An explosive recognition method

Abstract

The present invention relates to an explosive identification method, belonging to the technical field of security inspection, and solves the problem that existing methods cannot separately identify explosives in different states. The method includes: an artificial intelligence identifies the bounding box of a container; segments three-dimensional objects in three-dimensional volume data and obtains the bounding box of each three-dimensional object; identifies explosives: matches the bounding box of each three-dimensional object with the bounding box of the container identified by the artificial intelligence to determine whether the three-dimensional object is wrapped by a container; if so, the three-dimensional object enters the liquid and powdery explosive identification library for matching; if not, the three-dimensional object enters the solid explosive identification library for matching. This method can quickly identify liquid and powdery explosives and solid explosives separately, improving the accuracy of explosive identification.

Description

Technical Field

[0001] The present invention relates to the technical field of security inspection, and in particular, to a method for identifying explosives. Background Art

[0002] With the increasing demand for public security, security inspections are required in crowded public areas such as subways and airports. There are many types of explosives, including liquid explosives, powdered explosives, and solid explosives according to types. At present, the main technical methods for detecting explosives are divided into two categories: trace explosive detection and bulk explosive detection. Among them, trace explosive detection refers to the technology of sampling and analyzing trace or ultratrace explosive residues. The basic principle is that explosives will always leave residues in the form of gases or solid particles during the processing. These residues are collected and analyzed using relevant detection technologies to determine whether explosives exist. Common trace explosive detection technologies include electrochemical technology and laser Raman spectroscopy technology.

[0003] Bulk explosive detection refers to detecting visible amounts of explosives. It usually includes X-ray, γ-ray imaging technology and nuclear-based technology. Both X-ray and γ-ray detect characteristic quantities such as the density and atomic number of substances through high-energy electromagnetic waves. Nuclear technology mainly includes nuclear quadrupole resonance technology and neutron technology.

[0004] Dual-energy CT technology reconstructs tomographic images of the scanned object containing density and atomic number from the projection data of X-rays. By analyzing these characteristic data, hidden objects can be effectively identified, and dangerous items in the scanned object can be recognized. However, density and atomic number are not the only attributes of explosives. Some common safety items may also have similar density and atomic number to explosives. If the state differences of explosives are not considered, the false alarm rate will increase significantly. Summary of the Invention

[0005] In view of the above analysis, embodiments of the present invention aim to provide a method for identifying explosives to solve the problem that existing methods cannot separately identify explosives in different states.

[0006] On the one hand, embodiments of the present invention provide a method for identifying explosives, and the method for identifying explosives includes:

[0007] Artificially intelligent recognition of the bounding box of the container;

[0008] Segmenting three-dimensional objects in three-dimensional volume data and obtaining the bounding box of each three-dimensional object;

[0009] Identifying explosives: Match the bounding box of each three-dimensional object with the bounding box of the container identified by the artificial intelligence to determine whether the three-dimensional object is wrapped by a container; if so, the three-dimensional object enters the liquid and powdery explosive identification library for matching; if not, the three-dimensional object enters the solid explosive identification library for matching.

[0010] Preferably, the artificial intelligence identifies the bounding box of the container, including: performing real-time online detection through CT to obtain three-dimensional orthographic DR images at different angles; using the artificial intelligence model to identify the DR images, and performing intersection operation based on the identification results at different angles to obtain the bounding box of the three-dimensional identification of the container.

[0011] Preferably, before the real-time online detection through CT, artificial intelligence model training is performed, including:

[0012] (1) Obtaining data identified by the artificial intelligence

[0013] According to the density and atomic number data reconstructed by CT, perform orthographic projection at different angles to obtain high- and low-energy projection data, and perform color assignment to obtain a colored DR image;

[0014] (2) Artificial intelligence model training

[0015] Pre-collect the CT reconstruction data of the container. Using the DR image obtained in step (1) as the training set, use the artificial intelligence model to perform prediction to obtain the spatial position of the container; calculate the loss function. If the loss function meets the requirements, then the artificial intelligence model in the current state can be used for container identification; if the loss function does not meet the requirements, then correct the artificial intelligence model and perform identification again until the conditions are met.

[0016] Preferably, segment the three-dimensional objects in the three-dimensional volume data and obtain the bounding box of each three-dimensional object, including:

[0017] Obtain the projection data of different angles of the scanned piece through dual-energy CT scanning, reconstruct the projection data to obtain multiple two-dimensional tomograms, and sequentially number the two-dimensional tomograms;

[0018] Segment the two-dimensional objects in each two-dimensional tomogram, obtain the features of each two-dimensional object in each two-dimensional tomogram, and sequentially number the two-dimensional objects in each two-dimensional tomogram;

[0019] Analyze and judge the three-dimensional body connectivity to obtain the connectivity result, complete the three-dimensional object segmentation, and obtain the features of each three-dimensional object;

[0020] The starting position and ending position of each tomography of a three-dimensional object are defined as Z1 and Z2 of the bounding box; the union of the minimum rectangular boxes of each tomography is taken, and the resulting result is defined as [X1, Y1, X2, Y2]; the bounding box of the final three-dimensional object is [X1, Y1, Z1, X2, Y2, Z2]; where X1, Y1, Z1, X2, Y2, and Z2 are the x, y, and z coordinates of the starting point and ending point of the minimum circumscribed cube of the three-dimensional object respectively.

[0021] Preferably, the image preprocessing includes image smoothing, image enhancement, image region segmentation, and dilation and erosion.

[0022] Preferably, the features of the two-dimensional object include mean density, mean square deviation of density, mean atomic number, mean square deviation of atomic number, object area, object perimeter, abscissa of the regional centroid, ordinate of the regional centroid, and minimum circumscribed rectangle.

