Vehicle interior information nondestructive detection method based on double projection technology

Abstract

This application relates to the fields of image recognition and image detection technology. Specifically, it relates to a non-destructive testing method for vehicle interior information based on dual-projection technology, comprising the following steps: Step 1: Acquire the horizontal X-ray projection of the vehicle under test to obtain a horizontal projection image; acquire the vertical X-ray projection of the vehicle under test to obtain a vertical projection image; collect the radiation intensity of the horizontal projection image of the vehicle under test to obtain a horizontal radiation map; collect the radiation intensity of the vertical projection image of the vehicle under test to obtain a vertical radiation map; Step 2: Input the horizontal and vertical projection images into an automatic segmentation model respectively to obtain a horizontal image of the cargo area, a horizontal image of the non-cargo area, a vertical image of the cargo area, and a vertical image of the non-cargo area; obtain standard horizontal and vertical projection images of the same model as the vehicle under test from a preset database. This technical solution effectively solves the problem of difficulty in quickly and accurately identifying anomalies when manually checking projection images.

Description

Technical Field

[0001] This application relates to the fields of image recognition and image detection technology, and more specifically, to a non-destructive detection method for vehicle interior information based on dual projection technology. Background Technology

[0002] The content in this section provides only background information related to this application and may not constitute prior art.

[0003] Nuclear power plants house nuclear reactors, and access to and from them must be strictly controlled. This is because it's necessary to prevent the unauthorized removal of nuclear waste and to prevent the unauthorized entry of prohibited items. Therefore, all vehicles entering and leaving nuclear power plants must undergo inspection to prevent the unauthorized entry and exit of contraband.

[0004] Currently, nuclear power plants commonly use radiation imaging technology for vehicle inspections. This technology utilizes the penetration of rays through vehicles and their cargo. Because different materials absorb rays to varying degrees, the signal intensity received by the detector also changes accordingly, thus generating an image reflecting the internal density and structure of the object. By carefully examining these images, contraband can be prevented from entering and leaving the nuclear power plant to the greatest extent possible.

[0005] However, current X-ray image inspections still rely on manual interpretation. To maximize throughput, the inspection time for each image is limited. Some individuals embed contraband into vehicle parts or transported goods, increasing concealment. Amidst the complex image information, inspectors often struggle to visually and quickly identify abnormal areas, leading to inaccurate screening of vehicle interiors and allowing some contraband to illegally enter and exit the nuclear power plant. Summary of the Invention

[0006] The summary section of this application is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description section below. This summary section is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0007] Some embodiments of this application propose a non-destructive testing method for vehicle interior information based on dual projection technology to solve the technical problems mentioned in the background section above.

[0008] As a first aspect of this application, some embodiments of this application provide a non-destructive testing method for vehicle interior information based on dual-projection technology, comprising the following steps:

[0009] Step 1: Obtain the horizontal X-ray projection of the vehicle under test to obtain the horizontal projection image, and obtain the vertical X-ray projection of the vehicle under test to obtain the vertical projection image.

[0010] The radiation intensity of the horizontal projection image of the vehicle under test is collected to obtain the horizontal radiation map;

[0011] The radiation intensity of the vertical projection image of the vehicle under test is collected to obtain the vertical radiation map;

[0012] Step 2: Input the horizontal and vertical projection images into the automatic segmentation model to obtain the horizontal image of the cargo area, the horizontal image of the non-cargo area, the vertical image of the cargo area, and the vertical image of the non-cargo area.

[0013] Step 3: Obtain the standard horizontal projection image and standard vertical projection image of the same model as the vehicle to be tested from the preset database;

[0014] The standard non-cargo area horizontal image and the non-cargo area horizontal image in the standard horizontal projection image are merged into a horizontal image group;

[0015] The standard non-cargo area vertical image and the non-cargo area vertical image in the standard vertical projection image are merged into a vertical image group;

[0016] Step 4: Input the horizontal image of the cargo area into the similarity segmentation model to obtain several horizontal cargo images;

[0017] The vertical images of the cargo area are input into the similarity segmentation model to obtain several vertical cargo images;

[0018] Step 5: Input the horizontal image group into the first comparison model, and generate a horizontal non-cargo area anomaly feature map based on the image differences at the same location in the horizontal image group;

[0019] All horizontal cargo images are input into the second comparison model, and an anomaly feature map of the horizontal cargo area is generated based on the image differences between the horizontal cargo images.

[0020] The vertical image group is input into the first comparison model, and an abnormal feature map of the vertical non-cargo area is generated based on the differences between the images at the same position in the vertical image group.

[0021] All vertical cargo images are input into the second comparison model, and an anomaly feature map of the vertical cargo area is generated based on the image differences between the vertical cargo images;

[0022] Step 6: Combine the horizontal non-cargo area anomaly feature map and the horizontal cargo area anomaly feature map to generate a horizontal anomaly feature map; combine the vertical non-cargo area anomaly feature map and the vertical cargo area anomaly feature map to generate a vertical anomaly feature map.

[0023] Step 7: Fuse the horizontal anomaly feature map with the horizontal radiation map to generate a horizontal high-risk information prompt map; fuse the vertical anomaly feature map with the vertical radiation map to generate a vertical high-risk information prompt map.

[0024] This technical solution effectively solves the problem of difficulty in quickly and accurately identifying anomalies when manually reviewing projected images. By acquiring horizontal and vertical X-ray projection images of the vehicle under test, comprehensive coverage of the vehicle's interior information is achieved. For the non-cargo area, the images of the vehicle under test are compared with images of standard vehicles of the same model to significantly highlight abnormal features in modified areas. For the cargo area, a similarity segmentation model is used to divide it into multiple units and perform internal comparisons to quickly locate cargo areas that differ significantly from other units (i.e., potential modification risks). Although views from two directions are required, the automated screening and anomaly feature extraction of the aforementioned model can accurately identify high-risk areas in the images, significantly reducing the workload of manual review and greatly improving the accuracy of contraband detection.

[0025] Existing vehicle X-ray image segmentation methods struggle to efficiently and accurately extract cargo and non-cargo areas from the complex projections of different vehicle models. Due to the wide variety of vehicle models and their significant structural differences, general segmentation models often require massive amounts of labeled data for training, resulting in limited generalization ability, low segmentation accuracy, high training costs, and inefficiency when processing new vehicle models, failing to meet the needs of rapid safety inspections of vehicles in nuclear power plants.

