Three-dimensional reconstruction management method based on multi-view X-ray security inspection images

Abstract

The application discloses a three-dimensional reconstruction management method based on multi-view X-ray security inspection images, and belongs to the technical field of security inspection imaging. First, multi-view X-ray security inspection images of an inspected object are collected, the preprocessed multi-view X-ray images are subjected to feature extraction and fusion, image feature matching is carried out based on the fused feature vectors, and then a three-dimensional grid reconstruction model of the inspected object that passes the evaluation is constructed through an artificial intelligence algorithm. Secondly, whether the inspected object is a prohibited article is identified based on a similarity algorithm and a machine learning algorithm, parameters generated in the identification process of the inspected object are evaluated, a dangerous goods detection management evaluation index is obtained, and human-computer interaction is carried out. Through the innovative multi-view image collection, feature extraction and matching, and three-dimensional reconstruction algorithm based on sparse representation, the application can realize high-precision three-dimensional reconstruction of security inspection objects, comprehensively display the internal structure of the objects, and effectively improve the identification capability for complex objects.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of security imaging, in particular to a three-dimensional reconstruction management method based on multi-view X-ray security images. BACKGROUND

[0002] In the current society, X-ray security inspection equipment, as a key means of security inspection, has been widely used in many areas with high security requirements, such as airports, stations, customs, and large-scale events, to detect luggage, goods, and various packages, aiming to accurately identify whether there are prohibited items hidden in them, thereby effectively preventing security incidents and ensuring public safety.

[0003] Traditional X-ray security inspection equipment usually only provides two-dimensional images, which cannot fully display the internal structure of the goods, resulting in the possibility of missing prohibited items hidden in the goods by security personnel, thereby reducing the accuracy and reliability of security inspection. Although some existing technologies attempt to obtain multiple two-dimensional images by increasing the scanning angle, there are many problems in processing these multi-view two-dimensional images to reconstruct the three-dimensional structure of the goods. On the one hand, the matching and alignment between multi-view images are difficult, making it difficult to accurately find the corresponding relationship between images, thereby affecting the accuracy of three-dimensional reconstruction. On the other hand, existing three-dimensional reconstruction algorithms have high computational complexity, requiring a large amount of computing resources and time, which is difficult to meet the real-time requirements in security inspection scenarios. In addition, the reconstructed three-dimensional model lacks effective management methods, and cannot be quickly compared and analyzed with the known prohibited item model library, which is not conducive to the accurate judgment of security personnel.

[0004] Therefore, there is an urgent need for an innovative three-dimensional reconstruction management method based on multi-view X-ray security images, which realizes high-precision three-dimensional reconstruction and rapid analysis of security inspection goods through innovative image matching, three-dimensional reconstruction algorithms, and efficient model management strategies, improves the accuracy and efficiency of security inspection, effectively identifies prohibited items, and ensures public safety. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a three-dimensional reconstruction management method based on multi-view X-ray security images to solve the problems raised in the above background.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a three-dimensional reconstruction management method based on multi-view X-ray security images, comprising:

[0007] S1: synchronously scanning the inspected goods by multiple X-ray emitting sources and detectors according to an asymmetric view layout, counting X-ray images that pass the view coverage evaluation verification, and obtaining a first target multi-view X-ray image set;

[0008] S2: denoising the first target multi-view X-ray image through wavelet transform, performing histogram equalization on the denoised first target multi-view X-ray image, and obtaining a second target multi-view X-ray image set;

[0009] S3: performing feature extraction and fusion on the second target multi-view X-ray image through a scale-invariant feature transform algorithm and a principal component analysis algorithm, performing image feature matching based on the fused feature vector, and obtaining a matching point pair set that passes the image feature matching evaluation;

[0010] S4: constructing a three-dimensional point cloud model of the inspected item based on the image feature matching points through a three-dimensional reconstruction algorithm based on sparse representation, and converting the three-dimensional point cloud data into a triangular mesh model by using a marching cubes algorithm and a Laplace algorithm to obtain a three-dimensional mesh reconstruction model of the inspected item that passes the evaluation;

[0011] S5: comparing and analyzing the three-dimensional mesh reconstruction model of the inspected item with three-dimensional models of various prohibited items in a standard library based on a similarity algorithm and a machine learning algorithm, identifying whether the inspected item is a prohibited item, and obtaining an identification result of the inspected item;

[0012] S6: evaluating parameters generated in the identification process of the inspected item through a big data analysis technology, obtaining a dangerous goods detection management evaluation index, and feeding back abnormal evaluation indexes to a system administrator terminal for human-computer interaction.

[0013] Technical effects and advantages of the present application:

[0014] 1. The present application can provide high-precision three-dimensional reconstruction accurate data support for security inspection items by innovative multi-view image acquisition and fusion of local and global feature extraction algorithms, comprehensively display the internal structure of the items, make the feature description more accurate, improve the accuracy and reliability of image matching, and effectively improve the identification ability of complex items and reduce the missed detection rate of prohibited items.