[0023] Preferably, the analysis for determining the connectivity of the three-dimensional body includes: based on the k-th two-dimensional object M in the i-th two-dimensional tomography, traverse each two-dimensional object in the (i - 1)-th two-dimensional tomography, and determine whether the differences in the central position, area, average density, and average atomic number between each two-dimensional object in the (i - 1)-th two-dimensional tomography and the two-dimensional object M ik meet the threshold requirements; if there is a two-dimensional object in the (i - 1)-th two-dimensional tomography that meets the threshold requirements, then according to the differences in central position, area, average density, and average atomic number, a comprehensive score is given to the two-dimensional object that meets the threshold requirements. The smaller the difference, the higher the score. The two-dimensional object with the highest comprehensive score is the connected region in the (i - 1)-th two-dimensional tomography that is connected to the two-dimensional object M ik ; if there is no two-dimensional object in the (i - 1)-th two-dimensional tomography that meets the threshold requirements, then traverse each two-dimensional object in the (i - j)-th two-dimensional tomography and repeat the above steps; where, i ≥ 2, k ≥ 1, i ≥ j ≥ 2; j takes values from small to large, and only when there is no two-dimensional object in the currently selected two-dimensional tomography that meets the threshold requirements, j further takes a value 1 greater than the current value; repeat all the steps of the above connectivity analysis until the connectivity analysis of each two-dimensional object in each two-dimensional tomography is completed; the connected two-dimensional objects are combined to form a three-dimensional object, and the three-dimensional object segmentation is completed. ik

[0024] Preferably, the method for matching explosives includes an autonomous judgment criterion or a judgment method using artificial intelligence.

[0025] Preferably, the explosive recognition method further includes result display. If the matching result is an explosive, then the output result shows the location of the explosive and the name, weight, volume, and density of the predicted explosive; if the matching results are not explosives, then the result is directly output.

[0026] Preferably, the position of the explosive can be determined according to the bounding box of the obtained three-dimensional object.

[0027] Compared with the prior art, the present invention can at least achieve the following beneficial effects:

[0028] 1. For the existing security inspection CT equipment, the present invention uses artificial intelligence to identify the container and output the spatial position where the container is located; segment the three-dimensional objects in the three-dimensional volume data to obtain the bounding box of the three-dimensional objects, and match the position features of the three-dimensional objects with the container position identified by artificial intelligence; determine whether each object is wrapped by the container. If so, it is a liquid or powdery object, and perform matching for liquid and powdery explosives; otherwise, it is a solid, and perform matching for solid explosives. This method can identify liquid and powdery explosives and solid explosives separately, improving the accuracy of explosive identification.

[0029] 2. The three-dimensional object segmentation method of the present invention performs connectivity judgment on each two-dimensional object in the two-dimensional tomogram as the object of connectivity judgment, and finally combines the two-dimensional objects with connectivity to form a three-dimensional object, that is, the present invention is two-dimensional region growing. Compared with the three-dimensional region growing method in the prior art, the present invention reduces the range of region growing from 26 pixels in three dimensions to 8 pixels in two dimensions. Therefore, the three-dimensional object segmentation method of the present invention is faster, and thus the explosive identification is faster.

[0030] 3. Compared with the threshold method for segmenting three-dimensional objects in the prior art, due to the presence of artifact phenomena in CT images, the threshold method cannot accurately segment all objects. However, the present invention is two-dimensional region growing, which can accommodate differences and reduce the influence of artifacts. Therefore, the three-dimensional object segmentation method of the present invention has higher accuracy in segmenting three-dimensional objects, and thus the explosive identification is more accurate.

[0031] 4. Compared with the boundary method in the prior art, since the boundary method requires first-order or second-order differential calculations and iterative operations, while the present invention does not require any iteration. Therefore, the three-dimensional object segmentation method of the present invention is faster, and thus the explosive identification is faster.

[0032] In the present invention, the above technical solutions can also be combined with each other to achieve more preferred combination schemes. Other features and advantages of the present invention will be described in the subsequent specification, and some advantages can be made obvious from the specification or understood by implementing the present invention. The objectives and other advantages of the present invention can be achieved and obtained from the content specifically pointed out in the specification and the drawings. Description of the Drawings

[0033] The accompanying drawings are only for the purpose of showing specific embodiments and are not considered to be a limitation of the present invention. Throughout the drawings, the same reference signs denote the same components.

[0034] Figure 1 It is a flow chart for the present invention to identify explosives;

[0035] Figure 2 It is a flow chart of the three-dimensional object segmentation method in the three-dimensional volume data of the present invention;

[0036] Figure 3 It is a schematic diagram before and after the number change of each two-dimensional object after connectivity analysis using the three-dimensional object segmentation method in the three-dimensional volume data of the present invention. Detailed Description of the Invention

[0037] The following will specifically describe the preferred embodiments of the present invention with reference to the accompanying drawings. The accompanying drawings form a part of this application and are used together with the embodiments of the present invention to explain the principles of the present invention, rather than to limit the scope of the present invention.

[0038] In the field of security inspection, the technology of identifying explosives through dual-energy CT technology calculates the average electron density and atomic number corresponding to each item through dual-energy CT technology, and then compares them with the known explosive information to determine whether it is an explosive. However, there are many types of explosives. If the state of the object is not considered, there will be many false alarms for explosives. Therefore, identifying the corresponding state of the explosives is the key.

[0039] Moreover, during the security inspection process, after the CT scan of the package is completed, it is conveyed to the user through a conveyor belt. In this short process, the explosives need to be identified. Therefore, rapid three-dimensional segmentation and identification of explosives for the items are necessary.

[0040] Thus, the present invention provides an explosive identification method, as Figure 1 shown, the explosive identification method includes:

[0041] 1. Artificial intelligence identifies the bounding box (spatial position) of the container

[0042] S1.1: Obtain the data identified by artificial intelligence

[0043] According to the density and atomic data reconstructed by CT, perform forward projections at different angles (such as 0 degrees, 45 degrees, 90 degrees, 135 degrees, 180 degrees) to obtain high- and low-energy projection data, and perform color assignment to obtain a colored DR image.

[0044] Among them, different colors represent different materials. Exemplarily, organic substances such as food and plastic are shown as orange; inorganic substances such as books and ceramics are shown as green; mixtures also mostly appear green, and stainless steel products and metals are shown as blue.

[0045] S1.2: Artificial Intelligence Model Training

[0046] Pre-collect the CT reconstruction data of the container, and then use the color DR images at different angles obtained in S1.1 as the training set. After prediction by an artificial intelligence model (such as an RCNN neural network model), obtain the spatial position of the container. Compare the obtained spatial position of the container with the pre-collected CT reconstruction data of the container, calculate the loss function. If the loss function meets the requirements, then the artificial intelligence model in the current state can be used for container recognition; otherwise, correct the artificial intelligence model and re-recognize until the conditions are met. Finally, output the bounding box of the container recognition result.

