Artificial Intelligence-Based Automatic Identification Method and System for Dangerous Items in Baggage Security Inspection

Abstract

The present invention relates to an automatic identification method and system for dangerous goods in luggage security inspection based on artificial intelligence, which is applied to the field of intelligent security inspection technology, and includes: obtaining a suspicious image from the X-ray image of the luggage to be identified; screening the suspicious image to obtain the image area to be investigated; since the selected image area to be investigated excludes the image area that does not contain the content of the suspicious image, it avoids the waste of computing power and time caused by identifying all image areas; extracting the suspicious image object in the image area to be investigated, calculating the similarity between the suspicious image object and the preset comparison image object one by one, so as to obtain the matching comparison image object, and extracting the corresponding comparison suspicious image content, and verifying the suspicious image content through the comparison suspicious image content, which can speed up the determination of the dangerous goods identification result, improve the security inspection efficiency, and thus realize automatic security inspection, reducing the interference of human factors such as human experience and fatigue on the detection and identification result.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent security inspection, and particularly to an automatic identification method and system for dangerous goods in luggage security inspection based on artificial intelligence. Background Art

[0002] In the field of security inspection, X-ray image recognition technology has always been a key security inspection means. With the increasing popularity of public transportation and air travel, it has become crucial to conduct fast and accurate security inspections on luggage.

[0003] However, traditional X-ray image recognition methods often rely on manual inspection. This method is not only inefficient but also easily affected by human factors such as the experience and fatigue of inspectors, resulting in misinspection or missed inspection. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide an automatic identification method and system for dangerous goods in luggage security inspection based on artificial intelligence, so as to solve the problem in the prior art that traditional X-ray image recognition methods often rely on manual inspection, which is not only inefficient but also easily affected by human factors such as the experience and fatigue of inspectors, resulting in misinspection or missed inspection.

[0005] According to the first aspect of the embodiments of the present invention, an automatic identification method for dangerous goods in luggage security inspection based on artificial intelligence is provided. The method includes:

[0006] Obtain the X-ray image of the luggage to be identified, extract one or more image regions from the X-ray image, and perform suspicious image content recognition on each of the one or more image regions to obtain one or more suspicious images;

[0007] Obtain the confidence levels of the suspicious image objects included in each of the one or more suspicious images respectively, and filter the one or more suspicious images based on a preset confidence level threshold to obtain one or more image regions to be investigated;

[0008] Extract the suspicious image objects in each of the one or more image regions to be investigated respectively, and calculate the similarity between the suspicious image objects and the comparison image objects of each dangerous goods image in a pre-built dangerous goods image library one by one;

[0009] Obtain the comparison image object that matches the suspicious image object in the pre-built dangerous goods image library according to the similarity calculation result, extract the comparison suspicious image content corresponding to the suspicious image object in the matching comparison image object, and verify the image region to be investigated where the suspicious image object is extracted through the comparison suspicious image content to obtain the verification result of the image region to be investigated;

[0010] Determine the hazardous item recognition result of the X-ray image of the luggage to be recognized according to the verification results of the one or more image areas to be checked.

[0011] Preferably,

[0012] Performing suspicious image content recognition on the one or more image areas respectively to obtain one or more suspicious images, including:

[0013] Performing rough classification on the one or more image areas according to the edge contours of the one or more image areas;

[0014] Obtain the target detection algorithms corresponding to the one or more image areas according to the rough classification results of the one or more image areas;

[0015] Perform suspicious image content detection on the image area through the target detection algorithm. If the target content is detected, the image area where the target content is detected is used as a suspicious image to obtain one or more suspicious images.

[0016] Preferably,

[0017] The performing suspicious image content detection on the image area through the target detection algorithm includes:

[0018] Obtain the feature type of the image area;

[0019] If the feature type of the image area is a color feature, perform color feature comparison on the image area through the corresponding target detection algorithm to determine the target color;

[0020] Perform supplementary feature detection on the image area through the target detection algorithm, and fuse the target color and the supplementary feature based on the distribution conditions of the target color and the supplementary feature in the image area respectively, and determine whether it is the target content based on the fusion result.

[0021] Preferably,

[0022] The respectively extracting suspicious image objects in the one or more image areas to be checked includes:

[0023] Determine the specific recognized item types of the one or more image areas to be checked;

[0024] Perform target detection on the one or more image areas to be checked respectively under the specific recognized item types;

[0025] Extract the texture features of the target object based on the target detection results;

[0026] Fuse the object shapes and corresponding texture features extracted from the one or more image regions to be investigated to obtain one or more suspicious image objects.

