Passive millimeter wave target detection method and system based on regional clustering

Abstract

The invention belongs to the technical field of target detection, and discloses a passive millimeter wave target detection method and system based on regional clustering, and the method comprises the steps: dividing a brightness temperature image into a plurality of superpixel regions through a superpixel segmentation algorithm according to the similarity between target regions and the similarity between background regions, calculating a saliency image based on the regional statistical characteristics; according to the spatial distribution difference of boundary superpixels and internal superpixels, regional local direction centrality measurement is introduced to carry out clustering analysis, so that a candidate target region is highlighted; edge compensation is carried out on the preliminarily screened candidate target area; building a total variation optimization model by using regularization construction constraints; the local features of the compensated candidate target region are enhanced by solving the total variation optimization model; and finally, realizing accurate target detection and extraction through threshold segmentation operation. According to the invention, the accuracy of target detection and the integrity of target contour extraction can be improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of target detection, and particularly relates to a passive millimeter wave target detection method and system based on region clustering. BACKGROUND

[0002] Passive millimeter wave imaging technology has important applications in remote sensing, military monitoring, medical imaging and security inspection. However, due to the physical characteristics of the passive millimeter wave imaging mechanism and the hardware limitations of the imaging system, the passive millimeter wave image generally has problems such as low resolution, blurred target edges and missing texture information, which seriously affect the accuracy and reliability of target detection.

[0003] At present, the target detection methods for passive millimeter wave images mainly fall into two categories. One is the traditional detection method based on hand-designed features, which can extract effective information to a certain extent, but has limitations such as strong dependence on data and sensitivity to data distribution. The other is the detection method based on deep learning, which can automatically learn effective feature representation from a large amount of data without human intervention, but the generalization ability of the model still needs to be improved due to the lack of sufficient scale of labeled data set support.

[0004] Therefore, how to realize high-accuracy and high-integrity target detection and contour extraction under the physical limitations of passive millimeter wave imaging is still a key problem to be solved.

[0005] Through the above analysis, the problems and defects of the prior art are:

[0006] Due to the physical characteristics of the passive millimeter wave imaging mechanism and the hardware limitations of the imaging system, the passive millimeter wave image generally has problems such as low resolution, blurred target edges and missing texture information, which seriously affect the accuracy and reliability of target detection. SUMMARY

[0007] In view of the problems existing in the prior art, the application provides a passive millimeter wave target detection method and system based on region clustering.

[0008] The application is implemented as follows: a passive millimeter wave target detection method based on region clustering comprises:

[0009] S1: using the similarity between target regions and the similarity between background regions, the brightness temperature image is divided into a plurality of superpixel regions by a superpixel segmentation algorithm, and a saliency image is calculated based on region statistical characteristics;

[0010] S2: according to the spatial distribution difference between boundary superpixels and internal superpixels, a region local direction centrality measure is introduced to carry out clustering analysis, so that the candidate target regions are highlighted;

[0011] S3: Edge compensation is performed on the candidate target region preliminarily screened out; a total variation optimization model is established by using a regularization construction constraint; local features of the compensated candidate target region are enhanced by solving the total variation optimization model; and finally, accurate target detection and extraction are realized through a threshold segmentation operation.

[0012] Further, the S1 comprises:

[0013] S11: The multi-polarization image is normalized and fused; a segmentation threshold of a target and a background is estimated according to a brightness temperature distribution characteristic of the fused image T0, and a preliminary binary mask T is obtained BM ; and a clean human body foreground image T r is extracted in combination with a morphological processing method.

[0014] S12: An over-pixel segmentation algorithm is used to divide the fused image into a plurality of irregular over-pixel sets with similar brightness temperature and texture characteristics;

[0015] S13: On the basis of over-pixel segmentation, a region-based saliency image M s is calculated by using a statistical characteristic of the brightness temperature of the fused image; and the saliency in the lth over-pixel block is calculated as follows:

[0016]

[0017] Wherein, N l represents the size of the lth over-pixel; I l,n represents the brightness temperature intensity of the nth pixel of the lth over-pixel; and L represents the number of over-pixels.

[0018] S14: The interference of an environmental background is eliminated, and an improved saliency image is constructed, and its expression is as follows:

[0019] Further, the S2 comprises:

[0020] S21: A region local direction centrality is used to distinguish boundary over-pixels and internal over-pixels; a connection criterion is combined with a reachable distance to realize the connection of internal over-pixels in the same cluster; and finally, the boundary over-pixels are assigned to the cluster to which the nearest internal over-pixels belong, so as to complete clustering.

