely used in computer vision and image processing, such as image registration, object recognition, and image retrieval. Currently, existing corner detection methods are mainly divided into three categories, including intensity-based corner detection methods, template matching-based corner detection methods, and edge contour-based corner detection methods.

[0003] The intensity-based corner detector mainly uses the first-order or second-order partial derivatives of the original grayscale value of the image to detect corners, but the differential operator is sensitive to noise. The model-based corner detector detects corners based on similarity by matching the image with a pre-designed model. The model is difficult to cover all types of corners. The edge-based corner detector is based on edge detection and uses the characteristics of corners on the edge to design the detector. Corner points are often located at the intersection of two or more edge contours, which is different from the linear edge with a simple structure. The curvature scale space-based method detects corners based on the local contour curvature. It is very sensitive to noise on the local grayscale changes on the contour, and there will be missed detections and false detections. On this basis, the anisotropic directional derivative filter embeds the directional intensity change into the contour-based corner detection framework, combining the advantages of the contour-based and intensity-based corner detection methods, and can more accurately describe the local differential change information at the corner point, improving the efficiency of corner detection, but still sensitive to mixed noise. When there is a difference in contrast between different regions in the image, it is easy to generate strong responses at some non-corner points, resulting in pseudo corners. In addition, this type of method only uses the directional differential information of each edge pixel itself, and the constructed corner point measurement is easily affected by edge glitches. Summary of the invention

[0004] The technical problem to be solved by the present invention is to overcome the deficiencies of the above-mentioned prior art and provide a noise-robust corner point detection method based on multi-scale local similarity which is simple in method, robust to mixed noise and has high corner point positioning accuracy.

[0005] The technical solution adopted to solve the above technical problems consists of the following steps:

[0006] (1) Extracting the noise bit plane and residual image

[0007] The original grayscale image I(n) is separated into the noise plane X(n) and the residual image D(n) according to the following formula:

[0008]

[0009]

[0010] where n is the pixel position in the image and τ is the threshold, τ∈{1,2,3}.

[0011] (2) Construct left and right sub-window filters

[0012] The anisotropic Gaussian directional derivative filters are constructed as follows: The left subwindow filter and right subwindow filter

[0013]

[0014]

[0015]

[0016]

[0017]

[0018]

[0019] where |·| is the absolute value, is an anisotropic Gaussian directional filter, σ is the Gaussian kernel scale factor, σ∈[1,6], ρ is the anisotropy factor, ρ∈[3,9], θ k is the kth discrete rotation direction, k∈{1,2,...,M}, M is the total number of discrete rotation direction samples, M∈{16,17,...,64}, The direction of rotation is θ k The rotation matrix of .

[0020] (3) Determine the directional derivative response vector

[0021] The directional derivative response vector ν(n) is determined as follows:

[0022]

[0023]

[0024]

[0025]

[0026]

[0027]

[0028] in is the anisotropic Gaussian directional derivative response, is the convolution operation, yes Weighted filtering of D(n) yes Weighted filtering of D(n) yes Weighted filtering of X(n), yes Apply weighted filtering to X(n).

[0029] (4) Extract edge chain code set

[0030] The image after noise suppression of X(n) and D(n) is obtained as follows:

[0031]

[0032]

[0033] where G(n) is a two-dimensional Gaussian filter, ||·|| 2 is the Euclidean norm, α is the standard deviation,

[0034] For images The edge graph E is extracted using the Canny edge detection method, and the edge chain code set C of the edge graph E is obtained as follows:

[0035] C={P 1 ,P 2 ,...,P i}

[0036] P i ={n 1i ,n 2i ,...,n q}

[0037] Where P i is the ith edge chain code in the edge chain code set C, i∈{1,2,...,Q}, Q is the total number of edge chain codes in the edge chain code set C, n q is the edge chain code P i The qth pixel position in i ,2 i ,...,S i},S i is the edge chain code P i The total number of pixels in the image.

[0038] (5) Determine the local similarity of the response vector

[0039] In the edge chain code P i n q The left support region and the right support region are formed by N pixels on both sides respectively. The average response vector ν in the left support region is determined by the following formula: L (n q ) and the average response vector ν in the right support region R (n q ) N (n q ):

[0040]

[0041]

[0042]

[0043] Where T represents transpose, n q-j is in n q The pixel position in the left support region, n q+j is in n q The pixel position in the right support region of , N is the scale of the support region, N∈{2,3,4}.