[0027] Preferably,

[0028] The performing target detection for the one or more image regions to be investigated under a specific identification object category respectively includes:

[0029] Obtain a pre - constructed target detection transfer network for a specific identification object category; the target detection transfer network is obtained by performing transfer learning based on the overall target detection network for the specific identification object category; the overall target detection network is debugged based on image examples carrying prior markings of object shapes under the specific identification object category;

[0030] Input the one or more image regions to be investigated into the pre - constructed target detection transfer network respectively, and perform target detection on the image regions to be investigated through the target detection transfer network;

[0031] The target detection transfer network is obtained by performing transfer learning based on the overall target detection network for the specific identification object category;

[0032] The overall target detection network is debugged based on image examples carrying prior markings of object shapes under the specific identification object category;

[0033] The extracting the texture features of the target object based on the target detection results includes:

[0034] Obtain a pre - constructed representative texture image set for the specific identification object category, and the representative texture image set is a set of typical texture samples under the specific identification object category;

[0035] Perform an image block splitting operation on the image regions to be investigated respectively to obtain an image block set;

[0036] Calculate the similarity between each image block in the image block set and each representative texture image in the representative texture image set;

[0037] Select the image blocks with a similarity greater than a preset first similarity threshold to the representative texture images as representative image blocks;

[0038] Extract the texture of the representative image blocks as the texture features of the target object.

[0039] Preferably,

[0040] The calculating the similarity between each image block in the image block set and each representative texture image in the representative texture image set respectively includes:

[0041] Save the representative texture images belonging to the same inspection attribute type in the representative texture image set in the same storage space, and each storage space corresponds to a different texture attribute type;

[0042] Obtain a pre-built image recognition network; the image recognition network is debugged according to an image sample set;

[0043] Classify the image area to be inspected through the image recognition network to obtain an image classification result;

[0044] Determine the storage space corresponding to the texture attribute type to which the corresponding representative texture image belongs according to the image classification result;

[0045] Calculate the similarity between each image block in the image block set and each representative texture image saved in the determined storage space.

[0046] Preferably,

[0047] The calculation of the similarity between the suspicious image object and each comparison image object of each dangerous goods image in the pre-built dangerous goods image library one by one includes:

[0048] Traverse each dangerous goods image in the dangerous goods image library, and calculate the object shape similarity between the object shape included in the obtained suspicious image object and the object shape included in each comparison image object of each dangerous goods image respectively to obtain an object shape similarity calculation result;

[0049] If the object shape similarity calculation result is greater than a preset second similarity threshold, add the dangerous goods image to the temporary set of dangerous goods images until the entire dangerous goods image library is traversed;

[0050] Obtain the part of the comparison image object of each dangerous goods image in the temporary set of dangerous goods images that matches the shape of the suspicious image object to obtain the comparison inspection content of the comparison image object of each dangerous goods image;

[0051] Calculate the similarity between the object texture included in the suspicious image object and the texture features of the comparison inspection content in the comparison image object of each dangerous goods image in the temporary set of dangerous goods images.

[0052] Preferably,

[0053] The traversal of each dangerous goods image in the dangerous goods image library, and the calculation of the object shape similarity between the object shape included in the obtained suspicious image object and the object shape included in each comparison image object of each dangerous goods image respectively to obtain an object shape similarity calculation result, includes:

[0054] By grouping the semantics of the object shapes included in each dangerous item image in the dangerous item image library, obtaining the grouping of each dangerous item image group, and acquiring the centroid of each dangerous item image group;

[0055] Obtaining the spatial distances between the object shapes included in the suspicious image object and the centroids of each group, and sorting the spatial distances to determine the order of traversing each dangerous item image group;

[0056] According to the traversal order of each dangerous item image group, respectively obtaining the similarity between the object shape included in the comparison image object of the dangerous item image in each dangerous item image group and the object shape included in the suspicious image object;

[0057] If, within a certain dangerous item image group, the similarity between the object shape included in the suspicious image object and the object shape included in the comparison image object of any dangerous item image in this group is greater than a preset second similarity threshold, then after traversing this dangerous item image group, do not continue to traverse the next dangerous item image group.

[0058] Preferably, it further includes:

[0059] If there is no comparison image object in the dangerous item image library that matches the suspicious image object, then obtaining the expert dangerous item recognition result obtained by an expert's inspection of the image area to be investigated;

[0060] Integrating the suspicious image object and the suspicious image content in the image area to be investigated according to the expert dangerous item recognition result to obtain a dangerous item image;

[0061] Adding the integrated dangerous item image to the dangerous item image library for iterative update of the dangerous item image library;

[0062] The acquisition of the dangerous item images in the dangerous item image library includes:

[0063] Extracting a plurality of comparison image areas from the comparison images in a pre-built comparison image library;

[0064] Respectively extracting the suspicious image content and the suspicious image object of each comparison image area;

[0065] Regarding the suspicious image content and the suspicious image object in the same comparison image area as candidate suspicious merge data, and determining the occurrence density corresponding to each candidate suspicious merge data;

[0066] Taking the candidate suspicious merge data with an occurrence density greater than a preset density threshold as the dangerous item image;

[0067] The step of separately identifying suspicious image content in the one or more image regions to obtain one or more suspicious images further includes:

[0068] Based on the comparison suspicious image content in the dangerous goods image where the compared image object is matched, annotate the suspicious image content in the image region to obtain an annotated image;

[0069] Display the annotated image and give an alarm.

[0070] According to the second aspect of the embodiments of the present invention, there is provided an automatic identification system for dangerous goods in baggage security inspection based on artificial intelligence. The system includes a memory and a processor;

[0071] The memory stores a computer program that can run on the processor, and when the processor executes the program, it implements the steps in the method described in any one of the above.