[0026] Furthermore, step 2 includes the following steps:

[0027] Step 21: Obtain the horizontal and vertical projection images;

[0028] Step 22: Train the automatic segmentation model using images of the pre-annotated edges of the vehicle's cargo compartment;

[0029] Step 23: Input the horizontal projection image into the automatic segmentation model, extract the edges of the carriage, and connect the carriage edges to generate a horizontal image of the cargo area and a horizontal image of the non-cargo area;

[0030] The vertical projection image is input into the automatic segmentation model, the vehicle edges are extracted from it, and the vehicle edges are connected to generate vertical images of the cargo area and the non-cargo area.

[0031] This approach significantly improves the accuracy and efficiency of cargo area extraction by training an automatic segmentation model using the edges of the cargo compartment / vehicle as key features (sample anchor points). This method effectively utilizes the relatively clear and stable edges of vehicle structures in X-ray images, greatly reducing the model's dependence on large amounts of labeled data and specific vehicle models, and enhancing the model's generalization ability across different vehicle types. This not only simplifies the model training process and reduces training difficulty, but also enables rapid and accurate segmentation of cargo and non-cargo areas from horizontal and vertical projection images, laying a reliable foundation for subsequent anomaly detection and thus improving the efficiency and robustness of the overall security inspection process.

[0032] When training an automatic segmentation model to separate cargo and non-cargo areas of a vehicle, traditional loss functions struggle to effectively learn the key discriminative properties of vehicle edge features. Due to the variety of vehicle models and the complexity of the projected image background, the model easily confuses regions that look similar but are of different categories (such as the edge of the cargo compartment and adjacent non-cargo structures), resulting in blurred segmentation boundaries, insufficient accuracy, and difficulty in reliably extracting accurate cargo compartment edges as the basis for segmentation.

[0033] Furthermore, the loss function L used by the automatic segmentation model during training is:

[0034]

[0035] Where a represents the anchor sample, p represents the positive sample, n represents the negative sample, i` represents the index of the triple, and N represents the number of triples. This represents the anchor feature vector of the i'th triple. This represents the positive sample feature vector of the i'th triplet. Let N' represent the negative sample feature vector of the i'th triplet, and N' represent the number of triplets.

[0036] This approach employs a triplet loss function to train the automatic segmentation model, significantly improving its ability to discriminate vehicle edge features and its robustness. This loss function forces the model to bring anchor point samples (cargo compartment edges) closer to positive samples (same type of edges) in the feature space, while simultaneously increasing their distance from negative samples (non-edge or out-of-type edges). This allows the model to more clearly distinguish subtle differences and accurately focus on the key structural features of the cargo compartment edges. This not only greatly improves the accuracy of cargo compartment edge extraction and enhances the model's generalization ability to complex projections of different vehicle models, but also lays a solid foundation for the subsequent reliable generation of images of cargo and non-cargo areas, thereby improving the accuracy and stability of the entire segmentation process.

[0037] Existing methods for detecting anomalies in vehicle non-cargo areas lack accurate and reliable reference standards when comparing images. Because the structures of the non-cargo areas (such as chassis, cab, and frame) differ significantly between different vehicle models, using standard images of generic or incorrect models for comparison can lead to ineffective alignment of the non-cargo area images. This results in numerous "false anomaly" signals caused by inherent structural differences, severely interfering with the identification of genuinely illegally modified areas and reducing the accuracy and reliability of detection.

[0038] Furthermore, step 3 includes the following steps:

[0039] Step 31: Pre-build a vehicle model database, which includes standard horizontal and vertical projection images of several standard vehicles of various models. The standard horizontal and vertical projection images are marked with cargo area dividing lines.

[0040] Step 32: Obtain the model of the vehicle to be tested, and match the standard horizontal projection image and standard vertical projection image of the standard vehicle with the same model as the vehicle to be tested from the vehicle model database;

[0041] Step 33: Merge the standard non-cargo area horizontal image and the non-cargo area horizontal image in the standard horizontal projection image into a horizontal image group; merge the standard non-cargo area vertical image and the non-cargo area vertical image in the standard vertical projection image into a vertical image group.

[0042] This solution pre-constructs a vehicle model database containing projected images of various standard vehicle models (with cargo area segmentation lines marked), and performs precise matching based on the model of the vehicle under test, ensuring that the standard image used for non-cargo area difference comparison is completely consistent with the model of the vehicle under test. This method maximizes the overlap of the inherent structure of the non-cargo area in the image, significantly reducing false alarms (false anomalies) caused by vehicle model differences. By merging the matched standard non-cargo area image with the non-cargo area image of the vehicle under test into an image group, highly comparable and structurally aligned input data is provided for the subsequent difference comparison model (first comparison model), thereby greatly improving the accuracy and reliability of non-cargo area anomaly detection (especially for illegal modifications).

[0043] Traditional segmentation methods face two major challenges when segmenting images of vehicle cargo areas for internal difference comparison: First, the shape, size, and arrangement height of the cargo in the cargo area are irregular and they occlude each other. Direct segmentation based on the original pixels is easily affected by noise, resulting in chaotic boundaries and poor internal consistency of the segmented units (cargo images), making it difficult to reflect the true independent units of the cargo. Second, conventional segmentation algorithms are prone to getting stuck in local optima and cannot optimize the segmentation results from a global perspective, resulting in low similarity between the segmented units and reducing the effectiveness of subsequent internal comparison to detect abnormal cargo.

[0044] Furthermore, similarity segmentation models include:

[0045] Feature point extraction networks are used to extract feature points from input images and transform the input images into feature point networks.

[0046] Image segmentation network, which segments a network of feature points into several images of cargo;

[0047] Among them, the image segmentation network uses the particle swarm optimization algorithm to segment the feature point network so that the similarity of each cargo image is maximized.

[0048] The similarity segmentation model designed in this scheme transforms high-dimensional pixel images into low-dimensional, more discriminative feature space representations through a feature point extraction network, effectively reducing data complexity and noise interference, thus laying the foundation for high-quality segmentation. Furthermore, the image segmentation network employs a particle swarm optimization (PSO) algorithm for global optimization search, aiming to maximize the internal similarity of segmentation units (cargo images) and intelligently partitioning the feature space. This method overcomes local optima, ensuring that each segmented cargo image possesses high internal consistency (i.e., "regularity") and more accurately corresponds to potential independent cargo units. This not only significantly improves the quality and physical meaning of the segmentation results but also provides highly reliable and comparable input data for subsequent steps (the second comparison model) to effectively compare differences within cargo areas and accurately locate abnormal cargo.

[0049] When using deep learning models (such as CNNs) to process high-resolution X-ray images of vehicle cargo areas for segmentation, two major bottlenecks are encountered: First, the model requires massive amounts of labeled data for supervised training, which is costly and difficult to generalize; second, directly processing raw high-dimensional pixel data is computationally complex and inefficient, making it difficult to meet the real-time requirements of security inspection scenarios. Especially for segmentation targets (regular areas of cargo) which are essentially differences in texture and structural patterns rather than fine object recognition, the complex feature extraction process of traditional CNNs appears redundant and inefficient.