[0044] The directional derivative response vector ν(n q ) and the average response vector in the entire support region Similarity between γ N (n q ):

[0045]

[0046]

[0047] (6) Determine the corner point measure of multi-scale local similarity

[0048] Repeat step (5) and replace the support region scale N in step (5) with the support region scale 2N and the support region scale 3N, respectively, to obtain ξ 2N (n q ) and γ 2N (n q ), 3N (n q ) and γ 3N (n q ), press the formula to get q The corner point measure Γ(n q ):

[0049]

[0050] η N (n q )=1-ξ N (n q )γ N (n q )

[0051] η 2N (n q )=1-ξ 2N (n q )γ 2N (n q )

[0052] η 3N (n q )=1-ξ 3N (n q )γ 3N (n q )

[0053] where η N (n q ) is in n q The corner point measure of the support region with scale N, η 2N (n q ) is in n q The corner point measure of the support region with a scale of 2N, η 3N (n q ) is in n q The support region scale is 3N.

[0054] (7) Determine noise-robust corner points

[0055] The noise robust corner point set U is determined as follows:

[0056] U={n q |Γ(n q )≥χ}

[0057] Where χ is the corner point measure decision value, χ∈[0.03,0.1].

[0058] In the step (1) of extracting the noise bit plane and the residual image of the present invention, the optimal value of the threshold τ is 2.

[0059] In the step (2) of constructing left and right sub-window filters of the present invention, the optimal value of the Gaussian kernel scale factor σ is 3, the optimal value of the anisotropy factor ρ is 6, and the optimal value of the total number of discrete rotation direction samples M is 32.

[0060] In the step (4) of extracting the edge chain code set of the present invention, the optimal value of the standard deviation α is 1.

[0061] In the step (5) of determining the local similarity of the response vector of the present invention, the optimal value of the scale N of the support area is 3.

[0062] In the step (7) of determining the noise robust corner point of the present invention, the optimal value of the corner point measurement decision value χ is 0.055.

[0063] The present invention separates the noise plane according to the influence of noise on the original image pixels, suppresses the influence of noise on the original image, and avoids the influence of pulse noise. The separated noise plane and the residual image are linearly weighted filtered to avoid the influence of Gaussian noise. The similarity between the response vectors of edge pixels and their left and right supporting areas is used to construct the corner point measurement, so that the final measurement is invariant to the image contrast, preventing the contrast difference between different areas in the image from causing a strong response at non-corner points and forming pseudo corner points. By fusing the corner point measurements at three scales, the noise robustness is enhanced, the corner point positioning accuracy is improved, and the influence of edge burrs on the constructed corner point measurement is avoided.

[0064] The method of the present invention is used to conduct a comparative simulation experiment with a multi-scale differential ratio corner detection method. The experimental results show that under the same test image and evaluation criteria, the method of the present invention is compared with the comparative method. In the case of mixed noise, the formation of pseudo corners is suppressed, the influence of edge burrs on corner detection is avoided, and the corner information in the image can be better extracted. The method of the present invention shows good performance in terms of correct detection rate and false detection rate, improves the accuracy of corner location, improves the accuracy and quality of corner detection, and can be used for noise-robust corner detection of multi-scale local similarity. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a flow chart of Example 1 of the method of the present invention.

[0066] Figure 2 It is the noise-free output image of the contrast method on the building block image.

[0067] Figure 3 It is the noise-free output image of the method of the present invention on the building block picture.

[0068] Figure 4 The output image of the comparison method adding 1% salt and pepper noise to the building block image.

[0069] Figure 5 This is the output image of the method of the present invention in which 1% salt and pepper noise is added to the building block image.

[0070] Figure 6 The output image of the comparison method adding 5% salt and pepper noise to the building block image.

[0071] Figure 7This is the output image of the method of the present invention in which 5% salt and pepper noise is added to the building block image.

[0072] Figure 8 is the noise-free output image of the comparison method in the laboratory picture.

[0073] Fig. 9 The method of the present invention outputs images without noise in laboratory pictures.

[0074] Fig.10 This is the output image of the comparison method after adding 1% salt and pepper noise to the laboratory image.

[0075] Fig.11 This is the output image of the method of the present invention when 1% salt and pepper noise is added to the laboratory image.

[0076] Fig.12 This is the output image of the comparison method after adding 5% salt and pepper noise to the laboratory image.