[0072] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0073] In this application, by obtaining the X-ray image of the baggage to be identified, extracting the image region from the X-ray image, identifying the suspicious image content in the image region to obtain a suspicious image; screening the suspicious image to obtain the image region to be checked; since the selected image region to be checked contains suspicious image content, it can exclude the image regions that do not contain suspicious image content, preventing the waste of computing power and time caused by identifying the image regions that do not contain suspicious image content; extracting the suspicious image object in the image region to be checked, calculating the similarity between the suspicious image object and the comparison image objects of each dangerous goods image in the pre-built dangerous goods image library one by one, so as to obtain the matched comparison image object, and extracting the corresponding comparison suspicious image content, verifying the image region to be checked through the comparison suspicious image content; determining the comparison suspicious image content through the matching mechanism with the comparison image objects in the dangerous goods image library, and verifying the suspicious image content according to the comparison suspicious image content, which can speed up the determination of the dangerous goods identification result, improve the security inspection efficiency, and thus realize automatic security inspection, reducing the interference of human factors such as human experience and fatigue on the detection and identification results.

[0074] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] The accompanying drawings herein are incorporated into the specification and constitute a part of this specification, showing embodiments consistent with the present invention and used together with the specification to explain the principles of the present invention.

[0076] Figure 1It is a schematic diagram of the overall process of an artificial intelligence-based automatic dangerous item recognition method for luggage security inspection shown according to an exemplary embodiment;

[0077] Figure 2 It is a schematic diagram of the process of suspicious image content recognition shown according to another exemplary embodiment;

[0078] Figure 3 It is a schematic diagram of the target detection process when the suspicious image content is color feature data shown according to another exemplary embodiment;

[0079] Figure 4 It is a schematic diagram of the process of extracting suspicious image objects from the image areas to be investigated shown according to another exemplary embodiment;

[0080] Figure 5 It is a schematic diagram of the target detection process under a specific type of recognized items shown according to another exemplary embodiment;

[0081] Figure 6 It is a schematic diagram of the object texture extraction process under a specific type of recognized items shown according to another exemplary embodiment;

[0082] Figure 7 It is a schematic diagram of the process of calculating the similarity between an image block and each representative texture image in a representative texture image set shown according to another exemplary embodiment;

[0083] Figure 8 It is a schematic diagram of the process of calculating the similarity one by one between a suspicious image object and each comparison image object in a pre-established dangerous item image library shown according to another exemplary embodiment;

[0084] Figure 9 It is a schematic diagram of the object shape similarity calculation process shown according to another exemplary embodiment;

[0085] Figure 10 It is a schematic diagram of the device of an artificial intelligence-based automatic dangerous item recognition system for luggage security inspection shown according to another exemplary embodiment; Detailed implementation manners

[0086] Here, the exemplary embodiments will be described in detail, and the examples are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementation manners described in the following exemplary embodiments do not represent all implementation manners consistent with the present invention. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0087] Embodiment 1

[0088] Figure 1 It is a schematic diagram of the overall process of an AI-based automatic identification method for dangerous items in luggage security inspection shown according to an exemplary embodiment. As Figure 1 shown, the method includes:

[0089] S1. Obtain the X-ray image of the luggage to be identified, extract one or more image regions from the X-ray image, and respectively perform suspicious image content recognition on the one or more image regions to obtain one or more suspicious images;

[0090] It can be understood that in practical applications, when a passenger's luggage passes through the X-ray machine, the security inspection system will capture and generate the X-ray image of the luggage in real time. These images contain the shapes and information of various items inside the luggage. To more accurately identify the contents of the luggage, the security inspection system needs to further process and analyze these images. First, the security inspection system uses image processing techniques, such as edge detection, threshold segmentation, etc., to extract one or more specific image regions from the overall X-ray image; these regions may be divided according to the characteristics of the item, such as shape, density, color, etc. For example, if a metal knife is placed in a luggage, the knife will appear as a highlighted region in the X-ray image because the absorption ability of its metal material to X-rays is different from that of other items. After extracting these specific image regions, the security inspection system uses suspicious image content recognition technology to perform preliminary analysis and recognition on these regions. In this process, machine learning models, such as convolutional neural networks (CNNs), can be used. This type of network is particularly suitable for processing image data; through training, the CNN can learn the method of extracting useful features from images and perform classification and recognition based on these features; taking the convolutional neural network as an example, the security inspection system can input the extracted image regions into the trained CNN model. The model automatically extracts features in the image, such as edges, textures, etc., and compares them with the previously learned features. If the model determines that a certain image region contains suspicious content, such as a potential dangerous item, then this region will be marked as suspicious and enter the next step of detailed investigation and analysis.

