A method for intelligent detection of contraband images based on a dual-projection radiometric imaging system

Abstract

This invention belongs to the field of contraband detection technology, specifically, it relates to an intelligent image detection method for contraband based on a dual-projection radiometric imaging system. This method generates dual-projection image pairs containing contraband through a projection simulation algorithm, then constructs a training dataset, and uses the algorithm to train side-view and bottom-view image datasets to obtain contraband detection models for the side-view and bottom-view images. During the side-view process, based on the distribution characteristics of different types of contraband in the dual-projection images, different weights are assigned to the detection results of the side-view and bottom-view images. Combining positional distribution and size information, the three-dimensional spatial distribution and probability of presence of contraband in the vehicle are estimated, thus optimizing the overall detection process, effectively improving detection speed and accuracy, and increasing the detection rate and recognition rate.

Description

Technical Field

[0001] This invention belongs to the field of contraband detection technology, specifically, it relates to an intelligent image detection method for contraband based on a dual-projection radiometric imaging system. Background Technology

[0002] A system designed for dual-projection large vehicle detection typically employs two sets of radiation sources and detectors, utilizing radiation imaging technology to scan and inspect large vehicles. However, due to the complex background structure of large vehicles, the scanned images contain a large amount of information, while contraband occupies a very small proportion of this background, and image stacking is prone to occur. Manual image review often makes it difficult to quickly and accurately determine the entire image. Therefore, intelligent image detection technology is needed for detection.

[0003] The main feature of the dual-projection large vehicle security inspection system is that it can utilize the correlation between the projected images from two directions to effectively unfold the stacked parts when encountering the problem of overlapping items in a single projection, thereby improving the detection capability. However, existing intelligent image detection algorithms mainly rely on a single image, and it is difficult for them to autonomously combine the correlation between two projections from different directions for detection. Therefore, it is necessary to design different detection methods, utilize the characteristics of the dual-projection images themselves and the actual security inspection situation, optimize the detection process, improve detection performance, increase the detection rate and recognition rate, and reduce the missed detection rate and false detection rate. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, the present invention aims to provide a method for intelligent detection of contraband images based on a dual-projection radiometric imaging system, which can achieve associated detection of two projected images from different angles, improve the performance of intelligent image detection networks, and obtain a higher detection rate.

[0005] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0006] The intelligent detection method for contraband images based on a dual-projection radiation imaging system described in this invention employs a dual-projection device, which includes a gantry. A side-mounted radiation source and two segmented polygonal detectors are located on the side of the gantry, while a bottom-mounted radiation source and three segmented polygonal detectors are located at the bottom of the gantry. The side-mounted and bottom-mounted radiation sources are discretized into spatial coordinates, and the rays are the lines connecting the side-mounted radiation source to the two segmented polygonal detectors, and the bottom-mounted radiation source to the three segmented polygonal detectors. The projection of the object's cross-section through the gantry into the linear array is obtained. By transforming the relative positions of the side-mounted radiation source, the bottom-mounted radiation source, and the two and three segmented polygonal detectors with the object, side-view and bottom-view image pairs of the object are obtained.

[0007] The method includes the following steps:

[0008] S1: Obtain the three-dimensional volume data of the target contraband. V ;

[0009] S2: Simulate the scanning process of the dual-projection system based on the system parameters of the dual-projection device, and process the three-dimensional volume data obtained in step S1. V Simulated scanning was performed to obtain side and bottom view images of the target contraband, i.e., simulated projection data. P ;

[0010] S3: Overlay the background image index of the actual vehicle to obtain a pair of scanned images containing the target contraband, and generate a labeling file based on the position of the target contraband in the scanned image pair;

[0011] S4: Change the type, position, and rotation angle of the target contraband, and repeat steps S2-S3 to obtain labeled double-projection image pairs to form a training dataset;

[0012] S5: Train the training dataset formed in step S4 using a neural network to generate two sets of detection networks for both upward and side views. Design a detection algorithm specifically for the dual-projection system. The detection algorithm includes:

[0013] The detection weights for the top and side views are defined based on the type of the target prohibited item. ;

[0014] Based on the size correlation and positional distribution information between the two projection images (side view and top view), the probability of the presence of the target contraband in different areas is defined. ;

[0015] right and The region with the highest value is analyzed to locate the target region.

[0016] S6: Replace the security inspection image to be inspected, and repeat steps S4-S5 to achieve intelligent detection of contraband security inspection images.

[0017] Further, in step S1, the three-dimensional volume data of the target contraband are analyzed based on the linear attenuation coefficient of the material under the system's radiation energy. V Perform the assignment.

[0018] Furthermore, in step S2, the Siddon algorithm is used to obtain the projection of the object's cross-section onto the linear array, thereby obtaining the simulated projection data of the object. P .

