Target detection method and device based on local region feature difference, equipment and storage medium

Abstract

The invention discloses a target detection method and device based on local region feature difference, equipment and a storage medium, and belongs to the technical field of target detection. The target detection method comprises the following steps: reconstructing an image based on background cancellation of a sensor; fitting brightness temperature distribution of the reconstructed image by adopting a Gaussian mixture model, and calculating to obtain a multi-order statistical characteristic parameter of each Gaussian component; calculating a Fisher vector of the superpixel by using the multi-order statistical characteristic parameters, and constructing a Fisher vector contrast; constructing a reference superpixel set of a local area based on the slow variation of background superpixels; on the basis of brightness temperature characteristics, extreme value suppression is utilized to construct suppression characteristics of background superpixels; in the local region, multiplying the Fisher vector contrast by the suppression feature of the background superpixel to obtain the heterosexual feature of the local region difference; and carrying out binarization processing on the difference characteristics by adopting a two-parameter method, obtaining target area pixels, and obtaining a target detection result. The target detection intersection-to-parallel ratio and quality factor performance can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of target detection, and more specifically, relates to a target detection method, device, equipment and storage medium based on local area feature differences. Background Art

[0002] The synthetic aperture radiometer imaging system has the advantages of sensitivity to metal and heat source targets, good concealment, natural forward observation capability and instantaneous imaging, which makes it widely used in target detection fields such as earth remote sensing, radio astronomy observation, military reconnaissance and human security inspection. Under the observation conditions of airborne platforms, the observed target has a certain texture structure, its brightness temperature distribution is non-uniform, the background area distribution is complex, and the reconstructed image is affected by noise, making it difficult to detect the complete target area. Therefore, many studies are devoted to solving the problem of target detection based on synthetic aperture radiometer imaging.

[0003] The existing target detection technology based on synthetic aperture radiometer imaging mainly relies on differential features and can be divided into two categories: one is based on the trajectory features of continuous frame images, detecting targets by extracting continuous target positioning points; the other is based on the features of single frame images, using the pixel brightness temperature difference or frequency characteristics under noise suppression for target detection. However, these two methods face the problem of false detection and missed detection of target pixels because they only use a single feature. Summary of the invention

[0004] In view of the defects of the related art, the purpose of the present invention is to provide a target detection method, device, equipment and storage medium based on local area feature differences, aiming to solve the problem of missed detection and false detection of target pixels caused by only using a single feature in the existing methods.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a target detection method based on local area feature differences, comprising:

[0006] S1. Perform background cancellation processing on the image acquired by the sensor to obtain a reconstructed image; perform superpixel segmentation on the reconstructed image, use a Gaussian mixture model to fit the brightness temperature distribution of the reconstructed image, and calculate the multi-order statistical characteristic parameters of each Gaussian component;

[0007] S2, calculating the Fisher vector of the superpixel according to the multi-order statistical characteristic parameters; constructing a local area of ​​a preset size and a reference superpixel set based on the slow variation of the background superpixels near the target area in the reconstructed image; calculating the feature difference between each superpixel point to be tested in the target area and each reference superpixel point in the local area according to the Fisher vector of each superpixel, and constructing a Fisher vector contrast;

[0008] S3, based on the brightness temperature characteristics of the target, using extreme value suppression to construct the suppression feature of the background superpixel; in the local area, multiplying the Fisher vector contrast and the suppression feature of the background superpixel to obtain the difference feature of the local area;

[0009] S4. Use a dual parameter method to perform binarization processing on the difference features of the local area, obtain pixels of the target area, and obtain a target detection result.

[0010] Optionally, S1 specifically includes:

[0011] S11, respectively acquiring a detection scene brightness temperature image and a background brightness temperature image based on a sensor, subtracting the background brightness temperature image from the detection scene brightness temperature image to perform background cancellation, and obtaining a reconstructed image;

[0012] S12, performing superpixel segmentation on the reconstructed image, and fitting the brightness temperature distribution of the reconstructed image using a Gaussian mixture model to obtain multi-order statistical characteristic parameters of each Gaussian component:

[0013]

[0014] Among them, x represents the brightness temperature of the pixel corresponding to the background cancellation reconstructed image, θ Q =[ω1,ω2,…,ω Q ,μ1,μ2,…,μ Q ,σ1,σ2,…,σ Q ] T represents the multi-order statistical characteristic parameters corresponding to the Gaussian mixture model, Q represents the number of components, ω q is the clustering weight of the component, μ q is the cluster center, σ q is the cluster width, is a Gaussian density function.