[0091] S2. Respectively obtain the confidence levels of the suspicious image objects included in the one or more suspicious images, and screen the one or more suspicious images based on a preset confidence level threshold to obtain one or more image regions to be investigated;

[0092] It can be understood that in practical applications, when a passenger's luggage passes through X-ray inspection, the image processing module of the security inspection system will first capture and generate an X-ray image of the interior of the luggage. Subsequently, the security inspection system divides these images into regions and separately identifies suspicious contents in each region. After the identification is completed, the security inspection system checks the identification results corresponding to each image region. These identification results are usually generated by a machine learning model, such as a convolutional neural network (CNN), based on features in the image (such as shape, texture, density, etc.). The model will output the probability or confidence level that each region contains suspicious contents. For example, the security inspection system identifies three image regions A, B, and C and gives their respective confidence levels of containing suspicious contents: 0.8 for region A, 0.3 for region B, and 0.9 for region C. The security inspection system will select the regions that need to be further investigated from these three regions according to a preset confidence threshold (such as 0.7). In this example, since the confidence levels of regions A and C are higher than the threshold, they will be selected as the image regions to be investigated for containing suspicious image contents, while the confidence level of region B is lower than the threshold, so it may not be selected as an image region to be investigated. Or, in a relatively simple implementation, the security inspection system directly deletes the image regions that do not contain suspicious image contents to obtain the image regions to be investigated for containing suspicious image contents. Next, the security inspection system continues to conduct a more detailed analysis and inspection on the selected image regions to be investigated through subsequent steps to determine whether there is a real security threat.

[0093] S3. Separately extract the suspicious image objects in the one or more image regions to be investigated, and calculate the similarity between each of the suspicious image objects and the comparison image objects of each dangerous item image in a pre-built dangerous item image library one by one;

[0094] It is understandable that the security inspection system conducts more in-depth processing and comparison on the suspected image areas screened in the previous steps. First, the security inspection system performs image segmentation and object extraction on these suspected image areas to determine specific suspicious image objects. For example, if a suspected image area is suspected of hiding a knife, the security inspection system uses image processing techniques such as edge detection and threshold segmentation to accurately extract the shape and contour of the knife, forming a suspicious image object. After extracting the suspicious image object, the security inspection system compares it with the comparison image objects in the pre-established dangerous goods image library. This dangerous goods image library contains image data of various known dangerous goods, such as knives, guns, flammable items, etc. Each dangerous good has a corresponding comparison image object, and these comparison image objects are usually carefully processed and labeled to accurately reflect the characteristics of the dangerous goods. During the comparison process, the security inspection system uses image processing and machine learning techniques to evaluate the similarity between the suspicious image object and the comparison image objects. For example, feature matching algorithms such as SIFT (Scale-Invariant Feature Transform) and SURF (Speeded-Up Robust Features) can be used to extract and compare key points and feature descriptors in the images. These algorithms can help the security inspection system accurately identify dangerous goods in a complex image background. If the suspicious image object highly matches a certain comparison image object in the dangerous goods image library, then the security inspection system will issue an alarm indicating the discovery of a potential dangerous good. At the same time, the security inspection system can also provide detailed comparison results and image evidence for security inspectors to further review and confirm. For example, assume that the security inspection system finds a suspicious knife-shaped object in a passenger's luggage. The security inspection system will compare this object with the knife image in the dangerous goods image library. If the comparison result shows a high similarity, the security inspection system will determine that the object is a knife and trigger an alarm. Security inspectors can then conduct a manual inspection and confirmation of this object to ensure safety. By accurately extracting and comparing suspicious image objects, it helps the security inspection system more accurately and efficiently identify potential dangerous goods, thus improving the security and reliability of the security inspection.

[0095] S4. Obtain, according to the similarity calculation result, a comparison image object that matches the suspicious image object in the pre-established dangerous goods image library, extract the comparison suspicious image content corresponding to the suspicious image object in the obtained matching comparison image object, and verify the suspected image area where the suspicious image object is extracted through the comparison suspicious image content to obtain the verification result of the suspected image area.

[0096] It is understandable that the security inspection system first locates the dangerous goods image where the matching comparison image object is located, and extracts the content of the comparison suspicious image from it. These contents usually refer to the specific item image or characteristic part corresponding to the suspicious image object in the dangerous goods image. Next, the security inspection system uses these contents of the comparison suspicious image to verify the area of the image to be checked of the extracted suspicious image object. The purpose of the verification is to confirm whether the content of the suspicious image matches the characteristics in the known dangerous goods image, so as to further determine the authenticity and danger of the content of the suspicious image. The verification process may involve various image processing and analysis techniques, such as feature point matching, contour analysis, or recognition by deep learning models. Taking the deep learning model as an example, the security inspection system can use a trained convolutional neural network (CNN) to identify the content of the suspicious image and compare it with the content of the comparison suspicious image. If the verification result shows that the content of the suspicious image is highly consistent with the content of the comparison suspicious image, the security inspection system can determine that the content of the suspicious image indeed represents a dangerous good and take corresponding security measures accordingly. For example, in the airport security inspection scenario, if the security inspection system finds through comparison that the items in a certain piece of luggage highly match the explosive image in the dangerous goods image library, then the security inspection system further verifies the content of this suspicious image. If the verification result shows that this item indeed has the characteristics of an explosive, the security inspection system will trigger an alarm and notify the security personnel to take immediate countermeasures. In this way, it is ensured that the security inspection system can, on the basis of the initial match, improve the accuracy and reliability of identifying dangerous goods through further verification, thus ensuring public safety.