[0021] S22: According to the preliminary binary mask, interference areas such as a background and clutter are filtered out, and an initial candidate target region is extracted, and its expression is as follows:

[0022] R={C i |ind(P(C i ),P(T BM,j ))=1};

[0023] Wherein, Ci Ci represents the i-th cluster of the clustering result C(i = 1, 2,..., M c ), T BM,j represents the binary mask T BM (j = 1, 2,..., M p ), P(·) represents a pixel coordinate set (or a pixel coordinate set in a cluster), and ind(·) represents an indication function. Specifically, if P(T BM,j ) ∈ P(C i ), then ind(P(C i ), P(T BM,j )) is equal to 1, otherwise it is equal to 0;

[0024] Further, the S3 comprises:

[0025] S31: Compensation is performed for an edge missing area of the initial candidate target region R, and the expression is as follows:

[0026] R + = V ∪ R;

[0027] Wherein, V represents a superpixel region set adjacent to the cluster in R. For each cluster in R, its surrounding adjacent superpixel regions are selected, and these superpixel regions are merged with the cluster. Further, a local measurement image is obtained, and the expression is as follows:

[0028] T p = bm(R + ) e T0;

[0029] Wherein, bm(R + ) represents that the elements of the corresponding candidate target region in R + are 1, and the elements of the remaining positions are 0.

[0030] S32: A constraint is constructed by using regularization, and a total variation optimization model is established, and the expression is as follows:

[0031]

[0032] Wherein, μ represents a regularization parameter; ||·||2 2 represents the square of the vector norm; f0 represents a local measurement vector (f0 = vec(T p )) with a size of PQ × 1; f represents a reconstruction variable

[0033] H represents a block circulant convolution matrix with a size of PQ × PQ, which is generated by a Gaussian blur kernel; D represents a two-dimensional finite difference operator, and ||Df||2 is defined as:

[0034]

[0035] wherein, [D x(y) f] i denotes the i-th element of the vector D x(y) f.

[0036] S33: solving the total variation optimization model by using a total variation minimization algorithm to obtain a reconstructed image Threshold segmentation is performed on the reconstructed image to accurately detect the target.

[0037] Another object of the present application is to provide a passive millimeter wave target detection system based on region clustering, comprising:

[0038] A division module is configured to divide an original image into a plurality of superpixel regions by using a superpixel segmentation algorithm, and calculate a saliency image based on region statistical characteristics;

[0039] A clustering module is configured to extract an initial candidate target region by using a region local direction centrality clustering algorithm;

[0040] A detection module is configured to perform edge compensation on the initial candidate target region, construct a constraint by using regularization, establish and solve a total variation optimization model, and realize accurate target detection and extraction by threshold segmentation.

[0041] Another object of the present application is to provide a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the computer program is executed by the processor to enable the processor to perform the steps of the passive millimeter wave target detection method based on region clustering.

[0042] Another object of the present application is to provide a computer readable storage medium storing a computer program, wherein the computer program is executed by a processor to enable the processor to perform the steps of the passive millimeter wave target detection method based on region clustering.

[0043] Another object of the present application is to provide an information data processing terminal for implementing the passive millimeter wave target detection system based on region clustering.

[0044] In combination with the above technical solutions and solved technical problems, the technical solution to be protected by the present application has the following advantages and positive effects:

[0045] The present application can significantly improve the accuracy and robustness of target detection, effectively solve the problem of difficult target detection under low signal-to-noise ratio conditions, and has high application value.

[0046] The application provides a passive millimeter wave target detection method and system based on region clustering. Firstly, according to the similarity between target regions and the similarity between background regions, the image is divided into sub-regions with similar local features by a superpixel segmentation algorithm, a region-based saliency image is calculated, discrete noise is effectively suppressed, the contrast of the PMMW image is improved, and the preliminary separation of the target and the background is realized. Considering the spatial distribution difference between internal superpixels and boundary superpixels, a clustering algorithm based on the local direction centrality of the region is used to highlight the candidate target region, and the preliminary binary mask is combined to filter out interference regions such as background and clutter, so that the complete extraction of the candidate target under low resolution conditions is realized. The edge of the candidate target is compensated, and then a regularization constraint is constructed to establish a total variation optimization model. The reconstructed image is obtained by solving the total variation optimization model. The threshold segmentation of the reconstructed image is realized to accurately and reliably detect the target. The application can realize higher target detection accuracy and more complete target contour extraction, and has better detection performance in complex scenes with multiple targets, multiple scales and multiple shapes. The problems of passive millimeter wave image, such as low resolution, fuzzy target edge and missing texture information, are effectively overcome to avoid false detection and missed detection.

