Target detection method based on millimeter-wave terahertz Fisher vector features

Abstract

The target detection method based on the millimeter wave terahertz Fisher vector feature is to solve the problem that the millimeter wave terahertz imaging detection has a high false detection rate, the target to be detected is difficult to separate from the environment, which makes it difficult to detect the target to be detected and reduces the imaging detection capability. The linear polarization images of the millimeter wave terahertz imaging of the target to be detected at multiple angles are obtained, and the linear polarization images of the multiple angles obtained are combined to obtain a combined image; the obtained combined image is segmented to obtain a segmented image; the Fisher vector feature quantity is calculated according to the obtained segmented image to obtain the Fisher vector feature quantity; the feature image of the target to be detected is generated according to the segmented image obtained in S2 and the Fisher vector feature quantity obtained in S3, and the feature image is threshold segmented to obtain the detection image of the target to be detected. It belongs to the field of electronic information and remote sensing detection technology.

Description

Technical Field

[0001] The invention relates to a target detection method, belonging to the technical field of electronic information and remote sensing detection. Background Art

[0002] Passive millimeter wave terahertz imaging system realizes remote sensing and detection of the observation scene by receiving the millimeter wave terahertz radiation spontaneously or reflected by the material. It has the advantages of working all day, quasi-all-weather, good concealment and no radiation hazard. Therefore, it has been applied to the fields of atmospheric remote sensing, ocean monitoring, soil and vegetation remote sensing, human security inspection, key site monitoring, military detection, etc.

[0003] In passive millimeter wave terahertz imaging applications, the grayscale difference of a single polarization image is usually used to detect and identify targets. Targets of different types and roughness usually have different physical properties. However, in single polarization imaging, the same brightness temperature grayscale may exist, making it difficult to distinguish different targets. Polarization imaging methods can produce multiple polarization images, thereby obtaining more information about the observed scene. Some studies have analyzed the brightness temperature grayscale of different targets from multiple dimensions through multi-polarization imaging, thereby improving detection and identification capabilities. However, existing millimeter wave terahertz imaging detection technology usually directly performs single or multiple threshold segmentation on the brightness temperature grayscale image, which has the problem of high false detection rate, especially the false detection of many scattered pixels, which seriously reduces the imaging detection capability; in addition, the existing technology does not use multi-polarization complementary information for comprehensive detection. The contrast of some objects is very low under single polarization, making it difficult to separate the target to be detected from the environment, resulting in difficulty in detecting the target to be detected. Therefore, how to reduce the false detection of scattered pixels during threshold segmentation and reasonably integrate polarization information is a technical problem that needs to be solved urgently. Summary of the invention

[0004] In order to solve the problems that the false detection rate of millimeter-wave terahertz imaging detection is high, the target to be detected is difficult to separate from the environment, resulting in difficulty in detecting the target to be detected and reduced imaging detection capability, the present invention further proposes a target detection method based on millimeter-wave terahertz Fisher vector features.

[0005] The technical solution adopted by the present invention is:

[0006] It includes the following steps:

[0007] S1, obtaining linear polarization images of the target to be measured at multiple angles of millimeter wave terahertz imaging, and combining the linear polarization images obtained at multiple angles to obtain a combined image;

[0008] S2, segmenting the obtained combined image to obtain a segmented image;

[0009] S3, calculating the Fisher vector feature quantity according to the obtained segmented image to obtain the Fisher vector feature quantity;

[0010] S4, generating a feature image of the target to be measured according to the segmented image obtained in S2 and the Fisher vector feature quantity obtained in S3, performing threshold segmentation on the feature image, and obtaining a detection image of the target to be measured.