[0050] Furthermore, the feature point network is generated as follows:

[0051] S1: Input the image into the feature point extraction network to extract all feature points in the image;

[0052] S2: Divide all feature points into n classes based on a preset similarity interval;

[0053] S3: Replace the descriptors of all feature points with the corresponding categories to generate a feature point image;

[0054] S4: Pre-set a rectangular window to divide the feature point image into several rectangular regions, and use the number of feature points, the location of feature points, and the type of feature points in the rectangular region as the dimensionality reduction operator for that rectangular region;

[0055] S5: Arrange all dimensionality reduction operators according to the rectangular window segmentation method to generate a feature point network.

[0056] The proposed feature point network generation method constructs an efficient, low-dimensional feature point network representation through unsupervised feature point clustering (based on a preset similarity interval) and spatial statistical dimensionality reduction (based on a rectangular window operator). This method first performs preliminary dimensionality reduction and semantic abstraction (category replacement) on feature points through clustering, and then uses rectangular window statistics (feature point quantity, location, and type) to achieve secondary dimensionality reduction and spatial pattern extraction. Although some detailed information is discarded, it highly focuses on key feature points in the image and their relative spatial distribution, and is particularly adept at capturing and representing texture variation patterns formed by the regular stacking of goods within cargo areas. Compared to convolutional networks that rely on a large number of annotations, this method requires no training, significantly improves computational efficiency, and can quickly and effectively segment regular cargo regions with internal consistency, providing highly optimized input for subsequent similarity comparison.

[0057] Traditional image segmentation methods based on texture similarity struggle to achieve globally optimal segmentation results when processing feature point networks in vehicle cargo areas. These methods are often limited to local region optimization, easily getting trapped in local optima. This results in low similarity within the segmented cargo images (units), irregular boundaries, and difficulty in effectively capturing and utilizing the overall spatial patterns generated by the regular stacking of goods (such as periodic arrangement) within the cargo area. This irregularity and low cohesion of the segmentation results severely reduce the accuracy and efficiency of subsequent inter-unit comparisons to identify abnormal cargo.

[0058] The segmentation method of feature point networks includes the following steps:

[0059] Z1: Obtain the feature point network S = {s1, s2, ..., s...} i …、s N}; i represents the index of the dimensionality reduction operator, N represents the total number of dimensionality reduction operators, s i =(x i y i ), (x i y i () represents the coordinates of the dimensionality reduction operator in the feature point network;

[0060] Z2: Generate M feasible solutions R based on the feature point network S. c ;c represents the index of a feasible solution;

[0061] R c ={r c1 r c2 …、r cj …、r cJ |g c}, where j represents the index of the cargo image, r cj Let J represent the center point of the j-th cargo image in the c-th feasible solution, and let J represent the feasible solution R. c The total number of images of cargo in China, rcj ={x cj y cj}, {x cj y cj} represents r cj The coordinates of the center point in the feature point network, g c Let represent the radius of the cargo image for the c-th feasible solution;

[0062] Z3: Set the number of particles H, inertia weight w, first acceleration constant c1, first acceleration constant c2, and maximum number of iterations T;

[0063] For each particle, an initial position R is randomly generated from M feasible solutions. h h represents the particle index;

[0064] R h ={r h1 r h2 …、r hj …、r hJ ;|g h}, r hj Let g represent the center point of the j-th cargo image in the h-th particle. h Let r represent the radius of the cargo image of the g-th particle. hj ={x hj y hj}, {x hj y hj} represents r hj The coordinates of the center point in the feature point network; x hj Randomly generated within the x-coordinate range of the feature point network, y hj Randomly generated within the ordinate range of the feature point network;

[0065] Set initial speed V h V h ={v h1 v h2 …、v hj …、v hJ ;|v h};v hj v represents the velocity of the center point of the j-th cargo image in the h-th particle. h The velocity representing the radius of the cargo image of the g-th particle;

[0066] Set the speed limit V min and the lower limit of speed V max v hj ∈[V min V max ];

[0067] Set the fitness function F(R)h The global optimum (Gbest) and local optimum (pbest) are obtained based on the fitness function. h ;

[0068] Z4: Iterate the particle swarm until the maximum number of iterations is reached, or the change in the global optimal position is less than the threshold.

[0069]

[0070] Among them, V h This represents the current velocity of the h-th particle. pbest represents the new velocity at the h-th particle iteration. h R represents the best position in the history of particle h. h This represents the current position of the h-th particle, where r1 and r2 represent the first and second random numbers, respectively.

[0071] Z5: Continuously iterates over the particles, constantly updating the global best position Gbest and the local best position pbest. h The global best position Gbest is output.

[0072] This scheme employs Particle Swarm Optimization (PSO) for global intelligent segmentation of the feature point network. This method efficiently searches the entire solution space by initializing the particle swarm (representing different segmentation schemes), defining positions (center point + radius of the cargo image) and velocities, setting a fitness function (maximizing intra-cell similarity), and iteratively optimizing. Its core advantage lies in its ability to escape local optima and find the optimal or near-optimal segmentation scheme from a global perspective, ensuring that each segmented cargo image (cell) has extremely high internal similarity and a regular shape (approximately circular). This highly regularized and cohesive segmentation result is particularly beneficial for revealing the spatial structure patterns formed by the periodic and regular stacking of goods in the cargo area. It provides highly consistent and comparable input data for subsequent steps (the second comparison model) to accurately compare inter-cell differences and quickly locate abnormal goods deviating from the pattern, significantly improving the reliability of anomaly detection.

[0073] When using particle swarm optimization to optimize cargo image segmentation, traditional similarity measurement methods (such as cosine similarity) have high computational complexity, making it difficult to meet the real-time segmentation requirements of large-scale feature point networks. These methods typically require calculating the angles or distances between vectors in high-dimensional space, involving numerous multiplication, division, and square root operations. When processing cargo images containing a massive number of dimensionality reduction operators, the computational overhead is enormous, severely limiting the efficiency and practicality of the segmentation algorithm and failing to meet the rapid response requirements of nuclear power plant vehicle security inspections.

[0074] Furthermore, the fitness function F(R)h )for:

[0075] n j Let represent the sum of all dimensionality reduction operators in the j-th cargo image. Let d represent the dimensionality reduction operator in the j-th cargo image, where D represents the total number of dimensionality reduction operators in the cargo image, and d represents the index of the dimensionality reduction operator.

[0076] μ represents the average of the dimensionality reduction operators for all cargo images in the h-th particle.