[0015] 2. The three-dimensional reconstruction algorithm based on sparse representation significantly reduces the computational complexity, improves the reconstruction speed, meets the strict real-time requirements in security inspection scenes, ensures the efficient performance of security inspection work, and realizes an improved image matching algorithm by combining a random sample consensus (RANSAC) algorithm and a geometric consistency verification algorithm, thereby improving the system matching point accuracy and robustness.

[0016] 3. The present application realizes accurate prohibited item identification and evaluation of parameters generated in the identification process of the inspected item by combining a similarity algorithm and a machine learning algorithm, obtains a dangerous goods detection management evaluation index, realizes effective management and analysis of three-dimensional models, improves the accuracy and management efficiency of security inspection work, and provides strong decision support for security inspection personnel. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a schematic diagram of the overall process of the present application.

[0018] Figure 2 is a schematic diagram of the process of the present application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0020] Please refer to Figure 1 The present application provides a multi-view X-ray security inspection image three-dimensional reconstruction management system, which comprises a multi-view X-ray security inspection image acquisition module, a multi-view image preprocessing module, a multi-view image feature matching evaluation module, a multi-view image three-dimensional reconstruction evaluation module, a multi-view image three-dimensional reconstruction management module, and a multi-view image three-dimensional reconstruction management evaluation module.

[0021] The multi-view X-ray security inspection image acquisition module is connected with the multi-view image preprocessing module, the multi-view image feature matching evaluation module is connected with the multi-view image preprocessing module and the multi-view image three-dimensional reconstruction evaluation module respectively, and the multi-view image three-dimensional reconstruction management module is connected with the multi-view image three-dimensional reconstruction evaluation module and the multi-view image three-dimensional reconstruction management evaluation module respectively.

[0022] The multi-view X-ray security inspection image acquisition module: synchronously performs X-ray scanning on the inspected object according to an asymmetric view layout through a plurality of X-ray emission sources and detectors, statistically evaluates and verifies X-ray images with a view coverage, acquires a first target multi-view X-ray image set, and transmits the first target multi-view X-ray image set to the multi-view image preprocessing module;

[0023] The multi-view image preprocessing module: performs denoising processing on the first target multi-view X-ray image through wavelet transform, performs histogram equalization on the denoised first target multi-view X-ray image, acquires a second target multi-view X-ray image set, and transmits the second target multi-view X-ray image set to the multi-view image feature matching evaluation module;

[0024] The multi-view image feature matching evaluation module: performs feature extraction and fusion on the second target multi-view X-ray image through a scale-invariant feature transformation algorithm and a principal component analysis algorithm, performs image feature matching based on the fused feature vector, and transmits a matching point pair set that passes the image feature matching evaluation to the multi-view image three-dimensional reconstruction evaluation module;

[0025] The multi-view image three-dimensional reconstruction evaluation module: through the three-dimensional reconstruction algorithm of sparse representation, the three-dimensional point cloud model of the inspected item is constructed based on the image feature matching points, and then the three-dimensional grid reconstruction is performed by using the marching cubes algorithm and the Laplace algorithm, so that the three-dimensional point cloud data is converted into a triangular mesh model, and the three-dimensional grid reconstruction model of the inspected item that passes the evaluation is transmitted to the multi-view image three-dimensional reconstruction management module;

[0026] The multi-view image three-dimensional reconstruction management module: based on the similarity algorithm and the machine learning algorithm, the three-dimensional grid reconstruction model of the inspected item is compared and analyzed with the three-dimensional models of various types of prohibited items in the standard library, whether the inspected item is a prohibited item is identified, and the identification result is transmitted to the multi-view image three-dimensional reconstruction management evaluation module;

[0027] The multi-view image three-dimensional reconstruction management evaluation module: through the big data analysis technology, the parameters generated in the identification process of the inspected item are evaluated, the dangerous goods detection management evaluation index is obtained, and the abnormal evaluation index is fed back to the system administrator terminal for human-computer interaction.