[0011] Preferably, in S1, linear polarization images of millimeter-wave terahertz imaging of the target to be measured at multiple angles are obtained, and the linear polarization images obtained at multiple angles are combined to obtain a combined image. The specific process is:

[0012] S11, obtaining linear polarization images of the target at four angles of 0°, +45°, +90° and +135° of millimeter wave terahertz imaging;

[0013] S12, obtaining a corresponding two-dimensional matrix according to each acquired linear polarization image, thereby obtaining four two-dimensional matrices;

[0014] S13, adding the four obtained two-dimensional matrices and taking the average value to obtain a mean image.

[0015] Preferably, the combined image is segmented in S2 to obtain a segmented image, and the specific process is as follows:

[0016] S21, performing median filtering on the mean image obtained in S13;

[0017] S22, using a bicubic interpolation method to change the length and width of the mean image after median filtering to G times the original length and width, G>1, to obtain a mean image after the bicubic interpolation method;

[0018] S23, segmenting the mean image after the bicubic interpolation method to obtain a segmented mean image.

[0019] Preferably, in S23, the mean image after the bicubic interpolation method is segmented to obtain the segmented mean image, and the specific process is:

[0020] S231, setting a double threshold to perform rough segmentation on the mean image after the bicubic interpolation method, eliminating the area that is not the target pixel to be measured, and obtaining the mean image after rough segmentation;

[0021] S232, performing super-pixel segmentation on the roughly segmented mean image to obtain a super-pixel image.

[0022] Preferably, in S3, the Fisher vector feature quantity is calculated according to the obtained segmented image to obtain the Fisher vector feature quantity, and the specific process is:

[0023] S31, inputting the mean image after the bicubic interpolation method obtained in S22 into the mixed Gaussian model for clustering, and obtaining the mixed Gaussian model after clustering:

[0024]

[0025] Among them, f Λ (x) represents the probability density function of the Gaussian mixture model, and x represents the pixel value of the mean image after the bicubic interpolation method;

[0026] Q represents the number of Gaussian components in the mixed Gaussian model, q = 1, 2, ..., Q;

[0027] ω q represents the mixing coefficient,

[0028] represents the qth Gaussian component in the mixed Gaussian model,

[0029] σ q represents the standard deviation of the qth Gaussian component;

[0030] μ q represents the mean of the qth Gaussian component;

[0031] exp[·] represents the e-exponential operator;

[0032] S32, calculating the Fisher vector according to the clustered mixed Gaussian model obtained in S31 and the super-pixel image obtained in S232 to obtain the Fisher vector;

[0033] S33. Calculate the Fisher vector feature quantity according to the obtained Fisher vector to obtain the Fisher vector feature quantity.

[0034] Preferably, in S32, the Fisher vector is calculated according to the clustered mixed Gaussian model obtained in S31 and the superpixel image obtained in S232 to obtain the Fisher vector, and the specific process is as follows:

[0035] S321, Fisher vector:

[0036] α l =[α l,ω1 ,α l,ω2 ,...,α l,ωQ ,α l,μ1 ,α l,μ2 ,...,α l,μQ ,α l,σ1 ,α l,σ2 ,...α l,σQ ] T(2)

[0037] Among them, α l represents the Fisher vector of the lth superpixel, l = 1, 2, ..., L;

[0038] α l,ωQ represents the mixing coefficient ω q The corresponding Fisher vector, q = 1, 2, ..., Q;

[0039] α l,μQ represents the mean value of Gaussian component μ q The corresponding Fisher vector, q = 1, 2, ..., Q;

[0040] α l,σQ represents the standard deviation of the Gaussian component σ q The corresponding Fisher vector, q = 1, 2, ..., Q;

[0041] S322, using α l =sign(α l )|α l | 1 / 2 The obtained Fisher vector is updated to obtain an updated Fisher vector. sign(*) represents a sign function. The sign(*) operator can set positive numbers to 1 and negative numbers to -1.