[0097] S5. Determine the dangerous goods recognition result of the X-ray image of the luggage to be recognized according to the verification result of the one or more areas of the image to be checked.

[0098] It is understandable that the security inspection system determines the identification result of dangerous items in the X-ray image of the luggage to be identified based on the verification results obtained in the previous steps. Specifically, the security inspection system comprehensively considers the verification results of all the image areas to be screened. If the verification result of a certain image area to be screened shows that its content highly matches the items in the dangerous item image library, then the security inspection system will mark it as a potential dangerous area. For example, in the airport security inspection scenario, if the security inspection system detects multiple suspicious areas in the X-ray image of the luggage and verifies these areas in the above steps, then the security inspection system synthesizes these verification results to determine whether there are dangerous items in the luggage. If the verification result of a certain area strongly indicates that it contains dangerous items such as explosives or knives, the security inspection system generates an identification result of dangerous items, triggers the corresponding alarm, and notifies the security inspection personnel to conduct further inspection on the luggage. In addition, the security inspection system can also adopt some additional technical means to improve the accuracy of identification. For example, the security inspection system may use image segmentation technology to more accurately locate the position of dangerous items, or use multi-modal identification methods to combine X-ray images and other sensor data (such as weight, density, etc.) for comprehensive judgment. This step relies on the accurate verification and comparison in the previous steps to ensure that the security inspection system can accurately and timely identify dangerous items, thereby effectively maintaining public safety.

[0099] It is understandable that in this application, by obtaining the X-ray image of the luggage to be identified, extracting the image areas from the X-ray image, performing suspicious image content identification on the image areas to obtain suspicious images, screening the suspicious images to obtain the image areas to be screened. Since the selected image areas to be screened contain suspicious image content, it can exclude the image areas that do not contain suspicious image content, preventing the waste of computing power and time caused by identifying the image areas that do not contain suspicious image content. Extract the suspicious image objects in the image areas to be screened, calculate the similarity between each suspicious image object and the comparison image objects of each dangerous item image in the pre-built dangerous item image library one by one, so as to obtain the matching comparison image objects, extract the corresponding comparison suspicious image content, and verify the image areas to be screened through the comparison suspicious image content. Determine the comparison suspicious image content through the matching mechanism with the comparison image objects in the dangerous item image library, and verify the suspicious image content based on the comparison suspicious image content, which can speed up the determination of the dangerous item identification result, improve the security inspection efficiency, thereby realizing automatic security inspection, and reducing the interference of human factors such as human experience and fatigue on the detection and identification results.

[0100] Preferably,

[0101] Performing suspicious image content identification on the one or more image areas respectively to obtain one or more suspicious images, as shown in the appendix Figure 2 shown, includes:

[0102] S101, coarsely classify the one or more image regions according to the edge contours of the one or more image regions;

[0103] S102, obtain the target detection algorithm corresponding to the one or more image regions according to the coarse classification results of the one or more image regions;

[0104] S103, detect suspicious image content in the image region through the target detection algorithm. If target content is detected, regard the image region where the target content is detected as a suspicious image to obtain one or more suspicious images;

[0105] It can be understood that before the security inspection system executes step S102, rough classification is performed based on the extracted image regions. The purpose of this classification is to preliminarily confirm which field the content contained in the image region belongs to. The judgment method can be based on the edge contour of the image region, and a corresponding object detection algorithm is set for each field. For example, for dangerous items such as knives, guns, and flammable items, the YOLO algorithm is used; after obtaining the object detection algorithm corresponding to the suspicious image content to be recognized, in the security inspection scenario, the object detection algorithm is usually used to identify specific items in the X-ray image, such as knives, guns, or other contraband. These algorithms may be deep learning-based models, such as YOLO (You Only Look Once), SSD (Single Shot MultiBox Detector), or Faster R-CNN (Faster Region-based Convolutional Neural Networks), etc.; these algorithms have been trained with a large amount of labeled data and can accurately identify specific targets in the image; next, for each of the one or more extracted image regions, object detection is performed according to the selected object detection algorithm. In this step, the security inspection system inputs each image region into the object detection algorithm, and the algorithm will analyze the pixels and features in these regions to determine whether there are suspicious items. For example, if a hidden knife is contained in an image region, then after being processed by the YOLO algorithm, the security inspection system can identify the possibility of a knife existing in this region. Finally, if the detection result indicates that the target content is detected, the security inspection system obtains an identification result indicating that the image region contains suspicious image content; in other words, if the object detection algorithm detects a suspicious item, such as a knife or a gun, in a certain image region, then the security inspection system will mark this region as containing suspicious image content and may trigger further inspections and alarms. For example, if the YOLO algorithm detects a knife in an image region and the knife matches the knife features learned by the algorithm before, then the security inspection system will generate an identification result indicating that there is a suspicious knife image in this region. This identification result will subsequently be used in subsequent investigation and verification steps to determine whether there is really a dangerous item.