[0077] Fig.13 This is the output image of the method of the present invention when 5% salt and pepper noise is added to the laboratory image.

[0078] Fig.14 This is the output image of the comparison method after adding Gaussian noise with a standard deviation of 15 to the building block image.

[0079] Fig.15 This is the output image of the method of the present invention in which Gaussian noise with a standard deviation of 15 is added to the building block image.

[0080] Fig.16 The output image of the comparison method is a mixed noise of Gaussian noise with a standard deviation of 15 and 1% salt and pepper noise added to the building block image.

[0081] Fig.17 The output image of the method of the present invention is a mixed noise of Gaussian noise with a standard deviation of 15 and 1% salt and pepper noise added to the building block picture.

[0082] Fig.18 This is the output image of the comparison method after adding Gaussian noise with a standard deviation of 15 to the laboratory image.

[0083] Fig.19 This is the output image of the method of the present invention when Gaussian noise with a standard deviation of 15 is added to the laboratory picture.

[0084] Fig. 20 This is the output image of the comparison method after adding Gaussian noise with a standard deviation of 15 and 1% salt and pepper noise mixed noise to the laboratory picture.

[0085] Fig.21 The output image of the method of the present invention is a mixed noise of Gaussian noise with a standard deviation of 15 and 1% salt and pepper noise added to the laboratory picture. DETAILED DESCRIPTION

[0086] The present invention is further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the following embodiments.

[0087] Example 1

[0088] exist Figure 1 In the embodiment, the noise robust corner point detection method based on multi-scale local similarity is composed of the following steps:

[0089] (1) Extracting the noise bit plane and residual image

[0090] The original grayscale image I(n) is separated into the noise plane X(n) and the residual image D(n) according to the following formula:

[0091]

[0092]

[0093] Where n is the pixel position in the image, τ is the threshold, τ∈{1,2,3}, and the value of τ in this embodiment is 2.

[0094] In this step, a noise plane separation is used to suppress the influence of noise on the original image and avoid the influence of pulse noise.

[0095] (2) Construct left and right sub-window filters

[0096] The anisotropic Gaussian directional derivative filters are constructed as follows: The left subwindow filter and right subwindow filter

[0097]

[0098]

[0099]

[0100]

[0101]

[0102]

[0103] where |·| is the absolute value, is an anisotropic Gaussian directional filter, σ is the Gaussian kernel scale factor, σ∈[1,6], the value of σ in this embodiment is 3, ρ is the anisotropy factor, ρ∈[3,9], the value of ρ in this embodiment is 6, θk is the kth discrete rotation direction, k∈{1,2,...,M}, M is the total number of discrete rotation direction samples, M∈{16,17,...,64}, and the value of M in this embodiment is 32. The direction of rotation is θ k The rotation matrix of .

[0104] (3) Determine the directional derivative response vector

[0105] The directional derivative response vector ν(n) is determined as follows:

[0106]

[0107]

[0108]

[0109]

[0110]

[0111]

[0112] in is the anisotropic Gaussian directional derivative response, is the convolution operation, yes Weighted filtering of D(n) yes Weighted filtering of D(n) yes Weighted filtering of X(n), yes Apply weighted filtering to X(n).

[0113] In this step, linear weighted filtering is performed on the separated noise plane and the residual image to avoid the influence of Gaussian noise.

[0114] (4) Extract edge chain code set

[0115] The image after noise suppression of X(n) and D(n) is obtained as follows:

[0116]

[0117]

[0118] where G(n) is a two-dimensional Gaussian filter, ||·|| 2 is the Euclidean norm, α is the standard deviation, In this embodiment, the value of α is 1.

[0119] For images The edge graph E is extracted using the Canny edge detection method, and the edge chain code set C of the edge graph E is obtained as follows:

[0120] C={P 1 ,P 2 ,...,P i}

[0121]

[0122] Where P i is the ith edge chain code in the edge chain code set C, i∈{1,2,...,Q}, Q is the total number of edge chain codes in the edge chain code set C, n q is the edge chain code P i The qth pixel position in i ,2 i ,...,S i},S i is the edge chain code P i The total number of pixels in the image.

[0123] (5) Determine the local similarity of the response vector

[0124] In the edge chain code P i n q The left support region and the right support region are formed by N pixels on both sides respectively. The average response vector ν in the left support region is determined by the following formula: L (n q ) and the average response vector ν in the right support region R (n q ) N (n q ):

[0125]

[0126]

[0127]

[0128] Where T represents transpose, n q-j is in n q The pixel position in the left support region, n q+j is in n q The pixel position in the right support area of ​​, N is the scale of the support area, N∈{2,3,4}, and the value of N in this embodiment is 3.