[0028] Please refer to Figure 2 As shown in the figure, based on the multi-view X-ray security inspection image three-dimensional reconstruction management method, including: S1: synchronously scanning the inspected item by multiple X-ray emitting sources and detectors according to the asymmetric view layout, evaluating and verifying the X-ray images with the view coverage, obtaining a first target multi-view X-ray image set; S2: denoising the first target multi-view X-ray image by wavelet transform, and histogram equalization is performed on the denoised first target multi-view X-ray image, obtaining a second target multi-view X-ray image set; S3: feature extraction and fusion of the second target multi-view X-ray image is performed by using the scale invariant feature transform algorithm and the principal component analysis algorithm, image feature matching is performed based on the fused feature vector, and a matching point pair set that passes the image feature matching evaluation is obtained; S4: through the three-dimensional reconstruction algorithm of sparse representation, the three-dimensional point cloud model of the inspected item is constructed based on the image feature matching points, and then the three-dimensional grid reconstruction is performed by using the marching cubes algorithm and the Laplace algorithm, so that the three-dimensional point cloud data is converted into a triangular mesh model, and the three-dimensional grid reconstruction model of the inspected item that passes the evaluation is transmitted to the multi-view image three-dimensional reconstruction management module;

[0029] S1: Using multiple X-ray sources and detectors, the inspected item is simultaneously scanned with X-rays according to an asymmetrical viewing angle layout. The X-ray images that pass the viewing angle coverage evaluation and verification are statistically analyzed to obtain a first target multi-view X-ray image set I1, where I1 = [I1...]. 1 I2 1 ,...,I n 1 ], I n This represents the X-ray image from the nth viewpoint, where n represents the number of views.

[0030] In this embodiment, it should be specifically noted that the asymmetric viewpoint layout satisfies 30°≤|θ1-θ2|≤60° and at least one θ>50°. 0 The oblique viewing angle (e.g., 30° / 45° / 60°), with multiple viewing angles of n and ≥3 (preferably 5 to 7 viewing angles), and the X-ray image of each viewing angle contains the transmission information of the inspected item at that angle.

[0031] This embodiment specifically explains how the view coverage evaluation function C(p) is used to verify whether the inspected item is covered by at least three viewpoints. p represents the coordinates of any point in the space of the inspected item, V ob Ω represents the total volume of the inspected item, and Ω represents the spatial area where the inspected item is located. i Let C(p) represent the projection area of ​​the i-th viewpoint, n represent the number of multiple viewpoints, and I(·) represent the indicator function, which is 1 if the condition is true and 0 otherwise; if C(p) ≥ the corresponding threshold C(p). th If the result is positive, the verification is successful; otherwise, the verification fails, and the asymmetric viewing angle is rearranged to simultaneously perform X-ray scanning on the inspected item; the corresponding threshold C(p) is applied. th Based on historical data specific to the situation, for example in security inspection image analysis, if the coverage evaluation function is used to measure the degree of coverage of the inspected items, it may be required that C(p) ≥ 0.8 to be considered as valid, that is, at least 80% of the area-related features or information have been effectively processed.

[0032] S2: Denoising the multi-view X-ray image of the first target using wavelet transform, and then performing histogram equalization on the denoised multi-view X-ray image of the first target to obtain a set of multi-view X-ray images of the second target, including the following steps:

[0033] S2.1: First, the multi-view X-ray image of the first target is denoised using wavelet transform to obtain the denoised image I. n de I n de =W -1 (λ×W(I n 1 (x,y))), In 1 (x,y) is the pixel value of the nth view X-ray image at coordinate (x,y), W() is the wavelet transform function, λ is the threshold parameter, used to filter out important wavelet coefficients in the wavelet domain, and suppress the wavelet coefficients corresponding to the noise, a smaller λ value may retain more details, but also retain part of the noise; a larger λ value removes more noise, but may lose some image details, W -1 is the inverse wavelet function, which converts the wavelet coefficients after thresholding back to the spatial domain to obtain the denoised image I n de ; Then the denoised image I n de is subjected to a histogram equalization process to obtain the enhanced image I n en , I n en =T(I n de (x,y)), , I n de (x,y) is the pixel value of the denoised image I n de at coordinate (x,y), rk is the image gray level, , L is the number of image gray levels, h i is the frequency of gray level i, M and N represent the number of rows and columns of the image respectively, T(rk) is the gray scale transformation function, which maps the original gray level rk to a new gray level, so that the gray scale distribution of the image is more uniform, and the contrast of the image is enhanced, to obtain the enhanced image I n en ;

[0034] S2.2: Traverse the n-view X-ray images to perform wavelet transform denoising and then histogram equalization to obtain the second target multi-view X-ray image set I2, I2=[I12, I22,..., In2];

[0035] It is specifically pointed out in this embodiment that the wavelet transform function and the histogram equalization function are prior art, through the wavelet transform function, the noise can be suppressed while the significant features of the X-ray image are retained; through the histogram equalization function, the dynamic range of the denoised image can be expanded and the contrast can be improved; for the characteristics of the X-ray security image, the image details and contrast are effectively enhanced while denoising, and the gray scale range is unified, which improves the preprocessing effect and provides high-quality images for subsequent feature extraction and matching.