[0047] S1.3: Online Automatic Recognition to Obtain Three-dimensional Spatial Position Information

[0048] Perform real-time online detection with CT, perform the S1.1 process on the three-dimensional data to obtain three-dimensional orthographic DR images at different angles, use the artificial intelligence model in S1.2 for recognition to obtain recognition results at different angles, and then perform an intersection operation based on the recognition results at different angles to obtain the bounding box of the three-dimensional recognition of the entire container.

[0049] It should be noted that the above steps S1.1 and S1.2 are to obtain an accurate artificial intelligence model for container recognition to ensure the accuracy during S1.3 online automatic recognition. During the actual security inspection process, steps S1.1 and S1.2 are carried out during the equipment debugging process. When performing real-time online security inspection, directly perform step S1.3.

[0050] 2. Segment the three-dimensional objects in the three-dimensional volume data and obtain the bounding box of each three-dimensional object;

[0051] S2.1: Obtain the projection data of the scanned piece at different angles (for example, from 0 to 360 degrees, with an angle set every 1 degree) through dual-energy CT scanning, reconstruct and invert the projection data to obtain high-energy, low-energy, density, and atomic number data images, obtain multiple two-dimensional tomograms of the three-dimensional volume data, and sequentially number the two-dimensional tomograms.

[0052] Furthermore, after obtaining multiple two-dimensional tomograms, perform image preprocessing on the two-dimensional tomograms to improve the signal-to-noise ratio of the images and make the segmentation results more accurate.

[0053] Specifically, the image preprocessing includes image smoothing, image enhancement, image region segmentation, and dilation and erosion.

[0054] S2.2: Segment the two-dimensional objects in each two-dimensional tomogram, obtain the features of each two-dimensional object in each two-dimensional tomogram, and sequentially number the two-dimensional objects in each two-dimensional tomogram.

[0055] Specifically, the method for segmenting two-dimensional objects in each two-dimensional slice can be a conventional method in the art. For example, the region growing method or the boundary method.

[0056] Exemplarily, the region growing method includes: traversing each two-dimensional slice and selecting seed points in each two-dimensional slice, then traversing the neighborhood of the seed points in the two-dimensional slice. In the neighborhood, points with a difference less than the threshold from the seed points are used as growth points for neighborhood growth, thereby completing the segmentation of two-dimensional objects in each two-dimensional slice.

[0057] Exemplarily, the difference threshold for the seed points of the density image segmentation is 0.1.

[0058] The features of the two-dimensional object include density mean, density variance, atomic number mean, atomic number variance, object area, object perimeter, abscissa of the regional centroid, ordinate of the regional centroid, and minimum circumscribed rectangle.

[0059] S2.3: Analyze and judge the connectivity of the three-dimensional volume:

[0060] S2.3.1: Based on the kth two-dimensional object M in the ith two-dimensional slice ik traverse each two-dimensional object in the (i - 1)th two-dimensional slice, and judge whether the differences in the central position, area, average density, and average atomic number between each two-dimensional object in the (i - 1)th two-dimensional slice and the two-dimensional object M ik meet the threshold requirements; if there is a two-dimensional object in the (i - 1)th two-dimensional slice that meets the threshold requirements, then according to the differences in central position, area, average density, and average atomic number, a comprehensive score is given to the two-dimensional object that meets the threshold requirements. The smaller the difference, the higher the score. The two-dimensional object with the highest comprehensive score is the connected region in the (i - 1)th two-dimensional slice that is connected to the two-dimensional object M ik ; if there is no two-dimensional object in the (i - 1)th two-dimensional slice that meets the threshold requirements, then traverse each two-dimensional object in the (i - j)th two-dimensional slice and repeat the above steps; where i≥2, k≥1, i≥j≥2; j takes values from small to large, and only when there is no two-dimensional object in the currently taken two-dimensional slice that meets the threshold requirements, j will further take a value 1 greater than the current value.

[0061] It should be noted that i≥2, that is, the connectivity of the first two-dimensional slice is not analyzed, and it is directly used as an existing connected region, and the analysis of connectivity starts directly based on the two-dimensional objects in the second two-dimensional slice.

[0062] Exemplarily, the difference thresholds for the central position, area, average density, and average atomic number are 14, 50, 0.1, and 1 respectively.

[0063] Exemplarily, the comprehensive score is calculated according to Equation (I):

[0064]

[0065] where s0 is the area of the two-dimensional object M ik and ρ0 is the average density of the two-dimensional object M ik and Z0 is the average atomic number of the two-dimensional object M ik and x0 and y0 are the abscissa and ordinate of the center position of the two-dimensional object M respectively ik ; s1 is the area of the two-dimensional object that meets the threshold requirements, ρ1 is the average density of the two-dimensional object that meets the threshold requirements, Z1 is the average atomic number of the two-dimensional object that meets the threshold requirements, and x1 and y1 are the abscissa and ordinate of the center position of the two-dimensional object that meets the threshold requirements respectively, and a, b, c, d are parameters greater than 0.

[0066] The values of a, b, c, d are the weighted ratios, which are determined according to the difference thresholds of the corresponding features. For example, a is 14 / 4, b is 50 / 4, c is 0.1 / 4, and d is 1 / 4, indicating that the weights of the four are the same under the above difference thresholds.

[0067] In the present invention, the two-dimensional tomography of the two-dimensional object that does not meet the threshold requirements is defined as the error tomography. In step S2.3.1, the number of layers of the error tomography allowed to be traversed is A. When j is equal to A, if there is no two-dimensional object that meets the threshold requirements in the (i - A)-th two-dimensional tomography, then stop further traversing upward; where A ≥ 1, for example, A is 1, 2, or 3.