[0106] Preferably,

[0107] The detection of suspicious image content in the image region by the object detection algorithm is as shown in the attached Figure 3 figure and includes:

[0108] S1031, obtaining the feature type of the image region;

[0109] S1032. If the feature type of the image area is a color feature, perform a color feature comparison on the image area through the corresponding target detection algorithm to determine the target color;

[0110] S1033. Perform supplementary feature detection on the image area through the target detection algorithm, and fuse the target color and the supplementary features based on the distribution of the target color and the supplementary features in the image area respectively, and determine whether it is the target content based on the fusion result;

[0111] It can be understood that when the suspicious image content to be recognized is color feature data, the security inspection system performs a color feature comparison on the selected image area according to the target detection algorithm, which means that the security inspection system analyzes the color values of the pixels in this area and compares them with the predefined suspicious color feature data. For example, some dangerous items may have specific color features. For example, a blue liquid may indicate a certain flammable substance. Therefore, the security inspection system first checks whether there is such a specific blue feature in the image area. Once the security inspection system detects the target color, it performs supplementary feature detection on the selected image area according to the target detection algorithm, such as the coverage range of the color. This step is to further confirm whether the initially detected color feature has sufficient significance and range to support the identification of suspicious items. For example, if the initially detected blue area is very small, it may not be sufficient to determine it as a dangerous item; while if the blue area is large and continuous, it is more likely to be a sign of a dangerous item; finally, when the security inspection system detects the supplementary feature (such as the coverage range of the color), it fuses the detected target shape and the distribution of the supplementary feature in the image area to obtain an identification result representing the detected target. This step comprehensively considers multiple features such as color and shape to enhance the accuracy of identification; for example, if the security inspection system not only detects a large area of blue area, but also detects that this area presents a certain specific shape (such as the outline of a bottled liquid), then the security inspection system will be more confident in determining that there is a suspicious item in this area; through steps such as color feature comparison, supplementary feature detection, and multi-feature fusion, a detailed detection of suspicious items in the selected image area is realized. This method combines multiple features such as color and shape, improving the identification ability of the security inspection system for dangerous items.

[0112] Preferably,

[0113] The suspicious image objects in the one or more image areas to be investigated are extracted respectively, as shown in the appendix Figure 4 shown, including:

[0114] S301. Determine the specific types of identifiable items in the one or more image areas to be investigated;

[0115] S302. Perform object detection for the one or more image regions to be inspected under a specific type of target item.

[0116] S303. Extract the texture features of the target object based on the object detection results.

[0117] S304. Fuse the object shapes extracted from the one or more image regions to be inspected and the corresponding texture features to obtain one or more suspicious image objects.

[0118] It can be understood that in this embodiment, when it is necessary to extract suspicious image objects from the image regions to be inspected, the security inspection system will perform a series of delicate operations to identify and confirm potential dangerous items. The security inspection system performs object detection for the selected image regions to be inspected under a specific type of target item. This means that the security inspection system will focus on searching for and locating specific types of items that may exist in the image, such as knives, guns, or other contraband. For example, if the security inspection system is set to detect knives, it will use object detection algorithms such as YOLO (You Only Look Once) or FasterR-CNN (Faster Region-based Convolutional Neural Network) to scan the image and locate all shapes suspected to be knives. When the security inspection system detects one or more object shapes under a specific type of target item and successfully extracts the corresponding object texture, it fuses these two parts of information. The fusion process is to combine the shape information and the texture information to form a more complete and specific representation of the suspicious image object. This can enhance the accuracy of object recognition because shape and texture usually complement each other and jointly improve the robustness of recognition. For example, if the security inspection system detects an object with the shape of a knife and the texture of the object also matches that of a knife, then the security inspection system can be more confident in determining that this object is a knife. By combining technical means such as object detection, texture extraction, and information fusion, the security inspection system can more accurately identify and extract suspicious image objects from a complex image background, thereby improving the efficiency and security of security inspection.

[0119] Preferably,

[0120] The performing object detection for the one or more image regions to be inspected under a specific type of target item, as shown in the appendix Figure 5 includes:

[0121] S3021, Obtain a pre-constructed target detection transfer network for a specific type of identifiable object; the target detection transfer network is obtained through transfer learning based on a target detection overall network for a specific type of identifiable object; the target detection overall network is debugged based on image examples carrying prior shape markings of objects under the specific type of identifiable object;

[0122] S3022, Input the one or more image regions to be inspected into the pre-constructed target detection transfer network respectively, and perform target detection on the image regions to be inspected through the target detection transfer network;

[0123] S3023, The target detection transfer network is obtained through transfer learning based on a target detection overall network for a specific type of identifiable object;

[0124] S3024, The target detection overall network is debugged based on image examples carrying prior shape markings of objects under the specific type of identifiable object;

[0125] It can be understood that the security inspection system obtains a pre-debugged target detection transfer network for a specific type of identifiable object. This transfer network is obtained through transfer learning, that is, fine-tuning on a pre-trained target detection overall network to adapt to a specific type of identifiable object. The pre-debugging, that is, the pre-trained target detection overall network, is obtained by debugging with a large number of image examples carrying prior shape markings of objects under the specific type of identifiable object (prior markings are pre-determined markings). For example, if knives are to be detected, then a large number of image examples marked with the shape of knives are used to train this network; specifically, assuming the specific item to be detected is a knife, then a deep learning model such as Faster R-CNN can be used as the target detection overall network. In the training stage, a large number of images containing knives are used, and the shape of the knives is accurately marked; through training with these marked image examples, the model can learn the shape features of the knives, so as to accurately identify the knives in subsequent detection tasks; Next, the security inspection system loads the selected image regions to be inspected into the pre-debugged target detection transfer network. This process is equivalent to inputting the image to be detected into the neural network, and then the network will output the detection result. Taking knife detection as an example, when an image region to be inspected is input into the transfer network, the network outputs whether there is a knife in this region and the specific position of the knife; By using the target detection network of transfer learning, the security inspection system can accurately detect specific items such as knives in the image regions to be inspected. This method not only improves the detection accuracy but also enhances the intelligence and automation of the security inspection system.