[0036] S3: performing feature extraction and fusion on the second target multi-view X-ray image by a scale invariant feature transform algorithm and a principal component analysis algorithm, performing image feature matching based on the fused feature vector to obtain a matching point pair set that passes the image feature matching evaluation, and comprising the following steps:

[0037] S3.1: first flattening the n-th view X-ray image I n 2 In the input scale invariant feature transform (SIFT) algorithm, the local feature f n SIFT , is obtained, where g(x, y, σ) is a Gaussian function used for smoothing the image, σ is the standard deviation of the Gaussian kernel that controls the smoothing degree, and ∇I n 2 (x, y) is the image gradient of the n-th view X-ray image. 8×4 The gradient direction statistics are arranged as a 4x4 sub-region, an 8-direction histogram, to form a 128-dimensional local feature vector; then the n-th view X-ray image I n 2 is divided into m blocks, each block having a size of p x q, and each block is flattened into an image vector matrix I n 2 (x, y) according to the row pixel gray value I n . n I m = [I1, I2,..., Im] T , where I n represents the vector matrix of the n-th view X-ray image, and I m is the image vector of the m-th block. n In the input principal component analysis (PCA) algorithm, the global feature f n PCA is obtained, where f n PCA = (I n - μ n )W n T μ n is the image mean vector of the m blocks, and W n T is the projection matrix obtained after the principal component analysis of X n . n SIFT Finally, the local feature f n PCA and the global feature f n fu are input into a feature fusion function f n fu = a1 x f n .SIFT +a2×f n PCA ,f n fu is the fusion feature of the nth view X-ray image, a1 and a2 represent the corresponding weights respectively, a1+a2=1 and a1>a2, which can be obtained by cross-validation;

[0038] It is particularly pointed out in this embodiment that each block is flattened as an image vector matrix X according to the row pixel gray value I n 2 (x,y) n Each element I(x,y) (1≤x≤p, 1≤y≤q) in the matrix represents the pixel gray value of the image at position (x,y), I n =[I(1,1),I(1,2),...,I(1,q),I(2,1),I(2,2),...,I(2,q),...,I(p,1),I(p,2),...,I(p,q)] T ; meanwhile, by innovatively fusing two kinds of features, the details and overall structure information of the image are comprehensively captured, the feature description is more accurate, and the accuracy and reliability of image matching are improved.

[0039] S3.2: First, based on image feature extraction and fusion, the second target multi-view X-ray image is traversed to obtain a feature vector set f fu , f fu =[f1 fu ,f2 fu ,...,f n fu ], the extracted feature point sets are f1, f2,..., f n ; then any two feature points (f i ,f j ) in any two feature vectors I and J in the feature vector set are substituted into the bidirectional matching distance ratio function r IJ (f), , ε is a constant to prevent the denominator from being 0, f i , f j and f k represent the feature vectors of feature points i, j and k respectively, d(·) is the Euclidean distance, f i =[x i 1 ,y i 1 ,1] T , f j =[x j 2 ,y j 2,1],for the second matching point homogeneous coordinates, screening out the r IJ (f) The feature point pair (i, j) with a value less than a threshold (e.g. 0.7) is a matching point pair; traverse all feature vector pairs (I, J), 1≤I≤J≤n, to obtain a candidate matching point pair set F(i, j) of all feature vector pairs, the number of matching point pairs is M1; secondly, input the M1 groups of matching points into the fundamental matrix function AF=0, A is an M1x9 matrix, each row corresponds to a pair of matching points, for the M1th group of matching points (f i ,f j ), A M1 =[x j 2 x i 1 ,x j 2 y i 1 ,x j 2 ,y j 2 x i 1 ,y j 2 x i 1 ,y j 2 ,x i 1 ,y i 1 ,1] A M1 represents the M1th row matrix element, singular value decomposition (SVD) is performed on the matrix A to obtain the fundamental matrix F, i.e. A=UΣV T , the fundamental matrix F is obtained by rearranging the last column v9 of V into a 3x3 matrix, then the optimal fundamental matrix F u , , I inlier () is the inlier judgment function of RANSAC, then the geometric consistency verification function I final (f i ,f j ) is used, and the output of 1 indicates that (f i ,f j ) is an inlier, and the output of 0 indicates that (f i ,f j ) is an outlier, , and e is a set threshold value for judging whether the matching point pair meets the geometric consistency;

[0040] S3.3: Traverse the candidate matching point pair set F(i,j), and use the statistical geometric consistency verification function I. final (f i ,f j The set of interior points F that outputs 1 u Given (i,j) and the number of inliers Na, calculate the feature matching success rate evaluation index FM, FM = Na / M1. If FM is greater than the corresponding threshold, it indicates that the image feature matching is good; otherwise, optimize feature extraction and image matching until FM is greater than the corresponding threshold. This yields the set of matching point pairs that have passed the image feature matching evaluation, i.e., the inlier set F. u (i,j);

[0041] This embodiment specifically explains that, firstly, the Scale Invariant Feature Transform (SIFT) algorithm is used to extract local feature points in the image, accurately describing the local details of the image. Simultaneously, Principal Component Analysis (PCA) is combined to extract global features, obtaining the overall structural information of the image. The local and global features are then fused to form a more representative feature vector. Furthermore, the RANSAC (Random Sample Consensus) algorithm and the geometric consistency verification function are combined to improve the robustness of the system.