[0047] The existing target detection methods are mainly divided into traditional detection methods based on manually designed features and detection methods based on deep learning. The former can extract features to a certain extent, but has the limitation of over-reliance on artificial experience and specific data. The latter can automatically learn effective feature representation from a large amount of data, but due to the problems of passive millimeter wave image annotation data scarcity, data distribution imbalance and the like, the generalization ability of the model in complex scenes is insufficient. In view of these technical limitations, the application first proposes a target detection method and system based on region clustering, which realizes the aggregation of similar features by unsupervised machine learning. By introducing the region-based local direction centrality measurement, edge compensation and total variation regularization constraint, accurate detection of the target and accurate extraction of the fuzzy contour are realized, and the accuracy and robustness of the target detection are significantly improved. The application fills the technical gap in the field of passive millimeter wave image region processing at home and abroad.

[0048] It has been a key research direction in the field of target detection to realize accurate and reliable passive millimeter wave target detection. However, due to the physical characteristics and hardware limitations of passive millimeter wave imaging, the passive millimeter wave image generally has low resolution and high signal-to-noise ratio, which seriously affects the accuracy and reliability of target detection. In view of this technical problem, the present application provides an innovative detection method and system based on regional clustering, which breaks through the limitation of high false detection rate of traditional pixel-level processing under low signal-to-noise ratio. The present application designs a novel regional local direction centrality feature to measure the spatial distribution difference between internal superpixels and boundary superpixels, and the boundary superpixels constrain the distribution of internal superpixels by forming a closed curve, thereby improving the segmentation accuracy of the transition area between the target contour and the background. Further, an image reconstruction model is established by combining total variation regularization constraint, which suppresses the influence of noise on target detection while effectively protecting the boundary of the target. The present application significantly improves the accuracy of target detection and the precision of contour extraction, and provides an effective technical solution for passive millimeter wave target detection in complex scenes, which has high application value. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 is a flow chart of the passive millimeter wave target detection method based on regional clustering provided by the embodiment of the present application;

[0050] Figure 2 is a structure block diagram of the passive millimeter wave target detection system based on regional clustering provided by the embodiment of the present application;

[0051] Figure 3 is a processing flow diagram provided by the embodiment of the present application; (a) polarization average map, (b) saliency map, (c) clustering result map, (d) initial candidate target, (e) compensated candidate target, (f) reconstructed candidate target, (g) detection result;

[0052] Figure 4 is a result map of different concealed object detection methods provided by the embodiment of the present application; (a) AD-GMM, (b) SSI, (c) MP-RMFV, (d) OE-RLDC;

[0053] Figure 5 is a PR curve diagram of different concealed object detection methods provided by the embodiment of the present application;

[0054] Figure 6 is a similarity coefficient box plot of different concealed object detection methods provided by the embodiment of the present application;

[0055] Figure 7 is a result map of different target detection methods provided by the embodiment of the present application; (a) original brightness temperature map, (b) LRC, (c) OE-RLDC;

[0056] Figure 8 PR curves of different target detection methods provided by embodiments of the present invention;

[0057] Figure 9 is a box plot of similarity coefficients of different target detection methods provided by an embodiment of the present invention; DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is described in detail below in conjunction with the 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.

[0059] Existing passive millimeter-wave security inspection systems are generally limited by low signal-to-noise ratios and complex backgrounds. The target brightness temperature is only 1–3K lower than the background, or even drowned out by background noise. Therefore, traditional thresholding methods cannot achieve reliable segmentation within the speed of industrial production lines. The present invention first performs four-way orthogonal fusion in the polarization domain to suppress discrete noise and provide a continuous brightness temperature gradient for subsequent superpixel segmentation.

[0060] This embodiment of the present invention utilizes the Simple Linear Iterative Clustering (SLIC) algorithm to adaptively generate 4,000-5,000 irregular superpixels in the brightness temperature-distance feature space. Its saliency metric is based on regional statistical characteristics rather than traditional pixel gradients, effectively reducing the influence of discrete noise and sparse clutter. Measured data shows that the differential signal-to-noise ratio (DSNR) of the saliency image is 15% higher than that of the original image, significantly enhancing image quality. Furthermore, the algorithm can achieve preliminary distinction between target and background within a 0.3-second frame period in automated security inspection lanes, meeting high-speed traffic requirements.