[0042] Preferably, in S33, the Fisher vector feature quantity is calculated according to the obtained Fisher vector to obtain the Fisher vector feature quantity, and the specific process is:

[0043] The Fisher vector feature is calculated based on the updated Fisher vector obtained in S322:

[0044]

[0045] Among them, FV l Represents the Fisher vector feature of the lth superpixel;

[0046] mean{} represents the mean;

[0047] The Fisher vector α represents the lth superpixel l and the Fisher vector α of the adjacent superpixel l* The Euclidean distance between

[0048] N(l) represents the set of superpixels in the neighboring area of ​​the lth superpixel.

[0049] Preferably, in S4, a feature image of the target to be measured is generated according to the segmented image obtained in S2 and the Fisher vector feature quantity obtained in S3, and a threshold segmentation is performed on the feature image to obtain a detection image of the target to be measured. The specific process is as follows:

[0050] S41, generating a feature image of the target to be measured according to the superpixel image obtained in S232 and the Fisher vector feature quantity obtained in S33;

[0051] S42, performing threshold segmentation on the feature image of the target to be detected to obtain a detection image of the target to be detected.

[0052] Preferably, the step S41 generates a feature image of the target to be measured according to the superpixel image obtained in S232 and the Fisher vector feature quantity obtained in S33, and the specific process is as follows:

[0053] The Fisher vector feature quantity obtained in S33 is superimposed on the pixel value of each pixel in the super-pixel image obtained in S232 to generate a feature image of the target to be measured.

[0054] Preferably, in S42, threshold segmentation is performed on the feature image of the target to be detected to obtain a detection image of the target to be detected. The specific process is as follows:

[0055] The Otsu method is used to perform threshold segmentation on the feature image of the target to be measured, and the detection image of the target to be measured is obtained.

[0056] Beneficial effects:

[0057] The present invention first obtains linear polarization images of four angles (0°, +45°, +90° and +135°) of millimeter-wave terahertz imaging of a target to be measured in a millimeter-wave terahertz imaging detection image, combines and processes the linear polarization images of the four angles to obtain a processed image, then performs two threshold segmentations on the processed image in sequence, finally, calculates Fisher vector feature quantities by using a clustered mixed Gaussian model and the image after the second threshold segmentation, superimposes the calculated Fisher vector feature quantities with the pixel value of each pixel in the image after the second threshold segmentation, generates a feature image of the target to be measured in the millimeter-wave terahertz imaging detection image, and performs a third threshold segmentation on the feature image of the target to be measured by using the Otsu method, thereby generating a detection image of the target to be measured in the millimeter-wave terahertz imaging detection image. The present invention combines linear polarization images of the target to be measured at different angles and then performs three threshold segmentation on the combined processed images, so that the feature difference between the target to be measured and the environment in the millimeter wave terahertz imaging detection image is more obvious, that is, the features and contours of the target to be measured are clearer, thereby separating the target to be measured from the environment and improving the detection effect of the target to be measured. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is the mean image of the present invention;

[0059] Figure 2 is a coarse segmentation image of the present invention;

[0060] Figure 3 is the characteristic image of the present invention;

[0061] Figure 4 is a detection image of the target to be detected of the present invention; DETAILED DESCRIPTION

[0062] Specific implementation method 1: Combination Figure 1-Figure 4 The present embodiment is described. The present embodiment describes a target detection method based on millimeter wave terahertz Fisher vector features, which includes the following steps:

[0063] S1. Obtain linear polarization images of the target at multiple angles of millimeter wave terahertz imaging, and combine the linear polarization images at multiple angles to obtain a combined image. The specific process is as follows:

[0064] S11. Obtain linear polarization images at four angles of 0°, +45°, +90° and +135° of the millimeter-wave terahertz imaging of the target to be detected in the millimeter-wave terahertz imaging detection image. The linear polarization images at these four angles can express all the physical information of the linear polarization, thereby being used for target detection; millimeter-wave terahertz imaging can be obtained using a millimeter-wave terahertz imaging device.