[0077] This innovative approach employs a fitness function F(Rh) based on the arithmetic sum of dimensionality reduction operators. This function directly calculates the vector sum of all dimensionality reduction operator values ​​within each cargo image (cell), using it as its representation. The standard deviation (σ) of the arithmetic sum of all cells measures the overall cohesion (similarity) of the segmentation scheme. Since each dimensionality reduction operator is highly abstract and incorporates local texture features (feature point descriptors, locations, and types), its values ​​exhibit high discriminative power. This design completely eliminates complex vector similarity calculations (such as cosine similarity), requiring only efficient addition operations, significantly reducing computational complexity by several orders of magnitude. Although some vector space relationship information is sacrificed, this function effectively captures the differences in overall numerical distribution between cells caused by the regular stacking of cargo, quickly evaluating the quality of the segmentation scheme. This greatly improves the optimization efficiency and real-time performance of the particle swarm optimization algorithm, providing a crucial guarantee for achieving high-quality segmentation within a limited time and accurately identifying abnormal cargo.

[0078] When comparing images of non-cargo areas (the vehicle under test and a standard vehicle of the same model) to detect illegal modifications, traditional image texture difference analysis methods (such as pixel-by-pixel comparison or feature map calculation based on deep learning) face efficiency bottlenecks and robustness issues. These methods have high computational complexity, making it difficult to meet the needs of rapid security inspections. They are also sensitive to minor translations, rotations, or changes in lighting between images, easily generating false alarms, and cannot reliably and efficiently capture substantial differences caused by structural modifications (such as welding, drilling, or the addition of components).

[0079] Furthermore, the first comparative model includes:

[0080] Feature point extractor, used to extract SIFT feature points from input images;

[0081] Image overlay tool, used to overlay input images one-to-one;

[0082] An anomaly generator generates horizontal or vertical anomaly feature maps of non-cargo areas based on the overlap and similarity rates of SIFT feature points in overlapping images.

[0083] The first comparison model designed in this scheme obtains scale- and rotation-invariant keypoints and their descriptors through a SIFT feature point extractor, and uses an image overlay device to achieve precise one-to-one spatial alignment between the test image and the standard image. The core innovation lies in the anomaly feature generator, which directly generates anomaly feature maps based on the overlap rate (number of matching points / total number of points) and similarity rate (average similarity of matching point descriptors) of SIFT feature points between the aligned images. This method fully utilizes the robustness of the SIFT algorithm to geometric transformations, enabling precise localization of regions with substantial structural differences (low overlap / similarity rate regions). Compared to texture analysis methods that rely on complex computations, this model only requires efficient SIFT feature matching operations, significantly improving the processing speed and efficiency of anomaly detection in non-cargo areas, while ensuring detection stability under minor vehicle pose changes, providing key technical support for the rapid and reliable identification of illegal modifications.

[0084] When detecting abnormal goods (such as contraband or dangerous goods) within a cargo area, single-dimensional image difference comparison methods have significant limitations: comparing only with the global average (basic difference) easily overlooks relative anomalies between local areas (e.g., a cell doesn't fit with its surrounding cells but the overall value doesn't deviate significantly); while comparing only with adjacent cells (proximity difference) may fail to identify abnormal cells whose overall characteristics deviate from the normal range (e.g., the entire area is contaminated). This single-dimensional detection strategy results in insufficient robustness in anomaly identification, easily leading to missed detections or false alarms, and making it difficult to comprehensively and reliably locate potential dangerous goods of various forms within the cargo area.

[0085] Furthermore, the second comparative model includes:

[0086] The basic anomaly feature calculator is used to calculate the difference between the dimensionality reduction operator and the average dimensionality reduction operator for all input cargo images, generating the basic difference coefficient for each cargo image.

[0087] The neighbor anomaly feature calculator calculates the average difference between the dimensionality reduction operator of each cargo image and the dimensionality reduction operators of neighboring cargo images, generating the neighbor difference coefficient for each cargo image.

[0088] The feature generator adds the base difference coefficient and the neighbor difference coefficient to generate the abnormal feature value for each cargo image; the abnormal feature value of all horizontal cargo images is obtained to obtain the abnormal feature map of the horizontal cargo area.

[0089] Obtain the abnormal feature values ​​of all vertical cargo images to obtain the abnormal feature map of the vertical cargo area.

[0090] The second comparative model designed in this scheme innovatively integrates a basic difference coefficient (difference from the global average dimensionality reduction operator) and a neighbor difference coefficient (average difference from the dimensionality reduction operators of adjacent cargo images) for dual evaluation. This model calculates the absolute deviation of each cargo image using a basic anomaly feature calculator, captures its degree of incongruity with the local context using a neighbor anomaly feature calculator, and finally, a feature generator adds the two to generate a comprehensive anomaly feature value. This dual mechanism complements each other: the basic difference ensures the identification of overall deviations, while the neighbor difference sensitively captures local anomalies. The resulting horizontal / vertical cargo area anomaly feature map can more comprehensively and robustly reveal various types of abnormal cargo (whether local abruptness or overall deviation), significantly reducing missed detections and false alarms, and providing security personnel with a highly reliable basis for cargo area risk location.

[0091] The technical solution of this application embodiment has at least the following advantages and beneficial effects:

[0092] This solution is the first to realize a full-stack solution that combines "vehicle-independent precise segmentation", "structural anomaly detection in non-cargo areas", and "regular deviation anomaly detection in cargo areas". It upgrades vehicle security inspection from a crude mode that relies on human experience to a precise perception system based on the collaboration of physical features and intelligent algorithms, providing a reliable technical barrier for high-risk locations such as nuclear power plants and ports. Attached Figure Description

[0093] Figure 1 This is a flowchart of a non-destructive testing method for vehicle interior information based on dual projection technology.

[0094] Figure 2 This is a horizontal projection image of the vehicle.

[0095] Figure 3 This is a vertical projection image of the vehicle.

[0096] Figure 4 This is a high-risk information alert image.

[0097] Figure 5 This is a vertical high-risk information display diagram. Detailed Implementation

[0098] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments. The same reference numerals in the accompanying drawings represent the same components. It should be noted that the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the described embodiments of this application without creative effort are within the scope of protection of this application.

[0099] Compared to the embodiments shown in the accompanying drawings, feasible embodiments within the scope of this application may have fewer components, other components not shown in the drawings, different components, differently arranged components, or components with different connections, etc. Furthermore, two or more components in the drawings may be implemented in a single component, or a single component shown in the drawings may be implemented as multiple separate components.

[0100] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” and similar terms used in this specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not necessarily indicate a quantity limitation. Terms such as “upper” and “lower” are used only to indicate relative positional relationships, and these relative positional relationships may change accordingly when the absolute position of the described object changes.