[0126] Extract the texture features of the target object based on the target detection result, as shown in the appendix Figure 6As shown, it includes:

[0127] S3031, obtain the pre-built representative texture image set under the specific identification object type, where the representative texture image set is a set of typical texture samples under the specific identification object type;

[0128] S3032, perform an image block splitting operation on the to-be-screened image regions respectively to obtain an image block set;

[0129] S3033, calculate the similarity between each image block in the image block set and each representative texture image in the representative texture image set;

[0130] S3034, select the image blocks with a similarity greater than a preset first similarity threshold to the representative texture image as representative image blocks;

[0131] S3035, extract the texture of the representative image blocks as the texture feature of the target object;

[0132] It can be understood that the security inspection system extracts the texture of the object under the specific identification object type for the same to-be-screened image region. Texture is an inherent property of the object surface and can provide important information about the object's material and structure. In this step, the security inspection system uses image processing technology to extract the surface texture features of the suspected item. For example, for the region suspected of containing a knife, the security inspection system will analyze the pixel distribution, gray-level co-occurrence matrix, etc. of this region to capture its unique texture pattern;

[0133] During this process, the security inspection system performs two key operations: obtaining a representative texture image set for a specific type of identifiable item, and splitting the selected image region to be inspected to obtain a set of image patches; First, the security inspection system obtains a representative texture image set for a specific type of identifiable item. This image set is a collection of typical texture samples of items of this type and is used for subsequent texture comparison and identification. For example, if the security inspection target is to identify carried knives, the security inspection system pre-collects and organizes typical texture images of the surfaces of various knives to form a representative texture image set. These texture images may include the surfaces of knives made of different materials and with different degrees of wear to ensure the accuracy and comprehensiveness of subsequent detections; Second, the security inspection system performs an image patch splitting operation on the selected image region to be inspected to obtain a set of image patches. This means that the security inspection system divides the entire image region into several small pieces, each of which is called an image patch. The purpose of splitting the image patches is to analyze the texture features in the image more carefully and improve the detection accuracy. For example, if the image region to be inspected is an X-ray image containing multiple items, the security inspection system will split this region into multiple small image patches, and each image patch may only contain a part of the item; For example, assume that the security inspection system is examining an X-ray image that contains a backpack and its contents. The security inspection system first obtains a representative texture image set of the dangerous items (such as knives, guns, etc.) that may be carried in the backpack. Then, the security inspection system will split the backpack region in the X-ray image into a series of image patches, and these image patches will be used for subsequent texture comparison to determine whether there are texture features that match the representative texture image set, so as to judge whether there are dangerous items hidden in the backpack. Through these two operations, the security inspection system lays a foundation for subsequent texture comparison and item identification, improving the accuracy and efficiency of security inspection; Next, the security inspection system performs a key operation: comparing each image patch in the obtained set of image patches with each representative texture image in the pre-collected representative texture image set. Specifically, the security inspection system traverses each image patch in the set of image patches, and then for each image patch, it will compare with all the representative texture images in the representative texture image set one by one. This comparison is achieved by calculating the similarity between the image patch and the representative texture image. There are various methods for calculating the similarity, such as the Structural Similarity Index (SSIM), cosine similarity, or feature point matching algorithms;For example, if the security inspection target is to identify knives, the representative texture image set will contain the textures of knives in various different styles and textures. During the comparison process, the security inspection system will calculate the similarity between each image block and these knife textures. If the texture of an image block is highly similar to a certain knife texture in the representative texture image set, then this image block is very likely to be part of a knife. This comparison process is crucial for ensuring the accuracy and efficiency of security inspection. Through meticulous comparison of image blocks, the security inspection system can more accurately identify potential dangerous items. Even if the item is partially blocked or overlapped with other items, it can be discovered through comparison of local textures. In practical applications, to improve the comparison efficiency and accuracy, the security inspection system can adopt efficient image processing and machine learning algorithms to accelerate the comparison process and reduce false alarms and missed detections. For example, deep learning models can be used to extract the features of image blocks and then compare them with the features in the representative texture image set to achieve fast and accurate identification; if any one of the image blocks in the image block set is paired with any one of the representative textures in the representative texture image set, the texture corresponding to the paired image block is used as the object texture obtained by extraction. If any one of the image blocks in the image block set is paired with any one of the representative textures in the representative texture image set (i.e., the similarity between the two exceeds a preset threshold), then the security inspection system uses the texture corresponding to this paired image block as the object texture obtained by extraction; specifically, when the security inspection system makes a comparison and finds that some image blocks are highly similar to the textures in the representative texture image set, this similarity is determined by calculating the similarity. For example, if the similarity between the texture of a certain image block and the representative texture image of a knife reaches more than 90% (this threshold can be adjusted according to the actual situation), then the security inspection system believes that this image block is paired with the knife texture. Once a pairing occurs, the security inspection system records the texture of this image block and uses it as the object texture extracted from the image area to be inspected. This object texture will play an important role in subsequent analysis and judgment because it may represent a part of a potential dangerous item. For example, in the inspection of X-ray security images, if the security inspection system detects that the texture of an image block is highly similar to the texture of a knife, then this texture will be extracted and saved as the object texture. Subsequently, security personnel can further determine whether this image block actually represents a knife based on this object texture and decide whether further security checks are required.;