[0019] Furthermore, in step S4, according to the category of the target contraband, the coordinate positions of the target contraband obtained by superimposing the scanned image pairs in step S3 are stored as txt tag files, and the scanned image pairs are paired with the tag files to form a dual-projection image pair containing multiple categories of target contraband, thus obtaining a training dataset.

[0020] Furthermore, in step S4, the scanned image pair can also be processed by random cropping, filling, scaling, or pasting.

[0021] Further, in step S5, one of the images, either from a low angle or a side view, is selected as the main detection image based on the type of the target contraband, and the detection weights of the low angle and side view are defined. ,when This indicates that all detections are performed using a bottom view. This indicates that all detections are performed using the side view.

[0022] Furthermore, in step S5, the probability of the target contraband being present in different areas is preset, and probability coefficients are defined. , The larger the value, the higher the likelihood that the target contraband is present in that area.

[0023] Furthermore, in step S5, the actual placement location of the target contraband inside the vehicle is preset, and a probability coefficient is defined. , The larger the value, the higher the likelihood that the target contraband is present in that area.

[0024] Furthermore, in step S5, the probability of the target contraband being present in different areas is preset, and probability coefficients are defined. , The larger the value, the higher the likelihood that the target contraband is present in that area;

[0025] Probability of existence , The larger the value, the higher the likelihood that the target contraband is present in that area.

[0026] Furthermore, the presence range of the target contraband is located in the stepping direction of the object, and the presence range is the intersection between the cone bundle formed by the side view and the cone bundle formed by the front view.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] (1) The present invention can effectively improve detection efficiency, and achieve rapid positioning of the target area by defining and optimizing the parameters during the detection process. It can also simultaneously improve detection accuracy and achieve more accurate detection of the target area.

[0029] (2) This invention makes full use of the advantages of the dual projection system. By jointly detecting the projected images in two directions, it maximizes the advantages of the dual projection system, simplifies situations that are difficult to handle with traditional case methods, improves the detection rate, solves the problem of smuggling contraband to the greatest extent, and ensures the safe and stable operation of society.

[0030] (3) This invention can be widely applied to various places and working conditions that require security checks. It improves security check efficiency and detection accuracy through intelligent monitoring, and, with the help of manual means, maximizes the safety of the target area. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the steps of the present invention;

[0032] Figure 2 This is a schematic diagram of the data of a prohibited substance in this invention;

[0033] Figure 3 This is a schematic diagram of the structure of the dual-projection device in this invention;

[0034] Figure 4 This is a type of simulated projection data in the present invention. P A schematic diagram;

[0035] Figure 5 This is a schematic diagram illustrating the structural information of a vehicle according to the present invention;

[0036] Figure 6 This is a schematic diagram of the three-dimensional positioning principle of the dual-projection detection system in this invention;

[0037] Figure 7 This is a schematic diagram of a training dataset containing images and label files in this invention;

[0038] Figure 8 These are images of a portion of the dataset used in this invention;

[0039] In the diagram: 1-Gantry; 2-Side radioactive source; 3-Two-segment zigzag detector; 4-Bottom radioactive source; 5-Three-segment zigzag detector. Detailed Implementation

[0040] The present invention will be further described below with reference to specific embodiments.

[0041] The intelligent detection method for contraband images based on a dual-projection radiometric imaging system described in this invention first employs a specialized dual-projection device, such as... Figure 1 As shown, the specific method is as follows.

[0042] First, after the detection process of the dual-projection vehicle inspection system is completed, two projected images of the object, one from the side and one from below, are obtained. These two images are then used for joint intelligent detection of prohibited items.

[0043] S1: Based on the main application scenarios of the dual-projection vehicle security inspection system, acquire three-dimensional volumetric data of common prohibited items such as explosives, police equipment, flammable and explosive contraband, and simulated weapons. V The volumetric data can be obtained through industrial CT or voxelization of vector design drawings, and the three-dimensional volumetric data can be determined based on the linear attenuation coefficient of the target contraband material under the system's radiation energy. V Perform the assignment, such as Figure 2 As shown, a schematic diagram of prohibited substance data is presented.

[0044] S2: Based on the system geometry and system parameters of the dual projection equipment, process the 3D volume data obtained in step S1. V Simulated scanning was performed to obtain side-view and upward-view projection images of the contraband, thus simulating projection data. P Dual projection devices such as Figure 3 As shown, it includes a gantry 1. The side of the gantry 1 is equipped with a side-mounted radiation source 2 and two segments of polygonal detectors 3. The bottom of the gantry 1 is equipped with a bottom-mounted radiation source 4 and three segments of polygonal detectors 5. The side-mounted radiation source 2 and the bottom-mounted radiation source 4 are discretized into spatial coordinates. The rays are the lines connecting the side-mounted radiation source 2 to the two segments of polygonal detectors 3, and the bottom-mounted radiation source 4 to the three segments of polygonal detectors 5. The Siddon algorithm is used to obtain the projection of the object's cross-section through the gantry 1 onto the linear array. By transforming the relative positions of the side-mounted radiation source 2, the bottom-mounted radiation source 4, the two segments of polygonal detectors 3, the three segments of polygonal detectors 5, and the object being measured, the side and bottom view image pairs of the object are obtained, which are the simulated projection data. P ,like Figure 4 As shown, a simulation projection data is presented. P A schematic diagram.