[0061] In complex scenarios, the noise caused by the mixing of human self-emission and environmental reflections leads to a high false alarm rate. To address this, the present invention uses unsupervised clustering on a regional saliency map, based on a regional local directional centrality metric and connectivity criteria, to automatically group superpixels with similar characteristics into the same cluster. This is further combined with a preliminary binary mask to filter out interfering areas such as background and clutter, resulting in initial candidate targets. After dilation compensation of the initial candidate targets using a 3×3 structuring element, the similarity coefficient between the detected target and the true structure is improved, meeting the industry standard for public safety detection level II.

[0062] The passive imaging system's antenna aperture limits the original spatial resolution to only about 2.0×2.0 cm, resulting in blurred target edges and missing target textures. This approach will use total variation regularization to constrain the gradient and establish The reconstruction model of the image is constructed; the alternating direction multiplier method (ADMM) is used to iteratively solve the problem and enhance small targets such as pliers and ceramic knives.

[0063] like Figure 1As shown, the present invention provides a passive millimeter wave target detection method based on regional clustering, comprising:

[0064] S1: Using the similarity between target regions and background regions, the brightness temperature image is divided into multiple superpixel regions through a superpixel segmentation algorithm, and a saliency image is calculated based on the regional statistical characteristics;

[0065] S2: Based on the spatial distribution differences between boundary superpixels and internal superpixels, the regional local direction centrality metric is introduced to conduct cluster analysis, so that the candidate target area is highlighted;

[0066] S3: Perform edge compensation on the candidate target areas that have been initially screened; use regularization to construct constraints and establish a total variation optimization model; by solving the total variation optimization model, enhance the local features of the compensated candidate target areas; finally, achieve accurate target detection and extraction through threshold segmentation operations.

[0067] In one embodiment, S1 includes:

[0068] S11: Normalize and fuse the multi-polarization image; estimate the target and background segmentation threshold according to the brightness temperature distribution characteristics of the fused image T0, and obtain the preliminary binary mask T BM ; Combined with morphological processing methods, extract the clean human foreground image T r .

[0069] S12: Use superpixel segmentation algorithm to divide the fused image into many irregular superpixel sets with similar brightness and texture characteristics;

[0070] S13: Based on superpixel segmentation, the statistical characteristics of the brightness temperature of the fused image are used to calculate the region-based saliency image M s ; The saliency within the lth superpixel block is calculated as follows:

[0071]

[0072] Among them, N l represents the size of the lth superpixel; I l,n represents the brightness temperature intensity of the nth pixel of the lth superpixel; L represents the number of superpixels;

[0073] S14: Eliminate the interference of the environmental background and construct an improved saliency image, which is expressed as:

[0074] Specifically, (1) multiple orthogonal polarization brightness temperature images are polarized and initially segmented, and sub-regions with similar local features are extracted using a superpixel segmentation algorithm. Then, a regional saliency image is calculated based on the statistical characteristics of the fused image. The specific steps are as follows:

[0075] (1-1)Collecting multiple polarization images, performing polarization average processing on 0°, 45°, 90° and 135° linear polarization images; according to the brightness temperature distribution characteristics of the fusion image T0, estimating the segmentation threshold of the candidate target region and the background, and obtaining a preliminary binary mask T BM ; and combining morphological hole filling and corrosion operation, extracting a clean human body foreground image T r .

[0076] (1-2) The fusion image is divided into a plurality of irregular superpixel sets with similar brightness temperature and texture characteristics by using a superpixel segmentation algorithm. In the present application, the superpixel segmentation algorithm is realized by a simple linear iterative clustering (SLIC) algorithm;

[0077] (1-3) On the basis of superpixel segmentation, a region-based saliency image M s is calculated according to the statistical characteristics of the brightness temperature of the fusion image. The saliency in the lth superpixel block is calculated as follows:

[0078]

[0079] Wherein, N l represents the size of the lth superpixel; I l,n represents the brightness temperature intensity of the nth pixel of the lth superpixel; and L represents the number of superpixels.