[0065] S12. Obtain a corresponding two-dimensional matrix according to each acquired linear polarization image, and obtain two-dimensional matrices corresponding to four linear polarization images, wherein each pixel value in each linear polarization image represents a numerical value in the two-dimensional matrix corresponding to each linear polarization image, that is, the size of the numerical value represents the pixel value of the pixel in the linear polarization image.

[0066] S13. Adding the four obtained two-dimensional matrices and taking the average thereof can obtain an average image of four linear polarization images.

[0067] The linear polarization images acquired at multiple angles may be arbitrarily mathematically combined, including many combination schemes. The above-mentioned addition and averaging of the linear polarization images at multiple angles is only one embodiment of the present invention.

[0068] S2. Segment the combined image to obtain a segmented image. The specific process is as follows:

[0069] S21, performing median filtering on the mean image obtained in S13, wherein the purpose of median filtering is to reduce the noise level of the combined image.

[0070] S22. Using a bicubic interpolation method, the length and width of the mean image after median filtering are changed to G times of the original, where G>1, to obtain a mean image after the bicubic interpolation method. In the present invention, G=4.

[0071] S23, segmenting the mean image after the bicubic interpolation method to obtain a segmented mean image, the specific process is:

[0072] S231, setting a double threshold to perform coarse segmentation on the mean image after the bicubic interpolation method. Coarse segmentation means a preliminary simple and rough segmentation. As the first segmentation of the mean image, the purpose is to preliminarily eliminate the area that is obviously not the target pixel to be tested, in preparation for the subsequent superpixel segmentation and the final accurate fine segmentation detection. Obtain the mean image after coarse segmentation;

[0073] S232, super-pixel segmentation is performed on the mean image obtained after rough segmentation. The mean image after rough segmentation is segmented again by super-pixel segmentation, with the purpose of aggregating local areas with adjacent positions and similar features such as brightness and texture, reducing the number of pixels without destroying the image boundary information, and reducing the processing complexity. As the second segmentation of the mean image, a super-pixel image is obtained.

[0074] S3. Calculate the Fisher vector feature quantity according to the obtained segmented image to obtain the Fisher vector feature quantity. The specific process is as follows:

[0075] S31, inputting the mean image after the bicubic interpolation method obtained in S22 into the mixed Gaussian model for clustering, and obtaining the mixed Gaussian model after clustering, that is:

[0076]

[0077] Among them, f Λ (x) represents the probability density function of the Gaussian mixture model, and x represents the pixel value of the mean image after the bicubic interpolation method;

[0078] Q represents the number of Gaussian components in the mixed Gaussian model, q = 1, 2, ..., Q;

[0079] ω q Represents the mixing coefficient, that is, the weight, and satisfies:

[0080] represents the qth Gaussian component in the mixed Gaussian model,

[0081] σ q represents the standard deviation of the qth Gaussian component;

[0082] μq represents the mean of the qth Gaussian component;

[0083] exp[·] represents the e-exponential operator.

[0084] The Gaussian mixture model (GMM) is composed of a plurality of single Gaussian model fittings. The Gaussian mixture model is used to cluster the mean image obtained after the bicubic interpolation method, so as to obtain the clustered Gaussian mixture model.

[0085] S32, calculating the Fisher vector according to the clustered mixed Gaussian model obtained in S31 and the super-pixel image obtained in S232, to obtain the Fisher vector, the specific process is:

[0086] S321, Fisher vector:

[0087] α l =[α l,ω1 ,α l,ω2 ,...,α l,ωQ ,α l,μ1 ,α l,μ2 ,...,α l,μQ ,α l,σ1 ,α l,σ2 ,...α l,σQ ] T (5)

[0088] Among them, α l represents the Fisher vector of the lth superpixel, l = 1, 2, ..., L;

[0089] α l,ωQ represents the mixing coefficient ω q The corresponding Fisher vector, q = 1, 2, ..., Q;

[0090] α l,μQ represents the mean value of Gaussian component μ q The corresponding Fisher vector, q = 1, 2, ..., Q;

[0091] α l,σQ represents the standard deviation of the Gaussian component σ q The corresponding Fisher vector, q = 1, 2, ..., Q;

[0092] S322, using formula α l =sign(α l )|α l | 1 / 2The obtained Fisher vector is updated to obtain an updated Fisher vector. The purpose of the update is to prevent the Fisher vector from being sparse. Here, sign(*) represents a sign function. The sign(*) operator can set positive numbers to 1 and negative numbers to -1.