[0129] The directional derivative response vector ν(n q) and the average response vector in the entire support region Similarity between γ N (n q ):

[0130]

[0131]

[0132] This step uses the similarity between the response vectors of edge pixels and their left and right supporting regions to construct the corner point measure, so that the final measure is invariant to the image contrast, preventing the contrast difference between different regions in the image from causing strong response at non-corner points and forming pseudo corner points.

[0133] (6) Determine the corner point measure of multi-scale local similarity

[0134] Repeat step (5) and replace the support region scale N in step (5) with the support region scale 2N and the support region scale 3N, respectively, to obtain ξ 2N (n q ) and γ 2N (n q ), 3N (n q ) and γ 3N (n q ), press the formula to get q The corner point measure Γ(n q ):

[0135]

[0136] η N (n q )=1-ξ N (n q )γ N (n q )

[0137] η 2N (n q )=1-ξ 2N (n q )γ 2N (n q )

[0138] η 3N (n q )=1-ξ 3N (n q )γ 3N (n q )

[0139] where η N (nq ) is in n q The corner point measure of the support region with scale N, η 2N (n q ) is in n q The corner point measure of the support region with a scale of 2N, η 3N (n q ) is in n q The support region scale is 3N.

[0140] This step enhances noise robustness and improves corner location accuracy by fusing corner measurements at three scales, thus avoiding the influence of edge glitches on the constructed corner measurements.

[0141] (7) Determine noise-robust corner points

[0142] The noise robust corner point set U is determined as follows:

[0143] U={n q |Γ(n q )≥χ}

[0144] Wherein χ is the corner point measurement decision value, χ∈[0.03,0.1], and the value of χ in this embodiment is 0.055.

[0145] Example 2

[0146] The noise robust corner point detection method based on multi-scale local similarity of this embodiment consists of the following steps:

[0147] (1) Extracting the noise bit plane and residual image

[0148] The original grayscale image I(n) is separated into the noise plane X(n) and the residual image D(n) according to the following formula:

[0149]

[0150]

[0151] Where n is the pixel position in the image, τ is the threshold, τ∈{1,2,3}, and the value of τ in this embodiment is 1.

[0152] (2) Construct left and right sub-window filters

[0153] The anisotropic Gaussian directional derivative filters are constructed as follows: The left subwindow filter and right subwindow filter

[0154]

[0155]

[0156]

[0157]

[0158]

[0159]

[0160] where |·| is the absolute value, is an anisotropic Gaussian directional filter, σ is the Gaussian kernel scale factor, σ∈[1,6], in this embodiment, σ is 1, ρ is the anisotropy factor, ρ∈[3,9], in this embodiment, ρ is 3, θ k is the kth discrete rotation direction, k∈{1,2,...,M}, M is the total number of discrete rotation direction samples, M∈{16,17,...,64}, and the value of M in this embodiment is 16. The direction of rotation is θ k The rotation matrix of .

[0161] (3) Determine the directional derivative response vector

[0162] This step is the same as in Example 1.

[0163] (4) Extract edge chain code set

[0164] The image after noise suppression of X(n) and D(n) is obtained as follows:

[0165]

[0166]

[0167] where G(n) is a two-dimensional Gaussian filter, ||·|| 2 is the Euclidean norm, α is the standard deviation, The value of α in this embodiment is

[0168] The other steps of this step are the same as those in Example 1.

[0169] (5) Determine the local similarity of the response vector

[0170] In the edge chain code P i n q The left support region and the right support region are formed by N pixels on both sides respectively. The average response vector ν in the left support region is determined by the following formula: L (n q ) and the average response vector ν in the right support regionR (n q ) N (n q ):

[0171]

[0172]

[0173]

[0174] Where T represents transpose, n q-j is in n q The pixel position in the left support region, n q+j is in n q The pixel position in the right support area of ​​, N is the scale of the support area, N∈{2,3,4}, and the value of N in this embodiment is 2.

[0175] The other steps of this step are the same as those in Example 1.

[0176] (6) Determine the corner point measure of multi-scale local similarity

[0177] This step is the same as in Example 1.