[0080] (1-4) Eliminate the interference of the environment background, and construct an improved saliency image, the expression is as follows:

[0081]

[0082] In one embodiment, S2 includes:

[0083] S21: distinguishing boundary superpixels and internal superpixels by using regional local direction centrality measurement; connecting internal superpixels in the same cluster by combining reachable distance and connection criteria; and finally assigning the boundary superpixels to the cluster to which the nearest internal superpixel belongs, thereby completing clustering;

[0084] S22: filtering out interference areas such as background and clutter according to the preliminary binary mask, and extracting an initial candidate region, the expression is as follows:

[0085]

[0086] Wherein, C i represents the ith cluster of the clustering result C(i=1, 2,..., M c ); and T BM,j represents the binary mask T BM (j=1, 2,..., Mp ), P(·) represents a pixel coordinate set (or a pixel coordinate set in a cluster), and ind(·) represents an indication function. Specifically, if P(T BM,j ) e P(C i ), then ind(P(C i ), P(T BM,j )) is equal to 1, otherwise equal to 0;

[0087] Specifically, (2) highlights the candidate target region by using an unsupervised region clustering algorithm, and filters out interference regions such as background and clutter in combination with a preliminary binary mask to obtain an initial candidate target region. The specific steps are as follows:

[0088] (2-1) highlights the candidate target region by using an unsupervised region clustering algorithm based on local direction centrality;

[0089] (2-1-1) introduces a novel region local direction centrality feature for measuring the spatial distribution difference between internal superpixels and boundary superpixels. In two-dimensional space, the region local direction centrality feature is the directional angle variance calculated based on the KNN algorithm, and the expression is as follows:

[0090]

[0091] wherein K represents the first K nearest superpixels determined by the KNN algorithm, and a i represents the directional angle formed by the i-th nearest superpixel and the center superpixel. The internal superpixels and the boundary superpixels are divided into L1 internal superpixels and L-L1 boundary superpixels by setting the proportion parameter r=L1 / L;

[0092] (2-1-2) carries out region clustering based on reachable distance and connection criteria. The minimum distance of the internal superpixels and all boundary superpixels is defined as the reachable distance, and the expression is as follows:

[0093]

[0094] wherein d(i,j) represents the distance measure of superpixel i and superpixel j.

[0095] The distance measure comprehensively considers the features of two dimensions of spatial distance and saliency distance, and the expression is as follows:

[0096]

[0097] wherein w sp represents the weight of the spatial distance, w si represents the weight of the saliency distance; (x i , y i ) and (xj ,y j ) represent the geometric centers of the i-th superpixel and the j-th superpixel respectively; m i and m j denote the saliency values ​​of the i-th superpixel and the j-th superpixel, respectively.

[0098] When the distance between two internal superpixels satisfies the connection criterion, the two superpixels can be considered to belong to the same cluster, that is, the distance between the two superpixels is less than the sum of their reachable distances, which is expressed as follows:

[0099] d(i,i')≤D i +D i' ;

[0100] Where d(i,i') represents the distance metric between superpixel i and superpixel i'.

[0101] The boundary superpixels are assigned to the cluster to which the nearest interior superpixel belongs, thus completing the clustering.

[0102] (2-2) Based on the preliminary binary mask, interference areas such as background and clutter are filtered out. The expression is as follows:

[0103] R={C i |ind(P(C i ),P(T BM,j ))=1};

[0104] Among them, C i Represents the clustering result C(i=1,2,...,M c )’s i-th cluster, T BM,j Represents the binary mask T BM (j=1,2,...,M p ), P(·) represents the pixel coordinate set (or the pixel coordinate set in a cluster), and ind(·) represents the indicator function. Specifically, if P(T BM,j )∈P(C i ), then ind(P(C i ),P(T BM,j )) is equal to 1, otherwise it is equal to 0.

[0105] In one embodiment, the S3 includes:

[0106] S31: Compensate for the edge missing area of ​​the initial candidate target region R. The expression is as follows:

[0107] R + =V∪R;

[0108] Where V represents the set of superpixel regions adjacent to the cluster in R. For each cluster in R, the superpixel regions immediately surrounding it are selected and merged with the cluster. The local measurement image is further obtained as follows:

[0109] T p =bm(R + )e T0;

[0110] Among them, bm(R + ) means to make R + The elements corresponding to the candidate target area are 1, and the elements at other positions are 0.

[0111] S32: Use regularization to construct constraints and establish a total variation optimization model. The expression is as follows:

[0112]

[0113] where μ represents the regularization parameter; ||·|| 2 represents the square of the vector norm; f0 represents the local measurement vector of size PQ×1 (f0=vec(T p )); f represents the reconstruction variable with size PQ×1

[0114] H represents a block-circular convolution matrix of size PQ×PQ, generated by a Gaussian blur kernel; represents the two-dimensional finite difference operator, ||Df||2 is defined as:

[0115]

[0116] Among them, [D x(y) f] i Represents vector D x(y) the i-th element of f;

[0117] S33: Solve the total variation optimization model using the total variation minimization algorithm to obtain the reconstructed image Threshold segmentation is performed on the reconstructed image to accurately detect the target.