[0093] S33, calculating the Fisher vector feature quantity according to the obtained Fisher vector to obtain the Fisher vector feature quantity, the specific process is:

[0094] The Fisher vector feature is calculated based on the updated Fisher vector obtained in S322:

[0095]

[0096] Among them, FV l The Fisher vector feature of the lth superpixel is represented, and the mean of the Euclidean distance is selected as the Fisher vector feature of the current lth superpixel;

[0097] mean{} represents the mean;

[0098] The Fisher vector α represents the lth superpixel l and the Fisher vectors of adjacent superpixels The Euclidean distance between

[0099] N(l) represents the set of superpixels in the neighboring area of ​​the lth superpixel.

[0100] S4, generating a feature image of the target to be measured according to the segmented image obtained in S2 and the Fisher vector feature quantity obtained in S3, performing threshold segmentation on the feature image, and obtaining a detection image of the target to be measured. The specific process is as follows:

[0101] S41. Generate a feature image of the target to be measured in the millimeter-wave terahertz imaging detection image according to the super-pixel image obtained in S232 and the Fisher vector feature quantity obtained in S33, that is, superimpose a certain Fisher vector feature quantity obtained in S33 with the pixel value of each pixel in the super-pixel image obtained in S232 to generate a feature image of the target to be measured.

[0102] S42, performing threshold segmentation on the feature image of the target to be detected by using the Otsu method to obtain a detection image of the target to be detected.

Claims

1. A target detection method based on millimeter-wave terahertz Fisher vector features, characterized in that: It includes the following steps: S1. Obtain linear polarization images of the target at multiple angles of millimeter wave terahertz imaging, and combine the linear polarization images at multiple angles to obtain a combined image. The specific process is as follows: S11, obtaining linear polarization images of the target at four angles of 0°, +45°, +90° and +135° of millimeter wave terahertz imaging; S12, obtaining a corresponding two-dimensional matrix according to each acquired linear polarization image, thereby obtaining four two-dimensional matrices; S13, adding the obtained four two-dimensional matrices and taking the average value to obtain a mean image; S2. Segment the combined image to obtain a segmented image. The specific process is as follows: S21, performing median filtering on the mean image obtained in S13; S22, using a bicubic interpolation method to change the length and width of the mean image after median filtering to G times the original length and width, G>1, to obtain a mean image after the bicubic interpolation method; S23, segmenting the mean image after the bicubic interpolation method to obtain a segmented mean image, the specific process is: S231, setting a double threshold to perform coarse segmentation on the mean image after the bicubic interpolation method, eliminating the area that is not the target pixel to be measured, and obtaining the mean image after coarse segmentation; S232, performing super-pixel segmentation on the mean image obtained after the rough segmentation to obtain a super-pixel image; S3, calculating the Fisher vector feature quantity according to the obtained segmented image to obtain the Fisher vector feature quantity; S4, generating a feature image of the target to be measured according to the segmented image obtained in S2 and the Fisher vector feature quantity obtained in S3, performing threshold segmentation on the feature image, and obtaining a detection image of the target to be measured.