[0045] S3: Repeat steps S1 and S2 to obtain a large number of upward / side-view projection image pairs of different contraband. Then, exponentially overlay the contraband projection image pairs with the vehicle background to obtain scan image pairs with the target contraband. Store the overlaid coordinate positions and contraband categories as JSON annotation files in COCO format. Pair the scan image pairs with the label JSON files to create a dataset of side-view and upward-view contraband images.

[0046] S4: Change the type and location of the contraband, and use methods such as rotation angle, random cropping, filling, scaling, and pasting to perform data augmentation, obtain a large number of labeled dual-projection scan image pairs, form a training dataset, solve the problems of class imbalance and target sparsity, and select an appropriate detector based on the aspect ratio of the contraband in the dataset.

[0047] S5: Train the training dataset formed in step S4 using a neural network to generate two sets of detection networks for both upward and side views. Design a detection algorithm specifically for the dual-projection system. The detection algorithm includes:

[0048] The detection weights for the top and side views are defined based on the type of the target prohibited item. This value represents the weighting of the top and side views during the detection process. These weights are manually set based on different projected image pairs. For example, if the object being detected is a handgun, it has a clear feature outline in the top view, while in the side view it's essentially a line. Therefore, the weighting of the top view for this contraband can be increased, while the weighting of the side view can be decreased. We can also use some general rules, such as: objects with more symmetrical planes have a higher weighting for the side view, and smaller objects have a higher weighting for the top view. This indicates that all inspections are performed using a bottom view. This indicates that all detection is performed using the side view. By properly configuring parameter values ​​and adjusting the detection weight coefficients of the two projected images, detection accuracy and detection capability can be improved.

[0049] Because the positional information in the bottom and side views is highly correlated—for example, even if the contraband is placed arbitrarily in space, its coordinates in the direction of movement must be strictly consistent—once a contraband is detected in one projection image, the contraband in the other projection image will inevitably appear on a specific strip. The probability of the contraband's presence in different areas is defined based on the size correlation and positional distribution information between the side and bottom projection images. ;

[0050] right and The region with the maximum value is analyzed, and image information extraction methods such as histograms are used to perform joint detection with the gray-level histogram of the target to achieve rapid localization of the target region;

[0051] Among them, the probability coefficient is defined as the probability of the presence of the target contraband in different areas. , The larger the value, the higher the likelihood that the target contraband is present in that area;

[0052] Besides the information in the image itself, the actual placement of objects within the corresponding vehicle is also a factor limiting the area of ​​focus in the projected image. For example, for large freight vehicles, the probability of the contraband being found in the cargo compartment is higher, while for small cars, the probability is higher in the driver's cab. Based on the structural information of the actual vehicle, the probability of the contraband being present in different areas is preset, and probability coefficients are defined. , The larger the value, the higher the likelihood that the target contraband is present in that area. Figure 5 As shown, a schematic diagram of vehicle structure information is given. In this schematic diagram, the probability of dangerous goods being present in the side view area is relatively high, while in the bottom view, the probability of dangerous goods being present in the tire area is relatively high.

[0053] By combining the probability information of each region in the image, the total probability coefficient of the region is calculated, and the probability of existence is determined. , The larger the value, the higher the likelihood of the presence of the target contraband in that area, thus improving detection efficiency.

[0054] S6: Replace the security inspection image to be inspected, and repeat steps S4-S5 to achieve intelligent detection of contraband security inspection images.

[0055] Based on the back-projection cone beam of the detected results, the area where the contraband is located in the stepping direction can be determined. Additionally, auxiliary methods such as manual inspection can be used to achieve three-dimensional localization of the contraband. Figure 6 As shown in the figure, a schematic diagram of the three-dimensional positioning principle of the dual-projection detection system is given. As can be seen from the figure, the contraband detection area A in the side view and the side source form a cone beam, and the contraband detection area B in the bottom view and the bottom source form a cone beam. The intersection of the two cone beams is the contraband positioning area C.

[0056] In step S4 above, as Figure 7 As shown, a schematic diagram of a training dataset containing images and label files is provided in this embodiment. This training dataset contains images of several different categories of hazardous materials, with a fixed pixel size of 320×320. The hazardous materials are randomly distributed in different parts of different images according to their three-dimensional spatial location, such as... Figure 8 The image shown is a partial dataset. The training dataset also contains information on the location, size, and category of several hazardous materials in the images, which is the label file information. The label file name is the same as the file name of the corresponding image.