[0118] like Figure 2 As shown, the present invention provides a passive millimeter wave target detection system based on regional clustering, comprising:

[0119] A segmentation module is used to divide the original image into multiple super-pixel regions using a super-pixel segmentation algorithm; and calculate a saliency image based on regional statistical characteristics;

[0120] Clustering module, used to extract initial candidate target regions using regional local direction centrality clustering algorithm;

[0121] The detection module is used for edge compensation on the initial candidate target region; meanwhile, a constraint is constructed by regularization, a total variation optimization model is established and solved, and accurate target detection and extraction are realized by threshold segmentation.

[0122] The method provided by the application can select multi-polarization image data collected by a W-waveband passive imaging system as a specific implementation case. First, orthogonal linear polarization images are used for normalization and polarization average processing, and a preliminary binary mask is obtained by further simple threshold segmentation, and after morphological processing, a clean human body foreground image is obtained. Figure 3 (a) shows a polarization average image.

[0123] Note that in specific implementation, other image fusion algorithms can also be used to enhance the image.

[0124] The interference of discrete noise is suppressed, and the fusion image is divided into a plurality of irregular superpixel sets with similar brightness temperature and texture characteristics by using a simple linear iterative clustering (SLIC) algorithm. On the basis of superpixel segmentation, according to the statistical characteristics of the brightness temperature of the fusion image, a region saliency image is calculated, and the influence of the environment background is eliminated, Figure 3 (b) shows a superpixel-based saliency image.

[0125] Note that in specific implementation, other superpixel segmentation algorithms can also be used to implement superpixel segmentation of the image.

[0126] Based on the saliency image obtained in step (2), the candidate target region is highlighted by using a clustering algorithm (OE-RLDC) based on the local direction center of the region, and the background and interference regions such as clutter are filtered out in combination with the preliminary binary mask, and the clustering result is as shown in Figure 3 (c) shows an initial candidate target region as shown in Figure 3 (d).

[0127] The initial candidate target region is edge compensated to restore the complete candidate target region, and the compensated candidate target region is as shown in Figure 3 (e). A total variation optimization model is established by using regularization to construct a constraint, considering the characteristics that the brightness temperature of the passive millimeter wave image continuously changes and the noise is in a Gaussian distribution, the vector norm of the objective function is selected as L2 norm.

[0128] Note that in specific implementation, other norms can also be used to establish the total variation optimization model.

[0129] The alternating direction multiplier method (ADMM) is used to solve the total variation / L2 norm optimization model in this embodiment, and a reconstructed image is obtained, as shown in Figure 3(f) shown. A simple threshold segmentation process is performed on the reconstructed image, and the target can be accurately extracted, as shown in Figure 3 (g) shown.

[0130] Note that in the specific implementation, other optimization solving methods can also be used to solve the total variation optimization model.

[0131] To verify the advantages of the present application, comparative experiments of the present application and other passive millimeter wave target detection methods in the Jaccard coefficient, detection accuracy (ACC) and precision-recall (PR) curve area (AUC) are set. The data of the present embodiment is used as the implementation case. In addition to the target detection method (OE-RLDC) proposed in the present application, other passive millimeter wave target detection methods are Gaussian mixture model (AD-GMM), saliency (SSI) and region Fisher vector (MP-RMFV). The detection results corresponding to AD-GMM, SSI, MP-RMFV and OE-RLDC are shown in Figure 4 , and the quantitative comparison results are shown in Table 1.

[0132] Table 1 Performance comparison of different target detection methods in the embodiment of the present application

[0133]

[0134]

[0135] The detection results obtained in the embodiments of the present application show that the passive millimeter wave target detection method based on region clustering (OE-RLDC) proposed above has the highest Jaccard coefficient, the highest detection accuracy and the largest PR curve area. Further analysis of the detection results shows that the passive millimeter wave target detection method based on region clustering performs better in complex scenes with multiple targets, multiple scales and multiple shapes, and can more completely extract the target contour, effectively overcoming the false detection and missed detection caused by the low resolution of passive millimeter wave images, the fuzzy target edge and the lack of texture information.

[0136] The present application can be specifically applied to passive millimeter wave security inspection systems and airborne interferometric radiometer ship detection fields.