[0015] Optionally, S2 specifically includes:

[0016] S21. Calculate the Fisher vector of the superpixel according to the multi-order statistical characteristic parameters:

[0017]

[0018] Among them, P l represents the number of pixels contained in the lth superpixel, is the Cholesky decomposition of the inverse matrix of Fisher information, l∈{1,2,…,L} represents the superpixel index, L represents the number of superpixels divided, Represents about θ Q The gradient operator, x l,p Indicates the brightness temperature corresponding to the pth pixel in the lth superpixel;

[0019] S22, constructing a local area of ​​a preset size and a reference superpixel set C based on the slow variation of the background superpixels near the target area in the reconstructed image l ;

[0020] Among them, C l It means that the geometric center of the first superpixel is centered on d w The set of superpixels covered by a rectangular box with a side length of d w =ρ·S represents the side length of the rectangular frame, ρ represents the magnification of the rectangular frame, represents the average superpixel size;

[0021] S23, calculating the feature difference between each superpixel point to be tested in the target area and each reference superpixel point in the local area according to the Fisher vector of each superpixel, and constructing the Fisher vector contrast using the power normalized Fisher vector and the Euclidean distance:

[0022]

[0023] in, represents the power normalized Fisher vector, ||·||2 represents the 2-norm of the vector, sgn(·) represents the element-wise sign function, represents the element-wise arithmetic square root, and l and l′ represent different superpixel indices.

[0024] Optionally, S3 specifically includes:

[0025] S31. For low brightness temperature target observation scenes, the suppression feature of background superpixels is constructed as follows:

[0026]

[0027] Among them, min{·,0} means taking the minimum value compared with 0. It indicates the median brightness temperature of pixels in the superpixel, l and l′ indicate different superpixel indexes;

[0028] For the high-brightness temperature target observation scene, the suppression feature of the background superpixel is constructed as follows:

[0029]

[0030] Among them, max{·,0} means to calculate the maximum value compared with 0;

[0031] S32, in the local area, combining the Fisher vector contrast by multiplication and the suppression feature δ(x l ,x l′ ), construct the difference features of the local area:

[0032]

[0033] Optionally, S4 specifically includes:

[0034] S41, using the dual parameter method to set the detection threshold as: in, Represents the mean of the local regional difference characteristics, represents the standard deviation of local regional difference characteristics, and α represents the threshold adjustment factor;

[0035] S42, binarizing the difference features of the local areas of all superpixels, obtaining pixels in the target area, and obtaining target detection results.

[0036] In a second aspect, the present invention further provides a target detection device based on local area feature differences, comprising:

[0037] A segmentation module is used to perform background cancellation processing based on the image acquired by the sensor to obtain a reconstructed image; perform superpixel segmentation on the reconstructed image, use a Gaussian mixture model to fit the brightness temperature distribution of the reconstructed image, and calculate the multi-order statistical characteristic parameters of each Gaussian component;

[0038] A calculation module, used to calculate the Fisher vector of the superpixel according to the multi-order statistical characteristic parameters; construct a local area of ​​a preset size and a reference superpixel set based on the slow variation of the background superpixels near the target area in the reconstructed image; calculate the feature difference between each superpixel point to be tested in the target area and each reference superpixel point in the local area according to the Fisher vector of each superpixel, and construct a Fisher vector contrast;

[0039] A construction module is used to construct a suppression feature of a background superpixel based on the brightness temperature characteristics of the target by using an extreme value suppression method; in the local area, a Fisher vector contrast is multiplied by the suppression feature of the background superpixel to obtain a difference feature of the local area;

[0040] The detection module is used to perform binarization processing on the difference features of the local area using a dual parameter method, obtain pixels of the target area, and obtain target detection results.

[0041] In a third aspect, the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method described in any one of the first aspects when executing the computer program.

[0042] In a fourth aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method as described in any one of the first aspects.