[0068] That is to say, in step S2.3.1, if there is a two-dimensional object that meets the threshold requirements in the (i - 1)-th two-dimensional tomography, then perform a comprehensive score on it and determine the connected region; if there is no two-dimensional object that meets the threshold requirements in the (i - 1)-th two-dimensional tomography, then traverse each two-dimensional object in the (i - j)-th two-dimensional tomography; i ≥ j ≥ 2, j takes values from small to large. Only when there is no two-dimensional object that meets the threshold requirements in the two-dimensional tomography with the currently taken value of j, j further takes a value 1 greater than the current value until a two-dimensional object (connected region) that meets the threshold requirements is found or j = A; if there is a two-dimensional object that meets the threshold requirements in the two-dimensional tomography with the currently taken value of j, then end this traversal. If a two-dimensional object that meets the threshold requirements is never found, then end this traversal.

[0069] Furthermore, if there is a two-dimensional object that meets the threshold requirements in the traversed two-dimensional tomography, after determining the two-dimensional object with the highest comprehensive score, the two-dimensional object M ikModify the number to be the same as that of the two-dimensional object with the highest comprehensive score; if no two-dimensional object that meets the threshold requirements is found in the traversed two-dimensional tomograms, add 1 to the largest number of the connectivity region that has completed the connectivity analysis, and use it as the two-dimensional object M ik number.

[0070] It should be noted that each two-dimensional object in the first tomogram can be regarded as a connectivity region that has completed the connectivity analysis.

[0071] S2.3.2: Repeat step S2.3.1 until the connectivity analysis of each two-dimensional object in each two-dimensional tomogram is completed; the two-dimensional objects with connectivity are combined to form three-dimensional objects, and the three-dimensional object segmentation is completed. It includes:

[0072] S2.3.2.1: Based on the (k + 1)-th two-dimensional object M in the i-th two-dimensional tomogram i(k+1) repeat step S2.3.1 until the connectivity analysis of each two-dimensional object in the i-th two-dimensional tomogram is completed;

[0073] S2.3.2.2: Based on the k-th two-dimensional object M in the (i + 1)-th two-dimensional tomogram (i+1)k repeat step S2.3.1 and step S2.3.2.1 until the connectivity analysis of each two-dimensional object in each two-dimensional tomogram is completed.

[0074] To further clearly illustrate the connectivity analysis in steps S2.1 - S2.3, take Figure 3 as an example for illustration.

[0075] S2.1: Through dual-energy CT scanning, obtain 4 two-dimensional tomograms ( Figure 3 only 4 two-dimensional tomograms are shown for the sake of illustration), and number the 4 two-dimensional tomograms ordinally from top to bottom as 1, 2, 3, 4;

[0076] S2.2: Segment the two-dimensional objects in each two-dimensional tomogram, and obtain the characteristics of each two-dimensional object in each two-dimensional tomogram. Number the two-dimensional objects in the first two-dimensional tomogram ordinally as 1, 2, 3, 4, number the two-dimensional objects in the second two-dimensional tomogram ordinally as 1, 2, 3, number the two-dimensional objects in the third two-dimensional tomogram ordinally as 1, 2, and number the two-dimensional objects in the fourth two-dimensional tomogram ordinally as 1, 2, 3;

[0077] S2.3: Analyze and judge the three-dimensional body connectivity:

[0078] S2.3.1: First analyze the connectivity of the second two-dimensional tomogram. Based on the first two-dimensional object M in the second two-dimensional tomogram 21Based on this, traverse each two-dimensional object in the first two-dimensional slice. If the first two-dimensional object in the first two-dimensional slice meets the threshold requirement and has the highest comprehensive score, then the first two-dimensional object in the first two-dimensional slice is the two-dimensional object M 21 's connected region, and change the number of the two-dimensional object M 21 to be the same as the number of the first two-dimensional object in the first two-dimensional slice (both are 1 originally, so no change is needed here); based on the second two-dimensional object M 22 in the second two-dimensional slice, traverse each two-dimensional object in the first two-dimensional slice. If the third two-dimensional object in the first two-dimensional slice meets the threshold requirement and has the highest comprehensive score, then the third two-dimensional object in the first two-dimensional slice is the two-dimensional object M 22 's connected region, and change the number of the two-dimensional object M 22 to be the same as the number of the third two-dimensional object in the first two-dimensional slice, that is, change the number of the two-dimensional object M 22 to 3; based on the third two-dimensional object M 23 in the second two-dimensional slice, traverse each two-dimensional object in the first two-dimensional slice. There is no two-dimensional object in the first two-dimensional slice that meets the threshold requirement, and the maximum number of the two-dimensional objects in the first two-dimensional slice is 4, so change the number of the two-dimensional object M 23 to 5; the connectivity analysis of the second two-dimensional slice is completed.

[0079] Then analyze the connectivity of the third two-dimensional slice. Based on the first two-dimensional object M 31 in the third two-dimensional slice, traverse each two-dimensional object in the second two-dimensional slice. There is no two-dimensional object in the second two-dimensional slice that meets the threshold requirement, so traverse each two-dimensional object in the first two-dimensional slice. The second two-dimensional object in the first two-dimensional slice meets the threshold requirement and has the highest comprehensive score, that is, the second two-dimensional object in the first two-dimensional slice is the two-dimensional object M 31 's connected region, and change the number of the two-dimensional object M 31 to be the same as the number of the second two-dimensional object in the first two-dimensional slice, that is, change the number of the two-dimensional object M 31 to 2; based on the second two-dimensional object M 32 in the third two-dimensional slice, traverse each two-dimensional object in the second two-dimensional slice. The two-dimensional object with number 5 in the second two-dimensional slice meets the threshold requirement and has the highest comprehensive score, that is, the two-dimensional object with number 5 in the second two-dimensional slice is the two-dimensional object M 32 's connected region, and change the number of the two-dimensional object M 32 to 5; the connectivity analysis of the third two-dimensional slice is completed.