[0101] refer to Figure 1 A non-destructive testing method for vehicle interior information based on dual projection technology includes the following steps:

[0102] Step 1: Obtain the horizontal X-ray projection of the vehicle under test to obtain a horizontal projection image; obtain the vertical X-ray projection of the vehicle under test to obtain a vertical projection image; collect the radiation intensity of the horizontal projection image of the vehicle under test to obtain a horizontal radiation map; collect the radiation intensity of the vertical projection image of the vehicle under test to obtain a vertical radiation map.

[0103] Radiation sources are installed above and to the side of the entrances and exits of the nuclear power plant. These sources, when projected onto the vehicle under test, produce corresponding vertical and horizontal projection images. Radiation intensity detectors are then installed at the locations of these radiation sources to detect the radiation intensity of the vehicle under test.

[0104] like Figure 2 and Figure 3 Projected images of the vehicle in the vertical and horizontal directions are given respectively.

[0105] Step 2: Input the horizontal and vertical projection images into the automatic segmentation model to obtain the horizontal image of the cargo area, the horizontal image of the non-cargo area, the vertical image of the cargo area, and the vertical image of the non-cargo area.

[0106] Vehicles are generally divided into cargo and non-cargo areas. Drivers leave the cab, and no cargo is stored there. Thus, all goods are in the cargo area (cargo box). In the non-cargo area, prohibited items are typically integrated into parts, such as by attaching or strapping them to the chassis. In the cargo area, prohibited items are added inside the cargo. For example, a watermelon might be hollowed out and filled with prohibited items to avoid inspection. Therefore, by differentiating the projected images for cargo and non-cargo areas, targeted comparisons can be made, increasing the efficiency of identifying abnormal areas.

[0107] Step 2 includes the following steps:

[0108] Step 21: Obtain the horizontal and vertical projection images;

[0109] Step 22: Train the automatic segmentation model using images of the pre-annotated edges of the vehicle's cargo compartment;

[0110] Step 23: Input the horizontal projection image into the automatic segmentation model, extract the edges of the carriage, and connect the carriage edges to generate a horizontal image of the cargo area and a horizontal image of the non-cargo area;

[0111] The vertical projection image is input into the automatic segmentation model, the vehicle edges are extracted from it, and the vehicle edges are connected to generate vertical images of the cargo area and the non-cargo area.

[0112] Accuracy is crucial when segmenting the cargo and non-cargo areas of a projected image. However, there are many vehicle models with varying internal structures. Creating a separate segmentation model for each carriage would be computationally and labor-intensive. Therefore, this solution extracts the carriage edges, reducing computational load. This is because, while the internal structures of different vehicles may differ, the edge structures of the carriages are generally similar, typically consisting of a frame-like steel box. By extracting the carriage edge structure and segmenting the cargo area, the remaining area can be considered the non-cargo area.

[0113] Automatic segmentation models include:

[0114] The input layer is used to input image information;

[0115] Convolutional layers are used to receive image information from the input layer and perform convolution, normalization, and activation operations on the image information.

[0116] Pooling layers adaptively pool the information processed by convolutional layers. Examples include global pooling or average pooling.

[0117] The fully connected layer connects the output information of the pooling layer;

[0118] The output layer outputs the edge labels of the cargo area in the image information.

[0119] The automatic segmentation model is a common convolutional network model that uses different types of projected images as data items and the annotations on the edges of the carriages in the projected images as labels. The data items and labels are merged into a single sample. The automatic segmentation model is trained using multiple samples, and the final trained automatic segmentation model can accurately segment the projected images.

[0120] The projected images here refer to the vertical and horizontal projected images. Because the edges of the carriage are highly symmetrical, although the image features of the non-cargo areas in the vertical and horizontal projected images are very different, the edge features of the cargo areas are highly similar. Therefore, training a single model is sufficient.

[0121] Furthermore, the loss function L used by the automatic segmentation model during training is:

[0122]

[0123] Where a represents the anchor sample, p represents the positive sample, n represents the negative sample, i` represents the index of the triple, and N represents the number of triples. This represents the anchor feature vector of the i'th triple. This represents the positive sample feature vector of the i'th triplet. Let N' represent the negative sample feature vector of the i'th triplet, and N' represent the number of triplets.

[0124] Among them, the anchor point sample is the edge of the vehicle's cargo area, the positive sample is the enhanced cargo area edge of the vehicle (feature points extracted using the SIFT algorithm), and the negative sample is the non-cargo area edge of the vehicle.

[0125] Step 3: Obtain the standard horizontal projection image and standard vertical projection image of the same model as the vehicle to be tested from the preset database;

[0126] The standard non-cargo area horizontal image and the non-cargo area horizontal image in the standard horizontal projection image are merged into a horizontal image group;

[0127] The standard non-cargo area vertical image and the non-cargo area vertical image in the standard vertical projection image are merged into a vertical image group;

[0128] When comparing vehicles, zone management is required. For non-cargo areas, the vehicle signal needs to be matched to the non-cargo area of ​​the same vehicle. For example, if a truck model is AGX, then standard horizontal and vertical projection images of truck model AGX need to be acquired. When acquiring horizontal and vertical images of the non-cargo areas of the same model of vehicle under test, the images of the same area are merged into one image group. The features in the image group are compared. If abnormal areas, that is, areas with low contrast, are found, it indicates that the vehicle under test has undergone significant human modification in that area, which needs to be investigated further. The specific method is as follows:

[0129] Step 3 includes the following steps:

[0130] Step 31: Pre-build a vehicle model database, which includes standard horizontal and vertical projection images of several standard vehicles of various models. The standard horizontal and vertical projection images are already labeled with cargo area dividing lines. Since the vehicle model database was collected in advance, the crucial cargo area dividing lines can be labeled beforehand.

[0131] Step 32: Obtain the model of the vehicle to be tested, and match the standard horizontal projection image and standard vertical projection image of the standard vehicle with the same model as the vehicle to be tested from the vehicle model database;

[0132] Step 33: Merge the standard non-cargo area horizontal image and the non-cargo area horizontal image in the standard horizontal projection image into a horizontal image group; merge the standard non-cargo area vertical image and the non-cargo area vertical image in the standard vertical projection image into a vertical image group.

[0133] The key to step 3 is to combine the images of the non-cargo areas of the standard vehicle and the vehicle under test into a single image set for subsequent comparison.