[0043] Compared with the prior art, the above technical solution conceived by the present invention can achieve the following beneficial effects:

[0044] The present invention provides a target detection method based on local area feature difference. First, an image is reconstructed based on background cancellation of a sensor, multi-order statistical characteristic parameters of components are obtained by using a Gaussian mixture model, Fisher vectors are calculated based on the description of superpixels by using multi-order statistical characteristics, Fisher vector contrast is constructed, and false alarm pixels are avoided and a signal-to-noise ratio is improved by a superpixel level processing method; local brightness temperature characteristics of the target and the slow variation of the background near the target are considered, background superpixels are suppressed by using an extreme value suppression method to reduce the background superpixel intensity, local area difference features are constructed in combination with Fisher vector contrast to improve the superpixel intensity of the target area; based on the local area difference features, the features are binarized by using a dual parameter method to achieve target detection, and pixels in the target area are obtained. In this scheme, multi-order statistical characteristics are used to describe superpixels, and the differences in feature vectors (Fisher vectors) composed of zero-order, first-order and second-order statistical information are considered for target detection. Compared with the traditional method that only relies on brightness temperature differences for target detection, multi-order statistical information is used to improve the difference compared to a single feature / single statistical information; the problems of high false detection rate and target pixel omission based on single feature target detection are solved, and the target detection intersection-over-union ratio and quality factor performance can be improved, and effective detection of extended source targets under the influence of noise can be performed. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of a target detection method based on local area feature differences in an embodiment of the present invention;

[0046] Figure 2 is an array configuration diagram of a data acquisition system used in an embodiment of the present invention;

[0047] Figure 3a is an optical image of a selected target in an embodiment of the present invention;

[0048] Figure 3b is a brightness temperature image reconstructed by background cancellation of the 15th frame of scene data in an embodiment of the present invention;

[0049] Figure 4 15 is a density function and component function diagram of the Gaussian mixture model fitting of the scene data of the 15th frame in an embodiment of the present invention;

[0050] Figure 5 is a local area difference feature map of the 15th frame of the scene data in an embodiment of the present invention;

[0051] Figure 6It is the local area difference feature of the dual parameter method of the 15th frame of the scene data in the embodiment of the present invention Figure 2 The valued test result graph;

[0052] Figure 7a is a detection result diagram of the 15th frame of scene data based on the windowed background cancellation binarization method in an embodiment of the present invention;

[0053] Figure 7b is a curve diagram showing the change of the overall I / O ratio of all 29 frames of scene data with the normalized threshold adjustment factor in an embodiment of the present invention;

[0054] Figure 7c It is a curve diagram showing the change of the overall quality factor of all 29 frames of scene data with the normalized threshold adjustment factor in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0056] The contents involved in the above embodiment are described below in conjunction with a preferred embodiment.

[0057] Embodiment 1

[0058] The present invention provides a target detection method based on local area feature differences, comprising:

[0059] S1. Perform background cancellation processing on the image acquired by the sensor to obtain a reconstructed image; perform superpixel segmentation on the reconstructed image, use a Gaussian mixture model to fit the brightness temperature distribution of the reconstructed image, and calculate the multi-order statistical characteristic parameters of each Gaussian component;

[0060] S2, calculating the Fisher vector of the superpixel according to the multi-order statistical characteristic parameters; constructing a local area of ​​a preset size and a reference superpixel set based on the slow variation of the background superpixels near the target area in the reconstructed image; calculating the feature difference between each superpixel point to be tested in the target area and each reference superpixel point in the local area according to the Fisher vector of each superpixel, and constructing a Fisher vector contrast;

[0061] S3, based on the brightness temperature characteristics of the target, using extreme value suppression to construct the suppression feature of the background superpixel; in the local area, multiplying the Fisher vector contrast and the suppression feature of the background superpixel to obtain the difference feature of the local area;

[0062] S4. Use a dual parameter method to perform binarization processing on the difference features of the local area, obtain pixels of the target area, and obtain a target detection result.

[0063] Based on the image acquired by the sensor, background cancellation processing is performed to obtain a reconstructed image; the brightness temperature distribution of the reconstructed image with background cancellation is fitted according to the Gaussian mixture model to obtain the multi-order statistical characteristic parameters of each Gaussian component, and the Fisher vector of the superpixel is calculated based on these parameters, and the Fisher vector contrast is constructed using the power normalized Fisher vector and the Euclidean distance. The superpixel-level processing method is used to avoid false alarm pixels and improve the signal-to-noise ratio. Considering the difference between the local differences of different background superpixels and the local differences of the target superpixels, the extreme value suppression method is used to suppress the background superpixels with local differences in brightness temperature distribution, reduce the background superpixel intensity, and use the superpixels covered by the local rectangular frame as the reference unit to suppress the background superpixels near the target area, improve the target superpixel intensity, construct the difference characteristics of the local area, and improve the signal-to-noise ratio. The detection threshold of the local area difference feature is obtained by the dual parameter method, and the target area pixels are detected to achieve target detection.