[0178] (7) Determine noise-robust corner points

[0179] The noise robust corner point set U is determined as follows:

[0180] U={n q |Γ(n q )≥χ}

[0181] Wherein χ is the corner point measurement decision value, χ∈[0.03,0.1], and the value of χ in this embodiment is 0.03.

[0182] Example 3

[0183] The noise robust corner point detection method based on multi-scale local similarity of this embodiment consists of the following steps:

[0184] (1) Extracting the noise bit plane and residual image

[0185] The original grayscale image I(n) is separated into the noise plane X(n) and the residual image D(n) according to the following formula:

[0186]

[0187]

[0188] Where n is the pixel position in the image, τ is the threshold, τ∈{1,2,3}, and the value of τ in this embodiment is 3.

[0189] (2) Construct left and right sub-window filters

[0190] The anisotropic Gaussian directional derivative filters are constructed as follows: The left subwindow filter and right subwindow filter

[0191]

[0192]

[0193]

[0194]

[0195]

[0196]

[0197] where |·| is the absolute value, is an anisotropic Gaussian directional filter, σ is the Gaussian kernel scale factor, σ∈[1,6], the value of σ in this embodiment is 6, ρ is the anisotropy factor, ρ∈[3,9], the value of ρ in this embodiment is 9, θ k is the kth discrete rotation direction, k∈{1,2,...,M}, M is the total number of discrete rotation direction samples, M∈{16,17,...,64}, and the value of M in this embodiment is 64. The direction of rotation is θ k The rotation matrix of .

[0198] (3) Determine the directional derivative response vector

[0199] This step is the same as in Example 1.

[0200] (4) Extract edge chain code set

[0201] The image after noise suppression of X(n) and D(n) is obtained as follows:

[0202]

[0203]

[0204] where G(n) is a two-dimensional Gaussian filter, ||·|| 2 is the Euclidean norm, α is the standard deviation, The value of α in this embodiment is

[0205] The other steps of this step are the same as those in Example 1.

[0206] (5) Determine the local similarity of the response vector

[0207] In the edge chain code P i n q The left support region and the right support region are formed by N pixels on both sides respectively. The average response vector ν in the left support region is determined by the following formula: L (n q ) and the average response vector ν in the right support region R (n q ) N (n q ):

[0208]

[0209]

[0210]

[0211] Where T represents transpose, n q-j is in n q The pixel position in the left support region, n q+j is in n q The pixel position in the right support area of ​​, N is the scale of the support area, N∈{2,3,4}, and the value of N in this embodiment is 4.

[0212] The other steps of this step are the same as those in Example 1.

[0213] (6) Determine the corner point measure of multi-scale local similarity

[0214] This step is the same as in Example 1.

[0215] (7) Determine noise-robust corner points

[0216] The noise robust corner point set U is determined as follows:

[0217] U={n q |Γ(n q )≥χ}

[0218] Wherein χ is the corner point measurement decision value, χ∈[0.03,0.1], and the value of χ in this embodiment is 0.1.

[0219] In order to verify the beneficial effects of the present invention, the inventors conducted experiments on images in the OpenCV corner detection image library using the noise robust corner detection method based on multi-scale local similarity of Example 1 of the present invention. Various test situations are as follows.

[0220] 1. Experimental conditions

[0221] The experimental test environment is a Dell computer with Windows 10 (64) as the operating system, which is configured with an Intel (R) Core (TM) i5-7200U processor, a quad-core CPU and 8GB of memory. The experimental operations are carried out on the MATLAB R2016a platform.

[0222] 2. Test pictures

[0223] A comparative test was conducted using one picture of building blocks and one picture of a laboratory.

[0224] 3. Experimental methods

[0225] (1) The method of Example 1 of the present invention (hereinafter referred to as the present method) and the multi-scale differential ratio corner detection method (hereinafter referred to as the comparison method) were used to perform corner comparison detection experiments on the test images under the conditions of no noise, adding 1% density salt and pepper noise, and adding 5% density salt and pepper noise. The experimental results are shown in Table 1. Figure 2-Figure 13 Table 1 gives the number of registered corner points and erroneous corner points. Figure 2-Figure 7 This is the result of the building block test. Figure 8-Figure 13 This is a graph of laboratory test results.