[0137] To further verify the technical advantages of the present application, statistical analysis is carried out using measured data. A challenging test scene is constructed in the experiment: 1) in terms of sample setting, three categories of 20 typical concealed objects are selected, including weapons (such as pistols, knives), dangerous goods (such as bottled alcohol), and daily necessities (such as mobile phones, Bluetooth earphones); 2) in terms of scene configuration, each tester carries 3-4 different concealed objects, which are randomly placed on different parts of the human body. In addition, the testers pass through the security equipment in various postures, such as natural standing, hands raised, hands behind the back, etc. This statistical experiment uses 20 groups of measured data as test cases. Figure 5 The PR curves of each detection method are shown, among which the PR curve of the present application method is closest to the upper right corner, showing the best overall detection performance. Figure 6 The box plot of the similarity coefficients of each detection method is shown. It can be seen that the present application method has the highest average similarity coefficient, which is 27.37% higher than AD-GMM, 29.29% higher than SSI, and 25.69% higher than MP-RMFV, indicating the advantages of the present application method in detection accuracy. At the same time, from Figure 6 It can also be observed that the present application method presents the most compact data distribution on all test cases, i.e. with the smallest variance, which fully demonstrates that the present application method has better robustness and stability. In addition, the average execution time of the present application method on all test cases is within sub-second, meeting the real-time requirements of passive millimeter wave concealed object detection systems in complex scenes.

[0138] In addition, the applicability of the present application method in long-distance target imaging detection tasks is verified, and the target detection results are compared with the most advanced detection method based on local region contrast (LRC), as shown in Figure 7 From Figure 7 it can be seen that the present application method can still detect the target completely under low signal-to-noise ratio conditions. Quantitative analysis results show that the present application method reaches 77.93% in JAC index, which is 8.61% higher than LRC; reaches 0.93 in AUC index, which is 0.15 higher than LRC; and reaches 99.74% in ACC index, which is slightly better than LRC. To further verify the robustness of the present application method in target imaging detection tasks, statistical analysis is carried out using multiple groups of data. From Figure 8 it can be seen that the present application method has a higher PR curve, indicating that it has better overall detection performance. From Figure 9 it can be seen that the present application method has a higher average similarity coefficient, which is 8.03% higher than LRC, indicating that the present application method has stronger detection accuracy. At the same time, from Figure 9It can also be observed that the method of the present application presents a smaller variance, which fully proves that the method of the present application is more robust. In addition, the average execution time of the method of the present application is within sub-second level, which meets the real-time requirement of specific application tasks.

[0139] It should be noted that the embodiments of the present application can be realized by hardware, software, or a combination of software and hardware. The hardware part can be realized by special logic; the software part can be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or a specially designed hardware. Those skilled in the art can understand that the above-mentioned devices and methods can be realized by computer executable instructions and / or included in processor control codes, for example, such codes are provided on a carrier medium, such as a magnetic disk, a CD or a DVD-ROM, a programmable memory, such as a read-only memory (firmware), or a data carrier, such as an optical or electronic signal carrier. The devices of the present application and their modules can be realized by hardware circuits, such as very large scale integrated circuits or gate arrays, semiconductors, such as logic chips, transistors, etc., or programmable hardware devices, such as field programmable gate arrays, programmable logic devices, etc., by software executed by various types of processors, or by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0140] The above description is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any modification, equivalent replacement and improvement within the technical range disclosed by the present application and within the spirit and principle of the present application should be covered within the protection scope of the present application.

Claims

1. A passive millimeter wave target detection method based on regional clustering, characterized in that: The following steps are involved: S1: Using the similarity between target regions and background regions, the brightness temperature image is divided into multiple superpixel regions through a superpixel segmentation algorithm, and a saliency image is calculated based on the regional statistical characteristics; S2: Based on the spatial distribution differences between boundary superpixels and internal superpixels, the regional local direction centrality metric is introduced to conduct cluster analysis, so that the candidate target area is highlighted; S3: Perform edge compensation on the candidate target areas that have been initially screened; use regularization to construct constraints and establish a total variation optimization model; and enhance the local features of the compensated candidate target areas by solving the total variation optimization model. Finally, accurate target detection and extraction are achieved through threshold segmentation operation.