[0042] S4: Using a sparse representation 3D reconstruction algorithm, a 3D point cloud model of the inspected item is constructed based on image feature matching points. Then, the moving cube algorithm and Laplacian algorithm are used for 3D mesh reconstruction, converting the 3D point cloud data into a triangular mesh model, resulting in a 3D mesh reconstruction model of the inspected item that has passed evaluation. This includes the following steps:

[0043] S4.1: A 3D reconstruction algorithm based on sparse representation inputs the 2D coordinates of M sets of candidate matching points into the sparse reconstruction objective function. In the middle, by solving the sparse error The minimum value is used to obtain the three-dimensional spatial coordinates of the inspected item. B is the three-dimensional spatial coordinate vector to be solved (containing x, y, z coordinate information), A is the measurement matrix, and the three-dimensional spatial coordinates are projected onto the two-dimensional image plane. γ is the regularization parameter. A suitable γ will make the solution B sparser, which helps to find a simpler representation of the three-dimensional point coordinates while meeting a certain data fitting accuracy. F s u (i,j) is the coordinate vector of M candidate matching points; when the sparsity error When the value is less than the corresponding threshold, the optimal three-dimensional spatial point coordinates are obtained, and a three-dimensional point cloud model is constructed.

[0044] In this embodiment, it should be specifically noted that the measurement matrix A can be transformed using the Direct Linear Transformation (DLT) algorithm for each 3D point X. i=[X i ,Y i ,Z i ,1] T , and its corresponding two-dimensional matching point x i =[x i ,y i ,1] T , the projection relationship is expressed as there is a non-zero scalar s i such that s i x i =AX i , obtaining the projection matrix A (3x4), and then singular value decomposition (SVD) is performed on the matrix A to obtain the final projection matrix A.

[0045] S4.2: surface reconstruction is performed by using a Marching Cubes sign(f(x,y,z)) algorithm, three-dimensional space point coordinates are converted into a three-dimensional grid reconstruction model of the detected object, f(x,y,z) is a point cloud density field function, , e1 represents an isosurface threshold value (determining the surface position), sign() represents a sign function, +1 indicates that the point (x,y,z) is inside or exactly on the isosurface, and -1 indicates that the point is outside the isosurface; then the vertex set {v1,v2,...,v n1} of the three-dimensional grid reconstruction model of the detected object is input into a Laplace smoothing algorithm, v n1 is the nth1 vertex, to obtain a smoothed vertex set {v1 p ,v2 p ,...,v n1 p}, v n1 p is the nth1 smoothed vertex, , and γ1 is a Laplace smoothing coefficient (the Laplace smoothing coefficient γ1 controls the degree of smoothing); for a given vertex v n1 , all vertices adjacent to the vertex v n1 are found by traversing all edges connected to the vertex, to form a neighborhood vertex set N(n1), v L is the coordinate vector of the lth neighborhood vertex in the neighborhood vertex set N(n1) of the vertex v n1 ; and finally, a smoothed three-dimensional grid reconstruction model of the detected object is obtained.

[0046] The embodiment needs to be specifically described as M groups of three-dimensional space point coordinates For any point (x,y,z) in space, the distance d i , i , , then the point cloud density field function , and +1 and -1 play a key role in distinguishing spatial regions, determining the position and shape of the isosurface in the marching cubes thresholding function, and are an important step in generating a smooth triangular mesh model; meanwhile, based on the three-dimensional point cloud model, the surface reconstruction is innovatively combined with the marching cubes algorithm, and the Laplace smoothing algorithm is used to optimize the triangular mesh, which not only effectively converts the point cloud data into a triangular mesh model, but also further optimizes the surface quality of the model, so that the reconstructed three-dimensional model is more smooth and realistic.

[0047] S4.3: Based on the three-dimensional grid reconstruction model of the detected object, the convex hull algorithm is used to calculate the reconstructed volume V of the detected object rec , and the volume error rate VER is obtained, VER=|V rec -V rea | / V rea , V rea is the true volume of the detected object, and the volume error rate VER is less than or equal to the corresponding threshold value, which indicates that the three-dimensional grid reconstruction model of the detected object has good accuracy, otherwise the three-dimensional grid reconstruction model is optimized in steps S4.1 and S4.2.

[0048] It needs to be specifically explained in this embodiment that the volume calculated by the convex hull algorithm is prior art, which means that the point set in three-dimensional space is wrapped to form the smallest convex polyhedron (convex hull) by using the convex hull concept in computational geometry, and the volume of the polyhedron is calculated.