[0226] Table 1 Experimental results under different density of salt and pepper noise

[0227]

[0228] From Table 1, Figure 2-Figure 7 It can be seen that the number of registered corner points of the method of the present invention and the comparison method in the noise-free building block image is 56, and the number of wrong corner points is 1; when 1% salt and pepper noise is added to the building block image, a small number of pseudo corners appear in the comparison method, the number of registered corner points is 54, and the number of wrong corner points is 8. Compared with the noise-free case, the method of the present invention maintains a relatively stable result, with the number of registered corner points being 57 and the number of wrong corner points being 1; when 5% salt and pepper noise is added to the building block image, a large number of pseudo corners appear in the comparison method, the number of registered corner points is 54, and the number of wrong corner points is 69. Compared with the noise-free case, the method of the present invention can still maintain a relatively stable corner detection result, with the number of registered corner points being 57 and the number of wrong corner points being 1, and overall shows a strong ability to suppress mixed noise and a relatively stable corner detection accuracy.

[0229] From Table 1, Figure 8-Figure 13It can be seen that the number of registered corner points of the method of the present invention and the comparison method in the noise-free laboratory image is 196, and the number of erroneous corner points is 61; when 1% salt and pepper noise is added to the laboratory image, a small number of pseudo corners appear in the comparison method, the number of registered corner points is 189, and the number of erroneous corner points is 90. Compared with the noise-free case, the method of the present invention maintains a relatively stable result, with the number of registered corner points being 189 and the number of erroneous corner points being 57; when 5% salt and pepper noise is added to the laboratory image, a large number of pseudo corners appear in the comparison method, with the number of registered corner points being 188 and the number of erroneous corner points being 146. Compared with the noise-free case, the method of the present invention can still maintain a relatively stable corner detection result, with the number of registered corner points being 196 and the number of erroneous corner points being 71, and overall it shows a strong ability to suppress mixed noise and a relatively stable corner detection accuracy.

[0230] (2) The present invention is used to perform corner point comparison detection experiments on the above two images under the conditions of no noise, Gaussian noise with a standard deviation of 15, 1% density salt and pepper noise, and mixed noise with a standard deviation of 15 Gaussian noise. The experimental results are shown in Table 2. Figure 14-Figure 21 Table 2 shows the number of matched corner points and wrong corner points obtained under different conditions for two test images. Figure 14-17 This is the result of the building block test. Figure 18-Figure 21 This is a picture of the laboratory test results.

[0231] Table 2 Experimental results under Gaussian noise and mixed noise

[0232]

[0233] From Table 2, Figure 14-17 It can be seen that the number of registered corner points of the method of the present invention and the comparison method in the noise-free building block image is 56, and the number of wrong corner points is 1; when 1% salt and pepper noise is added to the building block image, a small number of pseudo corners appear in the comparison method, the number of registered corner points is 57, and the number of wrong corner points is 7. Compared with the noise-free case, the method of the present invention maintains a relatively stable result, with the number of registered corner points being 57 and the number of wrong corner points being 2; when 5% salt and pepper noise is added to the building block image, a large number of pseudo corners appear in the comparison method, the number of registered corner points is 52, and the number of wrong corner points is 21. Compared with the noise-free case, the method of the present invention can still maintain a relatively stable corner detection result, with the number of registered corner points being 56 and the number of wrong corner points being 3, and overall it shows a strong ability to suppress mixed noise and a relatively stable corner detection accuracy.

[0234] From Table 2, Figure 18-Figure 21It can be seen that the number of registered corner points of the method of the present invention and the comparison method in the noise-free laboratory image is 196, and the number of erroneous corner points is 61; when 1% salt and pepper noise is added to the laboratory image, a small number of pseudo corners appear in the comparison method, the number of registered corner points is 199, and the number of erroneous corner points is 81. Compared with the noise-free case, the method of the present invention maintains a relatively stable result, with the number of registered corner points being 199 and the number of erroneous corner points being 70; when 5% salt and pepper noise is added to the laboratory image, a large number of pseudo corners appear in the comparison method, with the number of registered corner points being 190 and the number of erroneous corner points being 120. Compared with the noise-free case, the method of the present invention can still maintain a relatively stable corner detection result, with the number of registered corner points being 192 and the number of erroneous corner points being 72, and overall it shows a strong ability to suppress mixed noise and a relatively stable corner detection accuracy.

[0235] 4. Conclusion

[0236] Under the same test image and evaluation criteria, the method of the present invention, compared with the comparative method, suppresses the formation of pseudo corners in the case of mixed noise, avoids the influence of edge burrs on corner detection, and can better extract corner information in the image. The method of the present invention shows good performance in terms of correct detection rate and false detection rate, improves the accuracy of corner location, and improves the accuracy and quality of corner detection.