2. The method according to claim 1, wherein The significance value of each superpixel block in step S1 is the arithmetic average of the brightness temperature intensities of all pixels in the block; Said S1 comprises: S11: Normalize and fuse the multi-polarization image; estimate the target and background segmentation threshold according to the brightness temperature distribution characteristics of the fused image T0, and obtain the preliminary binary mask T BM ; Combined with morphological processing methods, extract the clean human foreground image T r ; S12: Use superpixel segmentation algorithm to divide the fused image into many irregular superpixel sets with similar brightness and texture characteristics; S13: Based on superpixel segmentation, the statistical characteristics of the brightness temperature of the fused image are used to calculate the region-based saliency image M s ; The saliency within the lth superpixel block is calculated as follows: Among them, N l represents the size of the lth superpixel; I l,n represents the brightness temperature intensity of the nth pixel of the lth superpixel; L represents the number of superpixels; S14: Eliminate the interference of the environmental background and construct an improved saliency image, which is expressed as:

3. The method according to claim 1, wherein The region clustering in step S2 uses the distance metric and the saliency metric as the comprehensive distance metric and applies the connection criterion to divide the superpixel set into several clusters; further combined with the preliminary binary mask to filter out the background and clutter interference areas; The S2 includes: S21: Use the regional local directional centrality metric to distinguish boundary superpixels from internal superpixels; combine the reachable distance and connection criterion to connect internal superpixels in the same cluster; finally, assign the boundary superpixel to the cluster to which the closest internal superpixel belongs, thus completing the clustering; S22: Filter out the background and clutter interference areas based on the preliminary binary mask and extract the initial candidate target area. The expression is as follows: R={C i |ind(P(C i ),P(T BM,j ))=1}; Among them, C i Represents the clustering result C(i=1,2,...,M c )’s i-th cluster, T BM,j Represents the binary mask T BM (j=1,2,...,M p )’s j-th pixel, P(·) represents a pixel coordinate set, or a pixel coordinate set in a cluster, and ind(·) represents an indicator function; if P(T BM,j )∈P(C i ), then ind(P(C i ),P(T BM,j )) is equal to 1, otherwise it is equal to 0.

4. The method according to claim 1, wherein The edge compensation in step S3 and the total variation regularization are used to establish an optimization model, which consists of two parts: The data fidelity term is the weighted sum of the squared errors between the reconstruction vector and the local observation vector; The regularization term is the weighted sum of the gradient of the reconstruction vector under the norm of one; The ratio between the two is controlled by the regularization parameter, and the optimal solution is obtained through the total variation iterative algorithm; The S3 includes: S31: Compensate for the edge missing area of ​​the initial candidate target region R. The expression is as follows: R + =V∪R; Where V represents the set of superpixel regions adjacent to the cluster in R. For each cluster in R, the superpixel regions immediately surrounding it are selected and merged with the cluster. The local measurement image is further obtained as follows: T p =bm(R + )e T0; Among them, bm(R + ) means to make R + The elements corresponding to the candidate target area are 1, and the elements at other positions are 0; S32: Use regularization to construct constraints and establish a total variation optimization model. The expression is as follows: where μ represents the regularization parameter; ||·|| 2 represents the square of the vector norm; f0 represents the local measurement vector of size PQ×1 (f0=vec(T p )); f represents the reconstruction vector of size PQ×1 H represents a block-circular convolution matrix of size PQ×PQ, generated by a Gaussian blur kernel; represents the two-dimensional finite difference operator, ||Df||2 is defined as: Among them, [D x(y) f] i Represents vector D x(y) the i-th element of f; S33: Solve the total variation optimization model using the total variation minimization algorithm to obtain the reconstructed image Threshold segmentation is performed on the reconstructed image to accurately detect the target.

5. A passive millimeter wave target detection system based on regional clustering according to the method of any one of claims 1 to 4, characterized in that: include: A segmentation module, used to divide the original image into multiple super-pixel regions using a super-pixel segmentation algorithm; And calculate the saliency image based on regional statistical characteristics; Clustering module, used to extract initial candidate target regions using regional local direction centrality clustering algorithm; The detection module is used to perform edge compensation on the initial candidate target area; at the same time, it uses regularization to construct constraints, establishes and solves the total variation optimization model; and achieves accurate target detection and extraction through threshold segmentation.

6. The system according to claim 5, wherein: The segmentation module adopts a superpixel segmentation algorithm based on brightness temperature-distance and gives a saliency score to each superpixel in the saliency calculation unit.

7. The system according to claim 5, wherein: The clustering module includes an unsupervised clustering unit and an interference area filtering unit.

8. The system according to claim 5, wherein: The detection module includes an edge compensation unit and a total variation regularization unit.

9. A computer-readable storage medium having instructions stored thereon, which, when executed by a processor, causes the processor to execute the method according to any one of claims 1 to 4.

Images