Claims

1. A method for intelligent detection of contraband images based on a dual-projection radiometric imaging system, characterized in that, This method employs a dual-projection device, which includes a gantry. A side-mounted radiation source and two segmented polygonal detectors are located on the side of the gantry, while a bottom-mounted radiation source and three segmented polygonal detectors are located at the bottom of the gantry. The side-mounted and bottom-mounted radiation sources are discretized into spatial coordinates, and the rays are the lines connecting the side-mounted radiation source to the two segmented polygonal detectors, and the bottom-mounted radiation source to the three segmented polygonal detectors. The projection of the object's cross-section through the gantry into the linear array is obtained. By transforming the relative positions of the side-mounted radiation source, the bottom-mounted radiation source, and the two and three segmented polygonal detectors with the object, the side and bottom view images of the object are obtained. The method includes the following steps: S1: Obtain the three-dimensional volume data of the target contraband. V ; S2: Simulate the scanning process of the dual-projection system based on the system parameters of the dual-projection device, and process the three-dimensional volume data obtained in step S1. V Simulated scanning was performed to obtain side and bottom view images of the target contraband, i.e., simulated projection data. P ; S3: Overlay the background image index of the actual vehicle to obtain a pair of scanned images containing the target contraband, and generate a labeling file based on the position of the target contraband in the scanned image pair; S4: Change the type, position, and rotation angle of the target contraband, and repeat steps S2-S3 to obtain labeled double-projection image pairs to form a training dataset; S5: Train the training dataset formed in step S4 using a neural network to generate two sets of detection networks for both upward and side views. Design a detection algorithm specifically for the dual-projection system. The detection algorithm includes: The detection weights for the top and side views are defined based on the type of the target prohibited item. ; Based on the size correlation and positional distribution information between the two projection images (side view and top view), the probability of the presence of the target contraband in different areas is defined. ; right and The region with the highest value is analyzed to locate the target region. S6: Replace the security inspection image to be inspected, and repeat steps S4-S5 to achieve intelligent detection of contraband security inspection images.

2. The intelligent detection method for contraband images based on a dual-projection radiometric imaging system according to claim 1, characterized in that, In step S1, the three-dimensional volume data of the target contraband are analyzed based on the linear attenuation coefficient of the material under the system's radiation energy. V Perform the assignment.

3. The intelligent detection method for contraband images based on a dual-projection radiometric imaging system according to claim 1, characterized in that, In step S2, the Siddon algorithm is used to obtain the projection of the object's cross-section onto the linear array, thereby obtaining the simulated projection data of the object. P .

4. The intelligent detection method for contraband images based on a dual-projection radiometric imaging system according to claim 1, characterized in that, In step S4, based on the category of the target contraband, the coordinate positions of the target contraband obtained by superimposing the scanned image pairs in step S3 are stored as txt tag files, and the scanned image pairs are paired with the tag files to form a dual-projection image pair containing multiple categories of target contraband, thus obtaining the training dataset.

5. The intelligent detection method for contraband images based on a dual-projection radiometric imaging system according to claim 1, characterized in that, In step S4, the scanned image pairs can also be processed by random cropping, filling, scaling, or pasting.

6. The intelligent detection method for contraband images based on a dual-projection radiometric imaging system according to claim 1, characterized in that, In step S5, based on the type of the target contraband, one of the images, either viewed from below or from the side, is selected as the main detection image, and the detection weights for the views from below and from the side are defined. ,when This indicates that all detections are performed using a bottom view. This indicates that all detections are performed using the side view.

7. The intelligent detection method for contraband images based on a dual-projection radiometric imaging system according to claim 1, characterized in that, In step S5, the probability of the target contraband being present in different areas is preset, and probability coefficients are defined. , The larger the value, the higher the likelihood that the target contraband is present in that area.

8. The intelligent detection method for contraband images based on a dual-projection radiometric imaging system according to claim 1, characterized in that, In step S5, the actual location of the target contraband inside the vehicle is preset, and a probability coefficient is defined. , The larger the value, the higher the likelihood that the target contraband is present in that area.

9. The intelligent detection method for contraband images based on a dual-projection radiometric imaging system according to claim 8, characterized in that, In step S5, the probability of the target contraband being present in different areas is preset, and probability coefficients are defined. , The larger the value, the higher the likelihood that the target contraband is present in that area; Probability of existence , The larger the value, the higher the likelihood that the target contraband is present in that area.

10. The intelligent detection method for contraband images based on a dual-projection radiometric imaging system according to claim 1, characterized in that, The presence range of the target contraband is located in the direction of the object's movement. The presence range is the intersection between the cone bundle formed by the side view and the cone bundle formed by the front view.

Images