Claims

1. A noise-robust corner detection method based on multi-scale local similarity, characterized by The following steps consist of: (1) Extracting the noise bit plane and residual image The original grayscale image I(n) is separated into the noise plane X(n) and the residual image D(n) according to the following formula: Where n is the pixel position in the image, τ is the threshold, τ∈{1,2,3}; (2) Construct left and right sub-window filters The anisotropic Gaussian directional derivative filters are constructed as follows: The left subwindow filter and right subwindow filter where |·| is the absolute value, is an anisotropic Gaussian directional filter, σ is the Gaussian kernel scale factor, σ∈[1,6], ρ is the anisotropy factor, ρ∈[3,9], θ k is the kth discrete rotation direction, k∈{1,2,...,M}, M is the total number of discrete rotation direction samples, M∈{16,17,...,64}, The direction of rotation is θ k The rotation matrix of (3) Determine the directional derivative response vector The directional derivative response vector ν(n) is determined as follows: in is the anisotropic Gaussian directional derivative response, is the convolution operation, yes Weighted filtering of D(n) yes Weighted filtering of D(n) yes Weighted filtering of X(n), yes Apply weighted filtering to X(n); (4) Extract edge chain code set The image after noise suppression of X(n) and D(n) is obtained as follows: where G(n) is a two-dimensional Gaussian filter, ||·||2 is the Euclidean norm, α is the standard deviation, For images The edge graph E is extracted using the Canny edge detection method, and the edge chain code set C of the edge graph E is obtained as follows: C={P1,P2,...,P i } Where P i is the ith edge chain code in the edge chain code set C, i∈{1,2,...,Q}, Q is the total number of edge chain codes in the edge chain code set C, n q is the edge chain code P i The qth pixel position in i ,2 i ,...,S i },S i is the edge chain code P i The total number of pixels in (5) Determine the local similarity of the response vector In the edge chain code P i n q The left support region and the right support region are formed by N pixels on both sides respectively. The average response vector ν in the left support region is determined by the following formula: L (n q ) and the average response vector ν in the right support region R (n q ) N (n q ): Where T represents transpose, n q-j is in n q The pixel position in the left support region, n q+j is in n q The pixel position in the right support region, N is the scale of the support region, N∈{2,3,4}; The directional derivative response vector ν(n q ) and the average response vector in the entire support region Similarity between γ N (n q ): (6) Determine the corner point measure of multi-scale local similarity Repeat step (5) and replace the support region scale N in step (5) with the support region scale 2N and the support region scale 3N, respectively, to obtain ξ 2N (n q ) and γ 2N (n q )、ξ 3N (n q ) and γ 3N (n q ), press the formula to get q The corner point measure Γ(n q ): or N (n q )=1-ξ N (n q )c N (n q ) or 2N (n q )=1-ξ 2N (n q )c 2N (n q ) or 3N (n q )=1-ξ 3N (n q )c 3N (n q ) where η N (n q ) is in n q The corner point measure of the support region with scale N, η 2N (n q ) is in n q The corner point measure of the support region with a scale of 2N, η 3N (n q ) is in n q The corner point measure at the support region scale is 3N; (7) Determine noise-robust corner points The noise robust corner point set U is determined as follows: U={n q |Γ(n q )≥x} Where χ is the corner point measure decision value, χ∈[0.03,0.1].

2. The noise robust corner detection method based on multi-scale local similarity according to claim 1, characterized in that: In the step (1) of extracting the noise bit plane and the residual image, the threshold τ is set to 2.

3. The noise robust corner detection method based on multi-scale local similarity according to claim 1, characterized in that: In the step (2) of constructing the left and right sub-window filters, the Gaussian kernel scale factor σ is taken as 3, the anisotropy factor ρ is taken as 6, and the total number of discrete rotation direction samples M is taken as 32.

4. The noise robust corner detection method based on multi-scale local similarity according to claim 1, characterized in that: In the step (4) of extracting the edge chain code set, the standard deviation α is set to 1.

5. The noise robust corner detection method based on multi-scale local similarity according to claim 1, characterized in that: In the step (5) of determining the local similarity of the response vector, the scale N of the support region is set to 3.

6. The noise robust corner detection method based on multi-scale local similarity according to claim 1, characterized in that: In the step (7) of determining the noise robust corner point, the corner point measurement decision value χ is taken as 0.055.

Images