[0080] Re-analyze the connectivity of the 4th two-dimensional slice, starting from the 1st two-dimensional object M in the 4th two-dimensional slice 41 as the basis, traverse each two-dimensional object in the 3rd two-dimensional slice. If there is no two-dimensional object in the 3rd two-dimensional slice that meets the threshold requirement, then traverse each two-dimensional object in the 2nd two-dimensional slice. If there is no two-dimensional object in the 2nd two-dimensional slice that meets the threshold requirement, then traverse each two-dimensional object in the 1st two-dimensional slice. If there is no two-dimensional object in the 1st two-dimensional slice that meets the threshold requirement, and it is determined that no two-dimensional object that meets the threshold requirement has been found, and the maximum number of the previously found connected regions is 5, then change the number of the two-dimensional object M 41 to 6; starting from the 2nd two-dimensional object M in the 4th two-dimensional slice 42 as the basis, traverse each two-dimensional object in the 3rd two-dimensional slice. If there is no two-dimensional object in the 3rd two-dimensional slice that meets the threshold requirement, then traverse each two-dimensional object in the 2nd two-dimensional slice. If there is no two-dimensional object in the 2nd two-dimensional slice that meets the threshold requirement, then traverse each two-dimensional object in the 1st two-dimensional slice. If there is no two-dimensional object in the 1st two-dimensional slice that meets the threshold requirement, and it is determined that no two-dimensional object that meets the threshold requirement has been found, and the maximum number of the previously found connected regions is 6, then change the number of the two-dimensional object M 42 to 7; starting from the 3rd two-dimensional object M in the 4th two-dimensional slice 43 as the basis, traverse each two-dimensional object in the 3rd two-dimensional slice. The two-dimensional object with the number 5 in the 3rd two-dimensional slice meets the threshold requirement and has the highest comprehensive score, then change the number of the two-dimensional object M 43 to 5. That is, complete the connectivity analysis of all two-dimensional objects in all two-dimensional slices.

[0081] In the prior art, the segmentation of three-dimensional objects usually adopts the three-dimensional region growing method, the threshold method or the boundary method. These methods either take a long time or the segmentation results of three-dimensional objects are inaccurate.

[0082] The above three-dimensional object segmentation method of the present invention judges the connectivity with each two-dimensional object in the two-dimensional tomogram as the object of connectivity judgment, and finally combines the two-dimensional objects with connectivity to form a three-dimensional object, that is, the present invention is two-dimensional region growing. Compared with the three-dimensional region growing method in the prior art, the present invention reduces the range of region growing required from 26 pixels in three dimensions to 8 pixels in two dimensions. Therefore, the three-dimensional object segmentation method of the present invention is faster, and thus the explosive recognition is faster. Compared with the threshold method for segmenting three-dimensional objects in the prior art, due to the presence of artifact phenomena in CT images, the threshold method cannot accurately segment all objects, while the present invention is two-dimensional region growing, which can accommodate differences and reduce the influence of artifacts. Therefore, the three-dimensional object segmentation method of the present invention has higher accuracy in segmenting three-dimensional objects, and thus the explosive recognition is more accurate. Compared with the boundary method in the prior art, since the boundary method requires first-order or second-order differential calculations and iterative operations, while the present invention does not require any iteration, therefore, the three-dimensional object segmentation method of the present invention is faster, and thus the explosive recognition is faster.

[0083] S2.4 Obtain the features of each three-dimensional object, including spatial position, volume, average density, average atomic number, density mean square deviation, atomic number mean square deviation, etc. The features corresponding to the three-dimensional object are as follows:

[0084]

[0085]

[0086]

[0087]

[0088]

[0089] Among them, S i represents the area corresponding to the i-th pixel of the three-dimensional object, Voxel represents the pixel size corresponding to the pixel, ρ i and Z i respectively represent the density value and atomic number value corresponding to the pixel; V, ρ mean , Z mean , σ ρ , σ Z respectively represent the volume, average density, average atomic number, density mean square deviation, and atomic number mean square deviation corresponding to the three-dimensional object.

[0090] In order to quickly obtain the mean square deviation feature, the density mean square deviation and the atomic number mean square deviation can be abbreviated as σ ρsimple and σ Zsimple :

[0091]

[0092]

[0093] Among them, S k,i , k,i , Z k,i They respectively represent the area, density, and atomic number corresponding to the i-th pixel in the k-th (k=1-K)-th slice.

[0094] S2.5 obtains the three-dimensional data and segments the bounding box of each object.

[0095] The starting and ending positions of each fault of the three-dimensional object are defined as Z1 and Z2 of the bounding box. The minimum rectangular box of each fault is taken as the union, and the result is defined as [X1, Y1, X2, Y2]; the bounding box of the final three-dimensional object is [X1, Y1, Z1, X2, Y2, Z2]. Among them, X1, Y1, Z1, X2, Y2, Z2 are the x, y, z coordinates of the starting point and end point of the minimum circumscribed cube of the three-dimensional object respectively.

[0096] 3. Identify explosives

[0097] S3.1 matches the bounding box of each three-dimensional object obtained in the three-dimensional data segmentation with the bounding box of the container identified by artificial intelligence to determine whether each three-dimensional object is wrapped in a container; if so, the three-dimensional object enters the liquid and powdered explosives identification library for matching; if not, the three-dimensional object enters the solid explosives identification library for matching.

[0098] It should be noted that the term "enclosed" means that the bounding box of the three-dimensional object is smaller than and is close to the bounding box of the container. Since liquid and powdered substances are fluid, they must be contained in a container and in contact with the inner wall of the container. Therefore, when the bounding box of the three-dimensional object is smaller than and is close to the bounding box of the container, it can be determined that the three-dimensional object is liquid or powdered, and it is entered into the liquid and powdered explosives recognition library for matching; otherwise, the three-dimensional object is solid, and it is entered into the solid explosives recognition library for matching.

[0099] It is worth noting that the "solid" of the present invention refers to a block solid that does not have fluidity and cannot completely contact the inner wall of the container in the container. The "container" of the present invention refers to a container that can hold liquid or powder substances, such as a bottle, a can, etc.

[0100] In the present invention, the method of matching explosives can be an autonomous judgment criterion or an artificial intelligence method.

[0101] Exemplarily, the autonomous judgment criteria include: autonomously judging based on the features, if the density deviation is less than a certain number, the atomic number deviation is less than a certain number, etc., then it is determined that the match is successful; otherwise, the match is unsuccessful.

[0102] Exemplarily, the determination using artificial intelligence includes: making a determination in the way of artificial intelligence according to the features, specifically as follows:

[0103] Perform feature classification of explosive data based on the eigenvalue obtained in step S2.3:

[0104] (1) Input: The selected feature variables and the corresponding explosive attributes.