[0064] Optionally, S1 specifically includes:

[0065] S11, respectively acquiring a detection scene brightness temperature image and a background brightness temperature image based on a sensor, subtracting the background brightness temperature image from the detection scene brightness temperature image to perform background cancellation, and obtaining a reconstructed image;

[0066] S12, performing superpixel segmentation on the reconstructed image, and fitting the brightness temperature distribution of the reconstructed image using a Gaussian mixture model to obtain multi-order statistical characteristic parameters of each Gaussian component:

[0067]

[0068] Among them, x represents the brightness temperature of the pixel corresponding to the background cancellation reconstructed image, θ Q =[ω1,ω2,…,ω Q ,μ1,μ2,…,μ Q ,σ1,σ2,…,σ Q ] T represents the multi-order statistical characteristic parameters corresponding to the Gaussian mixture model, Q represents the number of components, ω q is the clustering weight of the component, μ q is the cluster center, σ q is the cluster width, is a Gaussian density function.

[0069] The specific steps are as follows:

[0070] (1) The expectation maximization algorithm is used to estimate the multi-order statistical characteristic parameters corresponding to the Gaussian mixture model of the background cancellation image to obtain the density function of the fitting brightness temperature distribution;

[0071] (2) Establish the relationship between the density function of the brightness temperature distribution of the background cancellation reconstructed image after fitting and each Gaussian component. The expression is as follows:

[0072]

[0073] In the formula, x represents the brightness temperature of the pixel in the background cancellation reconstructed image, and the value of this variable is the brightness temperature value in the image. Q It represents the multi-order statistical characteristic parameters corresponding to the Gaussian mixture model. The parameter vector is a column vector composed of the cluster weights, cluster centers and cluster widths of each component in sequence. Q is equal to the number of components.

[0074] For the multi-order statistical characteristic parameters corresponding to the Gaussian mixture model:

[0075] θ Q =[ω1,ω2,…,ω Q ,μ1,μ2,…,μ Q ,σ1,σ2,…,σ Q ] T , where ω q , μ q and σ q are the clustering weight, cluster center and cluster width of the components respectively.

[0076] p(x|μ q ,σ q ) represents the Gaussian density function corresponding to each component, expressed as:

[0077]

[0078] Optionally, S2 specifically includes:

[0079] S21. Calculate the Fisher vector of the superpixel according to the multi-order statistical characteristic parameters:

[0080]

[0081] Among them, P l represents the number of pixels contained in the lth superpixel, is the Cholesky decomposition of the inverse matrix of Fisher information, l∈{1,2,…,L} represents the superpixel index, L represents the number of superpixels divided, Represents about θ Q The gradient operator, x l,pIndicates the brightness temperature corresponding to the pth pixel in the lth superpixel;

[0082] S22, constructing a local area of ​​a preset size and a reference superpixel set C based on the slow variation of the background superpixels near the target area in the reconstructed image l ;

[0083] Among them, C l It means that the geometric center of the first superpixel is centered on d w The set of superpixels covered by a rectangular box with a side length of d w =ρ·S represents the side length of the rectangular frame, ρ represents the magnification of the rectangular frame, represents the average superpixel size;

[0084] S23, calculating the feature difference between each superpixel point to be tested in the target area and each reference superpixel point in the local area according to the Fisher vector of each superpixel, and constructing the Fisher vector contrast using the power normalized Fisher vector and the Euclidean distance:

[0085]

[0086] in, represents the power normalized Fisher vector, ||·||2 represents the 2-norm of the vector, sgn(·) represents the element-wise sign function, represents the element-wise arithmetic square root, and l and l′ represent different superpixel indices.

[0087] Optionally, S3 specifically includes:

[0088] S31. For low brightness temperature target observation scenes, the suppression feature of background superpixels is constructed as follows:

[0089]

[0090] Among them, min{·,0} means taking the minimum value compared with 0. It indicates the median brightness temperature of pixels in the superpixel, l and l′ indicate different superpixel indexes;

[0091] For the high-brightness temperature target observation scene, the suppression feature of the background superpixel is constructed as follows:

[0092]

[0093] Among them, max{·,0} means to calculate the maximum value compared with 0;

[0094] S32, combining Fisher vector contrast by multiplication and the suppression feature δ(x l ,x l′), construct the difference features of the local area:

[0095]

[0096] Optionally, S4 specifically includes:

[0097] S41, using the dual parameter method to set the detection threshold as: in, Represents the mean of the local regional difference characteristics, represents the standard deviation of local regional difference characteristics, and α represents the threshold adjustment factor;

[0098] S42, binarizing the difference features of the local areas of all superpixels, obtaining pixels in the target area, and obtaining target detection results.