2. The target detection method based on millimeter wave terahertz Fisher vector features according to claim 1, characterized in that: In S3, the Fisher vector feature quantity is calculated according to the obtained segmented image to obtain the Fisher vector feature quantity. The specific process is as follows: S31, inputting the mean image after the bicubic interpolation method obtained in S22 into the mixed Gaussian model for clustering, and obtaining the mixed Gaussian model after clustering: Among them, f Λ (x) represents the probability density function of the Gaussian mixture model, and x represents the pixel value of the mean image after the bicubic interpolation method; Q represents the number of Gaussian components in the mixed Gaussian model, q = 1, 2, ..., Q; ω q represents the mixing coefficient, represents the qth Gaussian component in the mixed Gaussian model, σ q represents the standard deviation of the qth Gaussian component; μ q represents the mean of the qth Gaussian component; exp[·] represents the e-exponential operator; S32, calculating the Fisher vector according to the clustered mixed Gaussian model obtained in S31 and the superpixel image obtained in S232 to obtain the Fisher vector; S33. Calculate the Fisher vector feature quantity according to the obtained Fisher vector to obtain the Fisher vector feature quantity.

3. The target detection method based on millimeter wave terahertz Fisher vector features according to claim 2, characterized in that: In S32, the Fisher vector is calculated according to the clustered mixed Gaussian model obtained in S31 and the superpixel image obtained in S232 to obtain the Fisher vector. The specific process is as follows: S321, Fisher vector: α l ＝[α l,ω1 ,α l,ω2 ,...,α l,ωQ ,α l,μ1 ,α l,μ2 ,...,α l,μQ ,α l,σ1 ,α l,σ2 ,...α l,σQ ] T (2) Among them, α l represents the Fisher vector of the lth superpixel, l = 1, 2, ..., L; α l,ωQ represents the mixing coefficient ω q The corresponding Fisher vector, q = 1, 2, ..., Q; α l,μQ represents the mean value of Gaussian component μ q The corresponding Fisher vector, q = 1, 2, ..., Q; α l,σQ represents the standard deviation of the Gaussian component σ q The corresponding Fisher vector, q = 1, 2, ..., Q; S322, using α l =sign(α l )|α l | 1 / 2 The obtained Fisher vector is updated to obtain an updated Fisher vector. sign(*) represents a sign function. The sign(*) operator can set positive numbers to 1 and negative numbers to -1.

4. The target detection method based on millimeter wave terahertz Fisher vector features according to claim 3, characterized in that: In S33, the Fisher vector feature quantity is calculated according to the obtained Fisher vector to obtain the Fisher vector feature quantity. The specific process is as follows: The Fisher vector feature is calculated based on the updated Fisher vector obtained in S322: Among them, FV l Represents the Fisher vector feature of the lth superpixel; mean{} represents the mean; The Fisher vector α represents the lth superpixel l and the Fisher vectors of adjacent superpixels The Euclidean distance between N(l) represents the set of superpixels in the neighboring area of ​​the lth superpixel.

5. The target detection method based on millimeter wave terahertz Fisher vector features according to claim 4, characterized in that: In S4, a feature image of the target to be detected is generated according to the segmented image obtained in S2 and the Fisher vector feature quantity obtained in S3, and a threshold segmentation is performed on the feature image to obtain a detection image of the target to be detected. The specific process is as follows: S41, generating a feature image of the target to be measured according to the superpixel image obtained in S232 and the Fisher vector feature quantity obtained in S33; S42, performing threshold segmentation on the feature image of the target to be detected to obtain a detection image of the target to be detected.

6. The target detection method based on millimeter wave terahertz Fisher vector features according to claim 5, characterized in that: The S41 generates a feature image of the target to be measured according to the superpixel image obtained in S232 and the Fisher vector feature quantity obtained in S33. The specific process is as follows: The Fisher vector feature quantity obtained in S33 is superimposed on the pixel value of each pixel in the super-pixel image obtained in S232 to generate a feature image of the target to be measured.

7. The target detection method based on millimeter wave terahertz Fisher vector features according to claim 6, characterized in that: In S42, the characteristic image of the target to be detected is segmented by threshold value to obtain the detection image of the target to be detected. The specific process is as follows: The Otsu method is used to perform threshold segmentation on the feature image of the target to be measured, and the detection image of the target to be measured is obtained.

Images