[0105] (2) According to the features and attributes, adopt different classification methods, such as SVM, K-NN, to obtain a classification model.

[0106] (3) Calculate the result. If the result meets the requirements, then this model can be used to identify the material features by the artificial intelligence model in the current state; otherwise, correct the model and perform the identification again until the conditions are met.

[0107] (4) Output: The attributes of the corresponding object.

[0108] This matching method automatically makes a comparison in the way of artificial intelligence, and finally compares each object with the explosives in the database to obtain the matching result of each object.

[0109] S3.2 Result display

[0110] If the matching result is an explosive, then display the position of the explosive in the rendering result, and display information such as the predicted name, weight, volume, density, etc. of the explosive in the information box. If the matching results of the objects in a package are not explosives, then directly output the rendering result.

[0111] Among them, the position of the explosive can be directly output according to the result of S2.4.

[0112] Compared with the prior art, the present invention can at least achieve the following beneficial effects: The present invention uses the artificial intelligence method to identify the container for the existing security inspection CT equipment, and outputs the spatial position where the container is located; uses the fast three-dimensional segmentation technology to output features, and matches the position features of the three-dimensional object with the position of the container identified by artificial intelligence; determines whether each object is wrapped by the container. If so, it is a liquid or powdery object, and performs matching for liquid and powdery explosives; otherwise, it is a solid, and performs matching for solid explosives; this method can quickly identify liquid and powdery explosives and solid explosives separately, improving the accuracy of explosive identification.

[0113] The following further illustrates the explosive identification method of the present invention through specific embodiments.

[0114] Example 1

[0115] S1.1: Based on the density and atomic data reconstructed by CT, perform forward projections at different angles (including 0 degrees, 45 degrees, 90 degrees, 135 degrees, 180 degrees) to obtain projection data of high and low energies, and perform color assignment to obtain a colored DR image.

[0116] Among them, different colors represent different materials. Exemplarily, organic substances such as food and plastic are shown as orange; inorganic substances such as books and ceramics are shown as green; mixtures are also shown as green, and stainless steel products and metals are shown as blue.

[0117] S1.2: Pre-collect the CT reconstruction data of the container, and then use the obtained colored DR images at different angles as the training set. After prediction by an artificial intelligence model (such as an RCNN neural network model), obtain the spatial position of the container; compare the obtained spatial position of the container with the pre-collected CT reconstruction data of the container, calculate the loss function. If the loss function meets the requirements, then the artificial intelligence model in the current state can be used for container recognition; otherwise, correct the artificial intelligence model and re-recognize until the conditions are met. Finally, output the bounding box of the container recognition result.

[0118] S1.3: After the real-time online scanning process of the security inspection CT device, perform S1.1 processing on the three-dimensional data to obtain three-dimensional forward projection DR images at different angles, use the artificial intelligence model of S1.2 for recognition to obtain recognition results at different angles, and then perform an intersection operation based on the recognition results at different angles to obtain the bounding box of the three-dimensional recognition of the entire container.

[0119] S2.1: Obtain projection data of the scanned object at different angles (for example, from 0 to 360 degrees, with an angle set every 1 degree) through dual-energy CT scanning, reconstruct and invert the projection data to obtain high-energy, low-energy, and density and atomic number data images, obtain multiple two-dimensional tomographs of the three-dimensional volume data, and sequentially number the two-dimensional tomographs.

[0120] Furthermore, after obtaining multiple two-dimensional tomographs, perform image preprocessing on the two-dimensional tomographs to improve the signal-to-noise ratio of the images and make the segmentation results more accurate.

[0121] Specifically, the image preprocessing includes image smoothing, image enhancement, image region segmentation, and dilation and erosion.

[0122] S2.2: Segment the two-dimensional objects in each two-dimensional tomograph, obtain the features of each two-dimensional object in each two-dimensional tomograph, and sequentially number the two-dimensional objects in each two-dimensional tomograph.

[0123] Specifically, the method for segmenting the two-dimensional objects in each two-dimensional tomograph can be a conventional method in the art, for example, the region growing method or the boundary method.

[0124] Exemplarily, the region growing method includes: traversing each two-dimensional slice and selecting seed points in each two-dimensional slice, then traversing the neighborhood of the seed points in the two-dimensional slice, and in the neighborhood, points with a difference from the seed points less than a threshold are used as growing points for neighborhood growth, so as to complete the segmentation of the two-dimensional object in each two-dimensional slice.

[0125] Exemplarily, the difference threshold for the seed points of the density image segmentation is 0.1.

[0126] The features of the two-dimensional object include density mean, density variance, atomic number mean, atomic number variance, object area, object perimeter, abscissa of the regional centroid, ordinate of the regional centroid, and minimum circumscribed rectangle.

[0127] S2.3: Analyze and judge the three-dimensional body connectivity:

[0128] S2.3.1: Based on the k-th two-dimensional object M in the i-th two-dimensional slice ik traverse each two-dimensional object in the (i - 1)-th two-dimensional slice, and judge whether the differences in the central position, area, average density, and average atomic number between each two-dimensional object in the (i - 1)-th two-dimensional slice and the two-dimensional object M ik meet the threshold requirements; if there is a two-dimensional object in the (i - 1)-th two-dimensional slice that meets the threshold requirements, then according to the differences in the central position, area, average density, and average atomic number, a comprehensive score is given to the two-dimensional object that meets the threshold requirements. The smaller the difference, the higher the score. The two-dimensional object with the highest comprehensive score is the connected region in the (i - 1)-th two-dimensional slice that is connected to the two-dimensional object M ik ; if there is no two-dimensional object in the (i - 1)-th two-dimensional slice that meets the threshold requirements, then traverse each two-dimensional object in the (i - j)-th two-dimensional slice and repeat the above steps; where, i ≥ 2, k ≥ 1, i ≥ j ≥ 2; j takes values from small to large, and only when there is no two-dimensional object in the currently taken two-dimensional slice that meets the threshold requirements, j further takes a value 1 greater than the current value.

[0129] It should be noted that i ≥ 2, that is, the connectivity of the first two-dimensional slice is not analyzed, and the analysis of connectivity directly starts with the two-dimensional objects in the second two-dimensional slice.