[0099] In a specific embodiment, background cancellation and reconstructed brightness temperature image data collected by a synthetic aperture radiometer system can be selected as the original data. Figure 2 Indicates the array configuration of the system used.

[0100] In this embodiment, a civilian ship in Qinhuangdao waters is selected as the detection target. The optical image corresponding to the scene target is as follows: Figure 3a As shown in the figure, the scene data consists of 29 frames of continuous frame data, among which the background cancellation and reconstruction of the brightness temperature image of the 15th frame is as follows Figure 3b shown.

[0101] (1) The brightness temperature distribution of the background cancellation reconstructed image is fitted according to the Gaussian mixture model to obtain the multi-order statistical characteristic parameters of each Gaussian component. The Fisher vector of the superpixel is calculated based on these parameters, and the Fisher vector contrast is constructed using the sparsity and Euclidean distance of the Fisher vector.

[0102] The expectation maximization algorithm is used to estimate the multi-order statistical characteristic parameters corresponding to the Gaussian mixture model, and the density function and component functions of the brightness temperature distribution of the fitted background cancellation reconstructed image are obtained. There are Q component functions in total, among which the component function of the 15th frame is as follows Figure 4 shown.

[0103] A Fisher vector is calculated to represent each superpixel, where the dimension of the Fisher vector is 3Q × 1. The sparse Fisher vector is power normalized by a signed arithmetic square root, and then the Fisher vector contrast of each pair of superpixels is calculated using the Euclidean distance.

[0104] (2) Considering the differences between the local differences of different background superpixels and the local differences of target superpixels, the superpixels covered by the local rectangular box are used as reference units. The extreme value suppression method is used to suppress the background superpixels with local differences in brightness temperature distribution, and the local area difference features are constructed.

[0105] According to the slow variation of the background superpixels near the target area and the local high brightness temperature attribute of the background superpixels to be suppressed, the reference superpixels in the local area are selected by setting a rectangular box.

[0106] The brightness temperature distribution of superpixels in the local area is determined by median operation, and the background superpixels with high brightness temperature distribution in the local area are suppressed by extreme value suppression.

[0107] The Fisher vector contrast is multiplied by the background suppression feature to complete the suppression of background superpixels. The eigenvalues ​​corresponding to the low brightness temperature superpixels in the local area are converted into high intensity values ​​by squaring, and the local area difference features are obtained according to the contrast and brightness temperature distribution differences of the selected reference superpixels.

[0108] In this embodiment, the target area is characterized by low brightness temperature in the local area. The background superpixel suppression feature mainly suppresses the background superpixels with high brightness temperature distribution in the local area. The local area difference feature map of the 15th frame is shown in FIG. Figure 5 shown.

[0109] (3) Based on the local area difference features obtained in step (2), a dual parameter method is used to obtain a detection threshold and detect pixels in the target area.

[0110] The mean and standard deviation of the local area difference features are calculated to obtain the detection threshold.

[0111] The local area difference features of all superpixels are binarized to obtain the pixels contained in the target area.

[0112] In this embodiment, the local area difference characteristics are analyzed by the dual parameter method. Figure 2 The detection result of the 15th frame obtained by valuating is as follows Figure 6 shown.

[0113] Furthermore, other methods may be used to determine the detection threshold in specific implementations.

[0114] For further explanation, a comparison of the overall intersection-over-combination ratio and the overall quality factor between the present invention and other target detection methods for synthetic aperture radiometers is set.