[0134] Step 4: Input the horizontal image of the cargo area into the similarity segmentation model to obtain several horizontal cargo images;

[0135] The vertical images of the cargo area are input into the similarity segmentation model to obtain several vertical cargo images;

[0136] For images of cargo areas, the defining characteristic is the stacking of goods. In large-scale transportation, the goods being transported are typically of the same type. Therefore, the goods stacked within the cargo area are very neatly arranged, and when viewed through perspective, a large amount of regular, clearly periodic image information can be observed. Thus, for images of cargo areas, it is only necessary to segment them and then compare the segmentation results.

[0137] However, different goods have different characteristics, making it difficult and time-consuming for manual segmentation of cargo area images. While neural network models can be problematic due to inconsistent features acquired at different times, potentially causing system malfunction when transporting new goods, this application employs a similarity segmentation model. This model segments based on the changing patterns of similarity, requiring no training during segmentation and exhibiting high accuracy.

[0138] Similarity segmentation models include:

[0139] Feature point extraction networks are used to extract feature points from input images and transform the input image into a feature point network.

[0140] The feature points are SIFT feature points. In practice, the feature point extraction network can be either a neural network model or an algorithm. In this scheme, to increase computational efficiency, the feature point extraction network model is a neural network model.

[0141] First, the SIFT algorithm can be used to extract feature points from the image. These feature points are then labeled, and the labeled image is used to train the feature point extraction network. In this way, the feature point extraction network can automatically extract feature points from the input image. SIFT (Scale-Invariant Feature Transform) is a widely used local image feature extraction algorithm in computer vision. The feature points extracted by the SIFT algorithm have scale invariance, so the feature point extraction logic is the same for different types of images. Thus, the trained feature point extraction network can extract feature points from both vertical and horizontal projected images. After extracting all feature points, the input "horizontal image or vertical image of the cargo area" is transformed into a feature point network. The feature point network essentially replaces pixels in the image with feature points. Pixel information is directly filtered out; in practice, dimensionality reduction of the feature points is also necessary.

[0142] The feature point network is generated as follows:

[0143] S1: Input the image into the feature point extraction network to extract all feature points in the image.

[0144] S2: Divide all feature points into n classes based on a preset similarity interval.

[0145] S3: Replace the descriptors of all feature points with the corresponding categories to generate a feature point image;

[0146] For example, in a feature point network with 100 feature points, if a similarity interval of 0.1 is set, these 100 feature points can be directly classified into 10 categories. This simplifies the complex feature information corresponding to these 100 feature points.

[0147] S4: Pre-set a rectangular window to divide the feature point image into several rectangular regions, and use the number of feature points, the location of feature points, and the type of feature points in the rectangular region as the dimensionality reduction operator for that rectangular region.

[0148] The rectangular region here is a relatively small area, similar to a convolutional kernel in a convolutional network, mainly used for dimensionality reduction of images. In this scheme, the rectangular region is 5*5 pixels. That is, the dimensionality reduction operator describes how many feature points exist within a 5*5 pixel grid, where these feature points are located, and the corresponding feature point types.

[0149] S5: Arrange all dimensionality reduction operators according to the rectangular window segmentation method to generate a feature point network.

[0150] Thus, the feature point network can be understood as reducing complex image information to lower-dimensional grid point information through manual convolution. When segmenting with particle swarm optimization, the lower-dimensional grid point information is less likely to cause a computational explosion.

[0151] Image segmentation networks divide a network of feature points into several images of cargo.

[0152] Among them, the image segmentation network uses the particle swarm optimization algorithm to segment the feature point network so that the similarity of each cargo image is maximized.

[0153] Image segmentation networks segment images based on feature point networks, as follows:

[0154] The segmentation method of feature point networks includes the following steps:

[0155] Z1: Obtain the feature point network S = {s1, s2, ..., s...} i …、s N}; i represents the index of the dimensionality reduction operator, N represents the total number of dimensionality reduction operators, s i =(x i y i ), (x i y i () represents the coordinates of the dimensionality reduction operator in the feature point network;

[0156] Z2: Generate M feasible solutions R based on the feature point network S. c ;c represents the index of a feasible solution;

[0157] R c ={r c1 r c2 …、r cj …、r cJ |g c}, where j represents the index of the cargo image, r cj Let J represent the center point of the j-th cargo image in the c-th feasible solution, and let J represent the feasible solution R. c The total number of images of cargo in China, r cj ={x cj y cj}, {x cj y cj} represents r cj The coordinates of the center point in the feature point network, g c Let represent the radius of the cargo image for the c-th feasible solution;

[0158] Z3: Set the number of particles H, inertia weight w, first acceleration constant c1, first acceleration constant c2, and maximum number of iterations T;

[0159] For each particle, an initial position R is randomly generated from M feasible solutions. h h represents the particle index;

[0160] R h ={r h1 r h2 …、r hj …、r hJ ;|g h}, r hj Let g represent the center point of the j-th cargo image in the h-th particle. h Let r represent the radius of the cargo image of the g-th particle. hj ={x hj y hj}, {x hj y hj} represents r hj The coordinates of the center point in the feature point network; x hj Randomly generated within the x-coordinate range of the feature point network, y hj Randomly generated within the ordinate range of the feature point network;

[0161] Set initial speed V h V h ={v h1 v h2 …、v hj …、v hJ ;|v h};v hj v represents the velocity of the center point of the j-th cargo image in the h-th particle. hThe velocity representing the radius of the cargo image of the g-th particle;

[0162] Set the speed limit V min and the lower limit of speed V max v hj ∈[V min V max ];

[0163] Set the fitness function F(R) h The global optimum (Gbest) and local optimum (pbest) are obtained based on the fitness function. h ;

[0164] Fitness function F(R) h )for:

[0165] n j Let represent the sum of all dimensionality reduction operators in the j-th cargo image. Let d represent the dimensionality reduction operator in the j-th cargo image, where D represents the total number of dimensionality reduction operators in the cargo image, and d represents the index of the dimensionality reduction operator.

[0166] μ represents the average of the dimensionality reduction operators for all cargo images in the h-th particle.

[0167] Z4: Iterate the particle swarm until the maximum number of iterations is reached, or the change in the global optimal position is less than the threshold.

[0168]

[0169] Among them, V h This represents the current velocity of the h-th particle. pbest represents the new velocity at the h-th particle iteration. h R represents the best position in the history of particle h. h This represents the current position of the h-th particle, where r1 and r2 represent the first and second random numbers, respectively.

[0170] Z5: Continuously iterates over the particles, constantly updating the global best position Gbest and the local best position pbest. h The global best position Gbest is output.

[0171] The output result is: V h ={v h1 v h2 …、v hj …、v hJ ;|v h According to v h1 v h2…、v hj …、v hJ Find the center point of J cargo images, based on v h Using the dimensions of the cargo image as a reference, we can reconstruct the cargo image (horizontal cargo image, vertical cargo image) from the center point by using the dimensions.