[0049] S5: Based on the similarity algorithm and the machine learning algorithm, the three-dimensional grid reconstruction model of the detected object is compared and analyzed with the three-dimensional models of various types of prohibited articles in the standard library to identify whether the detected object is a prohibited article, and the identification result of the detected object is obtained, including the following steps:

[0050] S5.1: First, the similarity algorithm is used to calculate the similarity Similarity(M1,M2) between the three-dimensional grid reconstruction model M1 of the detected object and any one three-dimensional model M2 of the prohibited article in the standard library, , m1 and m2 are the number of sampling points on models M1 and M2 used to calculate the shape context descriptor, h c (M1) and h d(M2) is the shape context descriptor of the cth sampling point on model M1 and the dth sampling point on model M2, respectively, d() is a distance function (Euclidean distance or other suitable distance measurement function); then the feature vector M0 of the three-dimensional mesh reconstruction model of the inspected item is input into the trained SVM decision function Y(M), the feature vector includes the volume, vertex set and shape context descriptor of the reconstruction model, Y(M)=+1 represents contraband, Y(M)=-1 represents safe goods, , M C represents the cth support vector, N represents the number of support vectors, α C represents the Lagrange multiplier, α C >0 corresponds to the feature vector of the support vector, Y C (M) is the label of the cth support vector, that is, +1 or -1;

[0051] S5.2: Finally, whether the inspected item is a contraband item is identified by a comprehensive judgment function Alert, that is, the obtained identification result, if Alert is true, it means that it is a contraband item, if Alert is false, it means that it is a safe item, , e2 is a similarity threshold (such as 0.15), the shape matching and the classification result are fused to improve the identification accuracy;

[0052] It needs to be specifically explained in this embodiment that the shape description is a feature vector for describing the shape distribution around a point, and the machine learning algorithm support vector machine SVM model is a prior art, the feature vector M0 of the three-dimensional mesh reconstruction model of the historical inspected item and the label of the vector (including +1 and -1) are input into the training process of the support vector machine SVM model, in the training, the feature vector is processed by a kernel function K(M C , M0), M C is a support vector, M0 is a feature vector of a model to be detected, combined with the Lagrange multiplier α C , the label Y C (M) of the support vector and the bias term b, by solving the optimization problem of minimizing the identification error, a trained support vector machine SVM model is obtained.

[0053] S6: Through big data analysis technology, the parameters generated in the identification process of the inspected item are evaluated to obtain dangerous goods detection management evaluation indexes, and the abnormal evaluation indexes are fed back to the system administrator terminal for human-computer interaction, including the following steps:

[0054] S6.1: through big data analysis technology, collect the single multi-view X-ray image input in the preset management cycle to get the total processing time t (unit: seconds) of the identification result of the detected article, the number of dangerous goods nj correctly detected and the total number of actual dangerous goods ns, get the dangerous goods detection management evaluation index DR, , b1 and b2 represent the corresponding weights respectively, b1 < b2, E is the detection efficiency index, , N1 is the number of times of identification of the detected article, t a and t0 are the total processing time of the a-th detection and the single detection reference time (the reference time is obtained from historical data) respectively, G is the correct detection rate of dangerous goods, G=nj / ns, sigma (t) and mu (t) represent the standard deviation and mean value of the processing time period in the preset management cycle respectively;

[0055] S6.2: if the dangerous goods detection management evaluation index DR is greater than or equal to the corresponding threshold DR th , it indicates that the dangerous goods detection management evaluation is good, that is, the X-ray security check image three-dimensional reconstruction management is good, otherwise it indicates that the evaluation is abnormal, and the abnormal information is fed back to the system administrator terminal for human-computer interaction, the abnormal information includes abnormal evaluation index, X-ray security check image three-dimensional reconstruction model, identification result, security check time and security check place.

[0056] Secondly: the drawings in the disclosed embodiments of the application only involve the structures involved in the disclosed embodiments of the application, other structures can refer to the usual design, and in the case of no conflict, the same embodiment and different embodiments of the application can be combined with each other;

[0057] Finally: the above only describes the preferred embodiments of the application, and is not used to limit the application, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application should be included in the protection scope of the application.