[0115] All 29 frames of data in this embodiment are uniformly used, and the dual parameter method is used to detect the targets in all frames. In addition to the detection method (LRC) of the present invention, the detection results based on the windowed background cancellation binarization (BWDFT) method are also considered. The detection result of the 15th frame is as follows: Figure 7a As shown. The overall intersection over union (IoU) and the overall quality factor (FoM) are given by the normalized threshold adjustment factor α0 = (α-α min ) / (α max -αmin ), where the curves of the overall I / O ratio and the overall quality factor of the BWDFT ​​method and the present method under all 29 frames of data changing with the normalized threshold adjustment factor are shown as follows: Figure 7b and Figure 7c As shown in the figure. The larger the overall IoU ratio, the more accurate the pixel-level detection results in all frame data, and the larger the overall quality factor, the fewer false detection targets and the more real target detections in all frame data. The comparison of the average IoU ratio and the average quality factor within the entire normalized threshold adjustment factor range is shown in Table 1.

[0116] Table 1

[0117] Detection Methods BWDFT LRC Average Intersection-Union Ratio 0.1691 0.5099 Average quality factor 0.5517 0.6858

[0118] In the embodiment provided by the present invention, the detection results obtained show that the above-mentioned target detection method based on local area feature differences has high intersection-over-union ratio and quality factor performance, and can detect the complete target area in continuous frames, effectively solving the problems of false detection and missed detection of target pixels.

[0119] In the embodiment of the present invention, the image is reconstructed by background cancellation of the sensor, the multi-order statistical characteristic parameters of the components are obtained by using the Gaussian mixture model, the Fisher vector is calculated based on the multi-order statistical characteristic parameters, the Fisher vector contrast is constructed, the background superpixel is suppressed by extreme value suppression, the local area difference feature is constructed in combination with the Fisher vector contrast, and finally the feature is binarized by the dual parameter method to achieve target detection. The technical problem of missed detection and false detection of target pixels caused by only using a single feature in the existing method is solved. The beneficial effect of improving the intersection-over-union ratio and quality factor performance of target detection and effectively detecting the extended source target under the influence of noise is achieved.

[0120] Embodiment 2

[0121] The present invention also provides a target detection device based on local area feature differences, comprising:

[0122] A segmentation module is used to perform background cancellation processing based on the image acquired by the sensor to obtain a reconstructed image; perform superpixel segmentation on the reconstructed image, use a Gaussian mixture model to fit the brightness temperature distribution of the reconstructed image, and calculate the multi-order statistical characteristic parameters of each Gaussian component;

[0123] A calculation module, used to calculate the Fisher vector of the superpixel according to the multi-order statistical characteristic parameters; construct a local area of ​​a preset size and a reference superpixel set based on the slow variation of the background superpixels near the target area in the reconstructed image; calculate the feature difference between each superpixel point to be tested in the target area and each reference superpixel point in the local area according to the Fisher vector of each superpixel, and construct a Fisher vector contrast;

[0124] A construction module is used to construct a suppression feature of a background superpixel based on the brightness temperature characteristics of the target by using an extreme value suppression method; in the local area, a Fisher vector contrast is multiplied by the suppression feature of the background superpixel to obtain a difference feature of the local area;

[0125] The detection module is used to perform binarization processing on the difference features of the local area using a dual parameter method, obtain pixels of the target area, and obtain target detection results.

[0126] An object detection device based on local area feature differences provided by an embodiment of the present invention is used to execute an object detection method based on local area feature differences provided by any embodiment of the present invention, and has corresponding beneficial effects.

[0127] Embodiment 3

[0128] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the method described in any one of the first embodiments are implemented.

[0129] Embodiment 4

[0130] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the method described in any one of the first embodiments are implemented.

[0131] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A target detection method based on local area feature differences, characterized in that: include: S1, performing background cancellation processing based on the image acquired by the sensor to obtain a reconstructed image; Performing superpixel segmentation on the reconstructed image, using a Gaussian mixture model to fit the brightness temperature distribution of the reconstructed image, and calculating multi-order statistical characteristic parameters of each Gaussian component; S2, calculating the Fisher vector of the superpixel according to the multi-order statistical characteristic parameters; constructing a local area of ​​a preset size and a reference superpixel set based on the slow variation of the background superpixels near the target area in the reconstructed image; calculating the feature difference between each superpixel point to be tested in the target area and each reference superpixel point in the local area according to the Fisher vector of each superpixel, and constructing a Fisher vector contrast; S3, based on the brightness temperature characteristics of the target, using extreme value suppression to construct the suppression feature of the background superpixel; in the local area, multiplying the Fisher vector contrast and the suppression feature of the background superpixel to obtain the difference feature of the local area; S4. Use a dual parameter method to perform binarization processing on the difference features of the local area, obtain pixels of the target area, and obtain a target detection result.