[0172] Step 5: Input the horizontal image group into the first comparison model, and generate a horizontal non-cargo area anomaly feature map based on the image differences at the same location in the horizontal image group;

[0173] All horizontal cargo images are input into the second comparison model, and an anomaly feature map of the horizontal cargo area is generated based on the image differences between the horizontal cargo images.

[0174] The vertical image group is input into the first comparison model, and an abnormal feature map of the vertical non-cargo area is generated based on the differences between the images at the same position in the vertical image group.

[0175] All vertical cargo images are input into the second comparison model, and an anomaly feature map of the vertical cargo area is generated based on the image differences between the vertical cargo images;

[0176] The first comparative model includes:

[0177] Feature point extractor, used to extract SIFT feature points from input images;

[0178] The feature point extractor can directly use the preceding feature point extraction network to extract SIFT feature points. The input image here refers to either a vertical image group or a horizontal image group.

[0179] An image overlay tool is used to overlap input images one-to-one. For example, if the input is a vertical image group, it will overlap two images in the vertical image group.

[0180] An anomaly generator generates horizontal or vertical anomaly feature maps of non-cargo areas based on the overlap and similarity rates of SIFT feature points in overlapping images.

[0181] Because the two images are relatively close, the positions of their SIFT feature points are also close. The overlap and similarity rates of adjacent SIFT feature points can be used as the rendering values ​​in the anomaly feature map. A high overlap rate and high similarity rate result in a smaller outlier. The anomaly feature map can be a horizontal or vertical non-cargo area anomaly feature map.

[0182] The first comparison model can obtain the anomaly feature map of the non-cargo area, while the second comparison model is used to obtain the anomaly feature map of the cargo area.

[0183] The second comparative model includes:

[0184] The basic anomaly feature calculator is used to calculate the difference between the dimensionality reduction operator and the average dimensionality reduction operator for all input cargo images, generating the basic difference coefficient for each cargo image.

[0185] The basic difference coefficient is the similarity between the dimensionality reduction operator of the cargo image and the average dimensionality reduction operator. The higher the similarity, the lower the basic difference coefficient.

[0186] The neighbor anomaly feature calculator calculates the average difference between the dimensionality reduction operator of each cargo image and the dimensionality reduction operators of neighboring cargo images, generating the neighbor difference coefficient for each cargo image.

[0187] The nearest neighbor difference coefficient is the average difference (average similarity) between the dimensionality reduction operator of the cargo image and the dimensionality reduction operators of adjacent cargo images. The higher the similarity, the lower the nearest neighbor difference coefficient.

[0188] The feature generator adds the base difference coefficient and the neighbor difference coefficient to generate the abnormal feature value for each cargo image; it obtains the abnormal feature value of all horizontal cargo images to obtain the horizontal cargo area abnormal feature map; and it obtains the abnormal feature value of all vertical cargo images to obtain the vertical cargo area abnormal feature map.

[0189] Step 6: Combine the horizontal non-cargo area anomaly feature map and the horizontal cargo area anomaly feature map to generate a horizontal anomaly feature map; combine the vertical non-cargo area anomaly feature map and the vertical cargo area anomaly feature map to generate a vertical anomaly feature map.

[0190] The stitching here mainly involves re-stitching the non-cargo area images and the cargo area images according to their correspondence and segmentation relationships. The vertical anomaly feature map shows the distribution of anomaly features in the vertical projection image. The horizontal anomaly feature map shows the distribution of anomaly features in the horizontal projection image.

[0191] Step 7: Fuse the horizontal anomaly feature map with the horizontal radiation map to generate a horizontal high-risk information alert map; fuse the vertical anomaly feature map with the vertical radiation map to generate a vertical high-risk information alert map. For example... Figure 4 and Figure 5 As shown, Figure 4 and Figure 5 The rectangular area is the high-risk anomaly area that appears after the horizontal anomaly feature map and the horizontal radiation map are merged.

[0192] The fusion process here primarily involves multiplying the radiation values ​​from the horizontal radiation map as weighting coefficients with the anomaly feature map. If a location exhibits an abnormal response and also has a high radiation level, that area is marked as a high-risk region.

[0193] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A non-destructive testing method for vehicle interior information based on dual-projection technology, characterized in that, include: Step 1: Obtain the horizontal X-ray projection of the vehicle under test to obtain the horizontal projection image, and obtain the vertical X-ray projection of the vehicle under test to obtain the vertical projection image. The radiation intensity of the horizontal projection image of the vehicle under test is collected to obtain the horizontal radiation map; The radiation intensity of the vertical projection image of the vehicle under test is collected to obtain the vertical radiation map; Step 2: Input the horizontal and vertical projection images into the automatic segmentation model to obtain the horizontal image of the cargo area, the horizontal image of the non-cargo area, the vertical image of the cargo area, and the vertical image of the non-cargo area. Step 3: Obtain the standard horizontal projection image and standard vertical projection image of the same model as the vehicle to be tested from the preset database; The standard non-cargo area horizontal image and the non-cargo area horizontal image in the standard horizontal projection image are merged into a horizontal image group; The standard non-cargo area vertical image and the non-cargo area vertical image in the standard vertical projection image are merged into a vertical image group; Step 4: Input the horizontal image of the cargo area into the similarity segmentation model to obtain several horizontal cargo images; The vertical images of the cargo area are input into the similarity segmentation model to obtain several vertical cargo images; Step 5: Input the horizontal image group into the first comparison model, and generate a horizontal non-cargo area anomaly feature map based on the image differences at the same location in the horizontal image group; All horizontal cargo images are input into the second comparison model, and an anomaly feature map of the horizontal cargo area is generated based on the image differences between the horizontal cargo images. The vertical image group is input into the first comparison model, and an abnormal feature map of the vertical non-cargo area is generated based on the differences between the images at the same position in the vertical image group. All vertical cargo images are input into the second comparison model, and an anomaly feature map of the vertical cargo area is generated based on the image differences between the vertical cargo images; Step 6: Combine the horizontal non-cargo area anomaly feature map and the horizontal cargo area anomaly feature map to generate a horizontal anomaly feature map; combine the vertical non-cargo area anomaly feature map and the vertical cargo area anomaly feature map to generate a vertical anomaly feature map. Step 7: Fuse the horizontal anomaly feature map with the horizontal radiation map to generate a horizontal high-risk information prompt map; The vertical anomaly feature map is fused with the vertical radiation map to generate a vertical high-risk information prompt map; Similarity segmentation models include: Feature point extraction networks are used to extract feature points from input images and transform the input images into feature point networks. Image segmentation network, which segments the feature point network into several cargo images; Among them, the image segmentation network is based on the particle swarm optimization algorithm to segment the feature point network so that the similarity of each cargo image is maximized; The feature point network is generated as follows: S1: Input the image into the feature point extraction network to extract all feature points in the image; S2: Divide all feature points into n classes based on a preset similarity interval; S3: Replace the descriptors of all feature points with the corresponding categories to generate a feature point image; S4: Pre-set a rectangular window to divide the feature point image into several rectangular regions, and use the number of feature points, the location of feature points, and the type of feature points in the rectangular region as the dimensionality reduction operator for that rectangular region; S5: Arrange all dimensionality reduction operators according to the rectangular window segmentation method to generate a feature point network.