Claims

1. A three-dimensional reconstruction management method based on multi-view X-ray security inspection images, characterized in that: The method comprises the following steps: S1: synchronously scanning the inspected article by multiple X-ray emitting sources and detectors according to an asymmetric view layout to obtain an X-ray image of the inspected article that passes the view coverage evaluation function verification, thereby obtaining a first target multi-view X-ray image set; S2: performing denoising processing on the first target multi-view X-ray image by wavelet transform, and performing histogram equalization on the denoised first target multi-view X-ray image, thereby obtaining a second target multi-view X-ray image set; S3: performing feature extraction and fusion on the second target multi-view X-ray image by a scale-invariant feature transform algorithm and a principal component analysis algorithm, performing image feature matching based on the fused feature vector, and obtaining a matching point pair set that passes the image feature matching evaluation; S4: constructing a three-dimensional point cloud model of the inspected article based on the image feature matching points by a three-dimensional reconstruction algorithm based on sparse representation, and converting the three-dimensional point cloud data into a triangular mesh model by a marching cubes algorithm and a Laplace algorithm, thereby obtaining a three-dimensional mesh reconstruction model of the inspected article that passes the evaluation; S5: comparing and analyzing the three-dimensional mesh reconstruction model of the inspected article with three-dimensional models of various types of prohibited articles in a standard library based on a similarity algorithm and a machine learning algorithm, identifying whether the inspected article is a prohibited article, and obtaining an identification result of the inspected article; S6: evaluating parameters generated in the identification process of the inspected article by a big data analysis technology, obtaining a dangerous goods detection management evaluation index, and feeding back abnormal evaluation indexes to a system administrator terminal for human-computer interaction. 2.The method of claim 1, wherein: The asymmetric view layout in the S1 satisfies 30°≤|θ1-θ2|≤60° and at least one θ>50°, θ1 and θ2 are the angles of any two view angles, the multi-view angle is n and ≥3; whether the detected item is covered by at least 3 view angles is verified by a view coverage evaluation function C(p), if C(p)≥ the corresponding threshold C(p) th , it is indicated that the verification is passed, otherwise the verification is failed, and the non-symmetric view angles are re-laid out to synchronously perform X-ray scanning on the detected item. 3.The method of claim 1, wherein: The feature extraction and fusion in S3: first, the n-th view X-ray image I n 2 In the input scale invariant feature transform (SIFT) algorithm, the local feature f n SIFT is obtained n 2 Then the n-th view X-ray image I n 2 is divided into m blocks, each block has a size of p x q, p and q are the length and width of the pixel block, and each block is flattened into an image vector matrix I n (x,y) according to the row pixel gray value I n (x,y), where (x,y) is the coordinate value of the pixel point, and I m =[I1,I2,...,Im] T , I n represents the vector matrix of the n-th view X-ray image, I m is the image vector of the m-th block, and the image vector matrix I n is input into the principal component analysis (PCA) algorithm to obtain the global feature f n PCA Finally, the local feature f n SIFT and the global feature f n PCA are input into the feature fusion function f n fu , f n fu =a1 x f n SIFT +a2 x f n PCA , f n fu is the fusion feature of the n-th view X-ray image, a1 and a2 represent the respective weights, a1+a2=1 and a1>a2. 4.The method of claim 1, wherein the method further comprises: The image feature matching in S3: first, based on image feature extraction and fusion, traversing the second target multi-view X-ray image, obtaining a feature vector set f fu , f fu = [f1 fu , f2 fu ,..., f n fu ], the extracted feature point set is f1, f2,..., f n ; then, substituting any two feature points (f i , f j ) of any two feature vectors I and J in the feature vector set into the bidirectional matching distance ratio function r IJ (f), ε is a constant to prevent the denominator from being 0, f i , f j and f k represent the feature vectors of feature points i, j and k respectively, d(·) is the Euclidean distance, f i = [x i 1 , y i 1 , 1] T , is the homogeneous coordinates of the first matching point, f j = [x j 2 , y j 2 , 1], is the homogeneous coordinates of the second matching point, and the feature point pair (i, j) that satisfies r IJ (f) < threshold is selected as the matching point pair; traversing all feature vector pairs (I, J), 1 ≤ I ≤ J ≤ n, n is the number of multi-view X-ray image feature vectors, obtaining a candidate matching point pair set F(i, j) of all feature vector pairs, and the number of matching point pairs is M1; secondly, inputting the M1 matching points into the fundamental matrix function AF = 0, A is an M1 × 9 matrix, each row corresponds to a pair of matching points, and singular value decomposition SVD is performed on the matrix A to obtain a fundamental matrix F, the fundamental matrix F is obtained by rearranging the last column v9 of V into a 3 × 3 matrix, f i = [x i 1 , y i 1 , 1] T , is the homogeneous coordinates of the first matching point, f j = [x j 2 , y j 2 , 1], is the homogeneous coordinates of the second matching point, and then an optimal fundamental matrix F is obtained from the fundamental matrix F by a random sample consensus RANSAC algorithm u , then by a geometric consistency verification function I final (f i ,f j ), output is 1 to indicate that (f i ,f j ) is an interior point, output is 0 to indicate that (f i ,f j ) is an exterior point. 5.The method of claim 4, wherein: The matching point pair set passed the image feature matching evaluation in S3: traverse the candidate matching point pair set F(i,j), count the geometric consistency verification function I final (f i ,f j ) The output of the inner point set F u (i,j) and the number of inner points Na, calculate the feature matching point success rate evaluation index FM, FM=Na / M1, if FM is greater than the corresponding threshold, it means that the image feature matching is good, otherwise optimize the feature extraction and image matching, until FM is greater than the corresponding threshold, and the matching point pair set passed the image feature matching evaluation, that is, the inner point set F u (i,j). 6.The method of claim 1, wherein the method further comprises: The implementation of S4 comprises: S4.1: A 3D reconstruction algorithm based on sparse representation inputs the 2D coordinates of M sets of candidate matching points into the sparse reconstruction objective function. In the middle, by solving the sparse error The minimum value is used to obtain the three-dimensional spatial coordinates of the inspected item, B is the three-dimensional spatial coordinate vector to be solved, A is the measurement matrix, the three-dimensional spatial coordinates are projected onto the two-dimensional image plane, γ is the regularization parameter, and F s u (i,j) is the coordinate vector of Na interior points; when the sparsity error When the value is less than the corresponding threshold, the optimal three-dimensional spatial point coordinates are obtained, and a three-dimensional point cloud model is constructed. S4.2: The moving cube algorithm sign(f(x,y,z)) is used for surface reconstruction, which transforms the 3D spatial point coordinates into a 3D mesh reconstruction model of the inspected object. f(x,y,z) is the point cloud density field function. e1 represents the isosurface threshold, sign() represents the conformance function, +1 indicates that the point (x,y,z) is inside the isosurface or exactly on the isosurface, -1 indicates that the point is outside the isosurface; then the vertex set {v1,v2,...,v...} of the 3D mesh reconstruction model of the inspected object is obtained. n1 The input is fed into the Laplace smoothing algorithm, v n1 For the n1-th vertex, we obtain the smoothed vertex set {v1}. p v2 p ,...,v n1 p }, v n1 p The n1th smoothed vertex is obtained; finally, the smoothed 3D mesh reconstruction model of the inspected object is obtained. 7.The method of claim 1, wherein the method further comprises: determining a three-dimensional (3D) model of the object based on the plurality of X-ray images; and determining a 3D model of the object based on the plurality of X-ray images. The implementation in the S4 further includes: S4.3: based on the three-dimensional grid reconstruction model of the inspected item, calculating the reconstruction volume V of the inspected item through a convex hull algorithm rec , obtaining a volume error rate VER, VER=|V rec -V rea | / V rea , V rea is the real volume of the inspected item, and the volume error rate VER is less than or equal to a corresponding threshold value, which indicates that the three-dimensional grid reconstruction model of the inspected item is good in accuracy, and otherwise, the three-dimensional grid reconstruction model is optimized in steps S4.1 and S4.