[0130] Exemplarily, the difference thresholds for the central position, area, average density, and average atomic number are 14, 50, 0.1, and 1 respectively.

[0131] Exemplarily, the comprehensive score is calculated according to formula (I):

[0132]

[0133] where, s0 is the two-dimensional object Mik The area, ρ0 is the average density of the two-dimensional object M ik and Z0 is the average atomic number of the two-dimensional object M ik ; x0 and y0 are the abscissa and ordinate of the center position of the two-dimensional object M respectively ik ; s1 is the area of the two-dimensional object that meets the threshold requirement, ρ1 is the average density of the two-dimensional object that meets the threshold requirement, Z1 is the average atomic number of the two-dimensional object that meets the threshold requirement, x1 and y1 are the abscissa and ordinate of the center position of the two-dimensional object that meets the threshold requirement respectively, and a, b, c, d are parameters greater than 0

[0134] The values of a, b, c, d are the weighted ratios, which are determined according to the difference thresholds of the corresponding features. For example, a is 14 / 4, b is 50 / 4, c is 0.1 / 4, d is 1 / 4, indicating that the four have the same weight under the above difference thresholds

[0135] In the present invention, the two-dimensional tomography of the two-dimensional object that does not meet the threshold requirement is defined as an error tomography. In step S2.3.1, the number of layers of the error tomography allowed to be traversed is A. When j is equal to A, if there is still no two-dimensional object that meets the threshold requirement in the (i - A)-th two-dimensional tomography, stop further traversing upward; where A ≥ 1 (for example, A = 3)

[0136] That is to say, in step S2.3.1, if there is a two-dimensional object that meets the threshold requirement in the (i - 1)-th two-dimensional tomography, perform a comprehensive score on it to determine the connected region; if there is no two-dimensional object that meets the threshold requirement in the (i - 1)-th two-dimensional tomography, traverse each two-dimensional object in the (i - j)-th two-dimensional tomography; i ≥ j ≥ 2, j takes values from small to large. Only when there is no two-dimensional object that meets the threshold requirement in the two-dimensional tomography with the current value of j, j further takes a value 1 greater than the current value until a two-dimensional object (connected region) that meets the threshold requirement is found or j = A; if there is a two-dimensional object that meets the threshold requirement in the two-dimensional tomography with the current value of j, stop traversing

[0137] Furthermore, if there is a two-dimensional object that meets the threshold requirement in the traversed two-dimensional tomography, after determining the two-dimensional object with the highest comprehensive score, modify the number of the two-dimensional object M ik to be the same as the number of the two-dimensional object with the highest comprehensive score; if no two-dimensional object that meets the threshold requirement is found in the traversed two-dimensional tomography, add 1 to the largest number of the connected region that has completed the connectivity analysis, and use it as the number of the two-dimensional object M ik

[0138] It should be noted that each two-dimensional object in the first tomography can be regarded as a connected region that has completed the connectivity analysis ​

[0139] S2.3.2: Repeat step S2.3.1 until the connectivity analysis of each two-dimensional object in each two-dimensional slice is completed; the two-dimensional objects with connectivity are combined to form three-dimensional objects, and the three-dimensional object segmentation is completed. It includes:

[0140] S2.3.2.1: Based on the (k + 1)-th two-dimensional object M in the i-th two-dimensional slice, repeat step S2.3.1 until the connectivity analysis of each two-dimensional object in the i-th two-dimensional slice is completed; i(k+1) Based on the (k + 1)-th two-dimensional object M in the i-th two-dimensional slice, repeat step S2.3.1 until the connectivity analysis of each two-dimensional object in the i-th two-dimensional slice is completed;

[0141] S2.3.2.2: Based on the k-th two-dimensional object M in the (i + 1)-th two-dimensional slice, repeat step S2.3.1 and step S2.3.2.1 until the connectivity analysis of each two-dimensional object in each two-dimensional slice is completed. (i+1)k Based on the k-th two-dimensional object M in the (i + 1)-th two-dimensional slice, repeat step S2.3.1 and step S2.3.2.1 until the connectivity analysis of each two-dimensional object in each two-dimensional slice is completed.

[0142] S2.4: Obtain the features of each three-dimensional object, including spatial position, volume, average density, average atomic number, density mean square deviation, atomic number mean square deviation, etc.

[0143] S2.5 Obtain the bounding box for each object segmented from the three-dimensional data: The start position and end position of the slices of each three-dimensional object are defined as Z1 and Z2 of the bounding box, and the union of the minimum rectangular boxes of each slice is taken, and the result is defined as [X1, Y1, X2, Y2]; the bounding box of the final three-dimensional object is [X1, Y1, Z1, X2, Y2, Z2]. Where X1, Y1, Z1, X2, Y2, Z2 are the x, y, z coordinates of the start point and end point of the minimum circumscribed cube of the three-dimensional object respectively.

[0144] S3.1: Match the bounding box of each three-dimensional object obtained from the three-dimensional data segmentation with the bounding box of the container identified by artificial intelligence to determine whether each three-dimensional object is wrapped by a container; if so, the three-dimensional object enters the liquid and powdered explosive recognition library for matching; if not, the three-dimensional object enters the solid explosive recognition library for matching.

[0145] S3.2: Result display

[0146] If the matching result is an explosive, then the position of the explosive is displayed on the rendering result, and information such as the predicted explosive name, weight, volume, density, etc. is displayed in the information box. If the matching results of all objects in a package are not explosives, then the rendering result is directly output.

[0147] Among them, the position of the explosive can be directly output according to the result of S2.5.

[0148] In this embodiment, on a 2.4 GHz computer, three-dimensional object segmentation is performed on a three-dimensional volume data of 512*512*343. The completion time of the three-dimensional object segmentation is 5 s, and 11 types of three-dimensional objects are segmented out.

[0149] Comparative Example 1

[0150] The same scan as in Example 1 is scanned. The difference is that the three-dimensional region growing method in the prior art is used to segment the three-dimensional objects in the three-dimensional volume data of the scan. On a 2.4 GHz computer, for a three-dimensional volume data of 512*512*343, it takes about 50 s for the region growing to complete.