2. The method according to claim 1, characterized in that S1 specifically includes: S11, respectively acquiring a detection scene brightness temperature image and a background brightness temperature image based on a sensor, subtracting the background brightness temperature image from the detection scene brightness temperature image to perform background cancellation, and obtaining a reconstructed image; S12, performing superpixel segmentation on the reconstructed image, and fitting the brightness temperature distribution of the reconstructed image using a Gaussian mixture model to obtain multi-order statistical characteristic parameters of each Gaussian component: Among them, x represents the brightness temperature of the pixel corresponding to the background cancellation reconstructed image, θ Q =[ω1,ω2,…,ω Q ,μ1,μ2,…,μ Q ,σ1,σ2,…,σ Q ] T represents the multi-order statistical characteristic parameters corresponding to the Gaussian mixture model, Q represents the number of components, ω q is the clustering weight of the component, μ q is the cluster center, σ q is the cluster width, is a Gaussian density function.

3. The method according to claim 1, characterized in that S2 specifically includes: S21. Calculate the Fisher vector of the superpixel according to the multi-order statistical characteristic parameters: Among them, P l represents the number of pixels contained in the lth superpixel, is the Cholesky decomposition of the inverse matrix of Fisher information, l∈{1,2,…,L} represents the superpixel index, L represents the number of superpixels divided, Represents about θ Q The gradient operator, x l,p Indicates the brightness temperature corresponding to the pth pixel in the lth superpixel; S22, constructing a local area of ​​a preset size and a reference superpixel set C based on the slow variation of the background superpixels near the target area in the reconstructed image l ; Among them, C l It means that the geometric center of the first superpixel is centered on d w The set of superpixels covered by a rectangular box with a side length of d w =ρ·S represents the side length of the rectangular frame, ρ represents the magnification of the rectangular frame, represents the average superpixel size; S23. Calculate the feature difference between each superpixel to be tested in the target area and each reference superpixel in the local area according to the Fisher vector of each superpixel, and construct the Fisher vector contrast using the power-normalized Fisher vector and the Euclidean distance: in, represents the power normalized Fisher vector, ||·||2 represents the 2-norm of the vector, sgn(·) represents the element-wise sign function, represents the element-wise arithmetic square root, and l and l′ represent different superpixel indices.

4. The method according to claim 1, characterized in that S3 specifically includes: S31. For low brightness temperature target observation scenes, the suppression feature of background superpixels is constructed as follows: Among them, min{·,0} means taking the minimum value compared with 0. It indicates the median brightness temperature of pixels in the superpixel, l and l′ indicate different superpixel indexes; For the high-brightness temperature target observation scene, the suppression feature of the background superpixel is constructed as follows: Among them, max{·,0} means to calculate the maximum value compared with 0; S32, in the local area, combining the Fisher vector contrast by multiplication and the suppression feature δ(x l ,x l′ ), construct the difference features of the local area:

5. The method according to claim 1, characterized in that S4 specifically includes: S41, using the dual parameter method to set the detection threshold is: in, Represents the mean of the local regional difference characteristics, represents the standard deviation of local regional difference characteristics, and α represents the threshold adjustment factor; S42, binarizing the difference features of the local areas of all superpixels, obtaining pixels in the target area, and obtaining target detection results.

6. A target detection device based on local area feature differences, characterized in that: include: A segmentation module is used to perform background cancellation processing based on the image acquired by the sensor to obtain a reconstructed image; The reconstructed image is segmented into super pixels, and a Gaussian mixture model is used to fit the brightness temperature distribution of the reconstructed image to calculate the multi-order statistical characteristic parameters of each Gaussian component; A calculation module, used to calculate the Fisher vector of the superpixel according to the multi-order statistical characteristic parameters; construct a local area of ​​a preset size and a reference superpixel set based on the slow variation of the background superpixels near the target area in the reconstructed image; calculate the feature difference between each superpixel to be tested in the target area and each reference superpixel in the local area according to the Fisher vector of each superpixel, and construct a Fisher vector contrast; A construction module is used to construct a suppression feature of a background superpixel based on the brightness temperature characteristics of the target by using an extreme value suppression method; in the local area, a Fisher vector contrast is multiplied by the suppression feature of the background superpixel to obtain a difference feature of the local area; The detection module is used to perform binarization processing on the difference features of the local area using a dual parameter method, obtain pixels of the target area, and obtain target detection results.

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images