2. 8.The method of claim 1, wherein the method further comprises: determining a three-dimensional (3D) model of the object based on the plurality of X-ray images; and determining a 3D model of the object based on the plurality of X-ray images. The implementation of S5 comprises: S5.1: first, calculate the similarity Similarity(M1, M2) between the three-dimensional mesh reconstruction model M1 of the inspected article and any one of the three-dimensional models M2 of prohibited articles in the standard library by a similarity algorithm; then, input the feature vector M0 of the three-dimensional mesh reconstruction model of the inspected article into a trained SVM decision function Y(M), wherein the feature vector comprises the volume, vertex set and shape context descriptor of the reconstruction model, Y(M) = +1 represents a prohibited article, and Y(M) = -1 represents a safe article; S5.2: Finally, the object is identified as a prohibited item or not by the comprehensive judging function Alert, i.e. the obtained identification result, if Alert is true, it means the object is a prohibited item, if Alert is false, it means the object is a safe item, e2 is a similarity threshold value. 9.The method of claim 1, wherein the method further comprises: determining a three-dimensional (3D) model of the object based on the plurality of X-ray images; and determining a 3D model of the object based on the plurality of X-ray images. The implementation of S6 comprises: S6.1: Collecting the total processing time t of the identification result of the inspected item, the number nj of correctly detected dangerous goods and the total number ns of actual dangerous goods in a single multi-view X-ray image input in a preset management period by big data analysis technology to obtain a dangerous goods detection management evaluation index DR, b1 and b2 represent the corresponding weights respectively, b1 < b2, E is the detection efficiency index, G is the correct detection rate of dangerous goods, and σ(t) and μ(t) represent the standard deviation and mean value of the processing time period in the preset management period respectively; S6.2: If the hazardous materials inspection and management assessment index DR is greater than or equal to the corresponding threshold DR th If the assessment is satisfactory, it indicates that the management of hazardous materials detection is good, meaning that the management of 3D reconstruction of X-ray security images is good. Conversely, if the assessment is unsatisfactory, it indicates an abnormal assessment. The abnormal information will be fed back to the system administrator terminal for human-computer interaction. The abnormal information includes abnormal assessment indicators, 3D reconstruction model of X-ray security images, recognition results, security inspection time, and security inspection location.

Images