[0151] Comparative Example 2

[0152] The same scan as in Example 1 is scanned. The difference is that the threshold method in the prior art is used to segment the three-dimensional objects in the three-dimensional volume data of the scan. On a 2.4 GHz computer, for a three-dimensional volume data of 512*512*343, the threshold method can only segment out 6 types of three-dimensional objects.

[0153] Comparative Example 3

[0154] The same scan as in Example 1 is scanned. The difference is that the boundary method in the prior art is used to segment the three-dimensional objects in the three-dimensional volume data of the scan. On a 2.4 GHz computer, for a three-dimensional volume data of 512*512*343, the boundary method takes 45 s.

[0155] From the comparison of the results of Example 1 with those of Comparative Examples 1 and 3, it can be seen that compared with the existing three-dimensional region growing method and boundary method, the method of the present invention can complete three-dimensional object segmentation more quickly, and thus can identify explosives more quickly; from the comparison of the results of Example 1 with those of Comparative Example 2, it can be seen that compared with the existing threshold method, the method of the present invention has higher accuracy in segmenting three-dimensional objects, and thus has higher accuracy in identifying explosives.

[0156] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. An explosive identification method, characterized in that, The explosive identification method includes: the artificial intelligence identifies the bounding box of the container; Segment the three-dimensional objects in the three-dimensional volume data and obtain the bounding box of each three-dimensional object; Identify explosives: match the bounding box of each three-dimensional object with the bounding box of the container identified by the artificial intelligence to determine whether the three-dimensional object is wrapped by a container; if so, the three-dimensional object enters the liquid and powdery explosive identification library for matching; if not, the three-dimensional object enters the solid explosive identification library for matching; The artificial intelligence identifies the bounding box of the container, including: performing real-time online detection through CT to obtain three-dimensional orthographic DR images at different angles; using the artificial intelligence model to identify the DR images, and performing intersection operation according to the identification results at different angles to obtain the bounding box of the three-dimensional identification of the container; Before the real-time online detection through CT, artificial intelligence model training is performed, including: (1) Obtain the data identified by the artificial intelligence According to the density and atomic number data reconstructed by CT, perform orthographic projection at different angles to obtain high- and low-energy projection data, and perform color assignment to obtain a colored DR image; (2) Artificial intelligence model training Pre-collect the CT reconstruction data of the container. Using the DR image obtained in step (1) as the training set, use the artificial intelligence model to perform prediction to obtain the spatial position of the container; calculate the loss function. If the loss function meets the requirements, then the artificial intelligence model in the current state can be used for container identification; if the loss function does not meet the requirements, correct the artificial intelligence model and perform identification again until the conditions are met.

2. The explosive identification method according to claim 1, wherein The segmentation of the three-dimensional objects in the three-dimensional volume data and obtaining the bounding box of each three-dimensional object includes: Obtain the projection data of different angles of the scanned piece through dual-energy CT scanning, reconstruct the projection data to obtain multiple two-dimensional tomograms, and sequentially number the two-dimensional tomograms; Segment the two-dimensional objects in each two-dimensional tomogram, obtain the features of each two-dimensional object in each two-dimensional tomogram, and sequentially number the two-dimensional objects in each two-dimensional tomogram; Analyze and judge the three-dimensional body connectivity to obtain the connectivity result, complete the three-dimensional object segmentation, and obtain the features of each three-dimensional object; The start position and end position of the tomogram of each three-dimensional object are defined as Z1 and Z2 of the bounding box; the union of the minimum rectangular boxes of each tomogram is taken, and the obtained result is defined as [X1, Y1, X2, Y2]; the final bounding box of the three-dimensional object is [X1, Y1, Z1, X2, Y2, Z2]; where X1, Y1, Z1, X2, Y2, Z2 are the x, y, z coordinates of the starting point and ending point of the minimum circumscribed cube of the three-dimensional object respectively.

3. The explosive identification method according to claim 2, wherein Preprocess the two-dimensional tomogram image, and the image preprocessing includes image smoothing, image enhancement, image region segmentation, and dilation and erosion.

4. The explosive identification method according to claim 2, characterized in that The features of the two-dimensional object include density mean, density mean square deviation, atomic number mean, atomic number mean square deviation, object area, object perimeter, abscissa of the regional centroid, ordinate of the regional centroid, and minimum circumscribed rectangle.

5. The explosive identification method according to claim 2, wherein, The analysis and judgment of the three-dimensional body connectivity includes: taking the k-th two-dimensional object M in the i-th two-dimensional slice ik as the basis, traversing each two-dimensional object in the (i - 1)-th two-dimensional slice, and judging whether the differences in the central position, area, average density, and average atomic number between each two-dimensional object in the (i - 1)-th two-dimensional slice and the two-dimensional object M ik meet the threshold requirements; if there is a two-dimensional object in the (i - 1)-th two-dimensional slice that meets the threshold requirements, then according to the differences in the central position, area, average density, and average atomic number, a comprehensive score is given to the two-dimensional object that meets the threshold requirements. The smaller the difference, the higher the score. The two-dimensional object with the highest comprehensive score is the connected region in the (i - 1)-th two-dimensional slice that is connected to the two-dimensional object M ik ; if there is no two-dimensional object in the (i - 1)-th two-dimensional slice that meets the threshold requirements, then traverse each two-dimensional object in the (i - j)-th two-dimensional slice and repeat the above steps; where, i≥2, k≥1, i≥j≥2; j takes values from small to large, and only when there is no two-dimensional object in the currently taken two-dimensional slice that meets the threshold requirements, j will further take a value 1 greater than the current value; repeat all the steps of the above connectivity analysis until the connectivity analysis of each two-dimensional object in each two-dimensional slice is completed; the two-dimensional objects with connectivity are combined to form a three-dimensional object, and the three-dimensional object segmentation is completed.

6. The explosive identification method according to claim 1, characterized in that, The method of matching explosives includes the autonomous judgment criterion or the judgment by using the artificial intelligence method.

7. The explosive identification method according to any one of claims 1-6, characterized in that, The explosive identification method further includes result display. If the matching result is an explosive, then the output result shows the location of the explosive and the name, weight, volume, and density of the predicted explosive; if the matching results are not explosives, then the result is directly output.

8. The explosive identification method according to claim 7, wherein The location of the explosive can be determined according to the obtained bounding box of the three-dimensional object.

Images