2. The non-destructive testing method for vehicle interior information based on dual projection technology according to claim 1, characterized in that, Step 2 includes the following steps: Step 21: Obtain the horizontal and vertical projection images; Step 22: Train the automatic segmentation model using images of the pre-annotated edges of the vehicle's cargo compartment; Step 23: Input the horizontal projection image into the automatic segmentation model, extract the edges of the carriage, and connect the carriage edges to generate a horizontal image of the cargo area and a horizontal image of the non-cargo area; The vertical projection image is input into the automatic segmentation model, the vehicle edges are extracted from it, and the vehicle edges are connected to generate vertical images of the cargo area and the non-cargo area.

3. The non-destructive testing method for vehicle interior information based on dual projection technology according to claim 2, characterized in that, The loss function used during training of the automatic segmentation model for: ； Where a represents the anchor sample, p represents the positive sample, n represents the negative sample, and i` represents the index of the triple. This represents the anchor feature vector of the i'th triple. This represents the positive sample feature vector of the i'th triplet. Let N' represent the negative sample feature vector of the i'th triplet, and N' represent the number of triplets.

4. The non-destructive testing method for vehicle interior information based on dual projection technology according to claim 1, characterized in that, Step 3 includes the following steps: Step 31: Pre-build a vehicle model database, which includes standard horizontal and vertical projection images of several standard vehicles of various models. The standard horizontal and vertical projection images are marked with cargo area dividing lines. Step 32: Obtain the model of the vehicle to be tested, and match the standard horizontal projection image and standard vertical projection image of the standard vehicle with the same model as the vehicle to be tested from the vehicle model database; Step 33: Merge the standard non-cargo area horizontal image and the non-cargo area horizontal image in the standard horizontal projection image into a horizontal image group; merge the standard non-cargo area vertical image and the non-cargo area vertical image in the standard vertical projection image into a vertical image group.

5. The non-destructive testing method for vehicle interior information based on dual projection technology according to claim 1, characterized in that, The segmentation method of feature point networks includes the following steps: Z1: Obtain the feature point network S = {s1, s2, ..., s...} i ...、s N }; i represents the index of the dimensionality reduction operator, N represents the total number of dimensionality reduction operators, s i = (x i y i (x) i y i () represents the coordinates of the dimensionality reduction operator in the feature point network; Z2: Generate M feasible solutions R based on the feature point network S. c ;c represents the index of a feasible solution; R c ={r c1 r c2 ...、r cj ...、r cJ |g c }, where j represents the index of the cargo image, r cj Let J represent the center point of the j-th cargo image in the c-th feasible solution, and let J represent the feasible solution R. c The total number of images of cargo in China, r cj ={x cj y cj }, {x cj y cj } represents r cj The coordinates of the center point in the feature point network, g c Let represent the radius of the cargo image for the c-th feasible solution; Z3: Set the number of particles H, inertia weight w, first acceleration constant c1, first acceleration constant c2, and maximum number of iterations T; For each particle, an initial position R is randomly generated from M feasible solutions. h h represents the particle index; R h ={r h1 r h2 ...、r hj ...、r hJ ;|g h }, r hj Let g represent the center point of the j-th cargo image in the h-th particle. h Let r represent the radius of the cargo image of the g-th particle. hj ={x hj y hj }, {x hj y hj } represents r hj The coordinates of the center point in the feature point network; x hj Randomly generated within the x-coordinate range of the feature point network, y hj Randomly generated within the ordinate range of the feature point network; Set initial speed V h V h ={v h1 v h2 ...、v hj ...、v hJ ;|v h };v hj v represents the velocity of the center point of the j-th cargo image in the h-th particle. h The velocity representing the radius of the cargo image of the g-th particle; Set the speed limit V min and the lower limit of speed V max v hj ∈[V min V max ]; Set the fitness function F(R) h The global optimum (Gbest) and local optimum are obtained based on the fitness function. ; Z4: Iterate the particle swarm until the maximum number of iterations is reached, or the change in the global optimal position is less than the threshold. ； Among them, V h This represents the current velocity of the h-th particle. This represents the new velocity at the h-th particle iteration. R represents the best position in the history of particle h. h This represents the current position of the h-th particle, where r1 and r2 represent the first and second random numbers, respectively. Z5: Continuously iterates over the particles, constantly updating the global optimum (Gbest) and the local optimum. The global best position Gbest is output.

6. The non-destructive testing method for vehicle interior information based on dual projection technology according to claim 5, characterized in that, Furthermore, the fitness function F(R) h )for: ; ;n j Let represent the sum of all dimensionality reduction operators in the j-th cargo image. Let d represent the dimensionality reduction operator in the j-th cargo image, where D represents the total number of dimensionality reduction operators in the cargo image, and d represents the index of the dimensionality reduction operator. ; Let represent the average of the dimensionality reduction operators for all cargo images in the h-th particle.

7. The non-destructive testing method for vehicle interior information based on dual projection technology according to claim 1, characterized in that, The first comparative model includes: Feature point extractor, used to extract SIFT feature points from input images; Image overlay tool, used to overlay input images one-to-one; An anomaly generator generates horizontal or vertical anomaly feature maps of non-cargo areas based on the overlap and similarity rates of SIFT feature points in overlapping images.

8. The non-destructive testing method for vehicle interior information based on dual projection technology according to claim 1, characterized in that, The second comparative model includes: The basic anomaly feature calculator is used to calculate the difference between the dimensionality reduction operator and the average dimensionality reduction operator for all input cargo images, generating the basic difference coefficient for each cargo image. The neighbor anomaly feature calculator calculates the average difference between the dimensionality reduction operator of each cargo image and the dimensionality reduction operators of neighboring cargo images, generating the neighbor difference coefficient for each cargo image. The feature generator adds the base difference coefficient and the neighbor difference coefficient to generate the abnormal feature value for each cargo image; the abnormal feature value of all horizontal cargo images is obtained to obtain the abnormal feature map of the horizontal cargo area. Obtain the abnormal feature values ​​of all vertical cargo images to obtain the abnormal feature map of the vertical cargo area.

Images