[0134] Preferably,

[0135] The calculation of the similarity between each image block in the image block set and each representative texture image in the representative texture image set, as shown in the appendix Figure 7 shows, includes:

[0136] S30331, save the representative texture images belonging to the same troubleshooting attribute type in the representative texture image set in the same storage space, and each storage space corresponds to a different texture attribute type;

[0137] S30332, obtain a pre-built image recognition network; the image recognition network is debugged according to an image sample set;

[0138] S30333, classify the image area to be troubleshot through the image recognition network to obtain an image classification result;

[0139] S30334, determine the storage space of the texture attribute type to which the corresponding representative texture image belongs according to the image classification result;

[0140] S30335, calculate the similarity between each image block in the image block set and each representative texture image saved in the determined storage space;

[0141] It can be understood that a pre-debugged image recognition network is obtained, and the image recognition network is debugged based on an image sample set, and the image sample set includes image samples with image classification prior labels and the same type of texture attributes. Specifically, the image recognition network is trained through a data set containing a large number of image samples. This data set is the image sample set, which contains images of various objects, and these images have been labeled with corresponding categories, namely, image classification prior labels. These prior labels play a role of supervised learning in the training process, helping the network learn how to correctly classify different objects. In addition, the image samples of this image sample set have a diversity of texture attribute types, that is, the data set not only contains images of different types of objects, but also covers various texture changes of these objects as much as possible. For example, for the category of knives, the data set may contain images of knives of different materials, different shapes, and different degrees of wear to ensure that the network can learn various texture features, so as to more accurately identify knives in practical applications. During the training process, the image recognition network continuously adjusts its internal parameters to minimize the difference between the predicted category and the prior label. This process is achieved through optimization algorithms, such as the commonly used gradient descent algorithm. After a large number of After iterative training, the network will gradually learn the ability to extract effective features from images and accurately classify them. Finally, when the security inspection system needs to identify an object in a certain image area to be checked, it will call this pre-debugged image recognition network to perform classification prediction. In this way, the security inspection system can quickly and accurately identify the category of objects in the image, providing an important basis for subsequent security inspection processes; calling the previously acquired image recognition network, this network has been trained with a large number of image sample sets and has the ability to identify different categories of objects; now, suppose the security inspection system is checking a package and has been trained in some way (e.g. X-ray scanning) obtains an image of the interior of the package. In this image, the security inspection system detects an area suspected of containing dangerous goods. This is the area of the image that needs to be checked. The security inspection system inputs this area of the image to be checked into the image recognition network. The network processes this image area and extracts the features of the image through its internal convolutional layer, pooling layer and other structures. The image is then classified based on these features. The classification result may be a specific category of items, such as knives, "guns" or "flammable items". This result is based on the knowledge learned by the network from the training data, and it reflects the network's understanding of the content of the input image area.For example, if the image area to be inspected actually contains a knife, then after processing this area, the image recognition network can output "knife" as the classification result. In this way, the security inspection system knows the type of dangerous item that may exist in this area, and thus can take corresponding security measures. Based on the image classification result, which indicates the type of item in the image area to be inspected. For example, the security inspection system may have identified that this area contains a knife. Next, the security inspection system searches in its storage for the storage space corresponding to the texture attribute type of the classification result of "knife". This storage space has been previously set up to store all the representative texture images related to knives. These representative texture images were selected during the training process and can typically represent the texture characteristics of items of the knife category. For example, if the image classification result is "knife", then the security inspection system will look for the storage space dedicated to storing the representative texture images of knives. The images in this space may include various types of knives, such as kitchen knives, fruit knives, folding knives, etc. However, their common feature is that they all have the unique texture attributes of knives, such as sharp blades, metallic textures, etc. By comparing the classification result with the label of the storage space, the security inspection system quickly locates the correct storage space. The representative texture images in this storage space will be used in the subsequent image comparison step to help the security inspection system more accurately identify the specific item in the image. In this way, the security inspection system can efficiently screen out the texture images that match the classification result, providing an accurate data basis for subsequent detailed comparison and analysis. The security inspection system compares each image block in the image block set with the representative texture images stored in the previously determined storage space. This step aims to further verify and identify the specific item in the image block, ensuring the accuracy and security of the security inspection.Specifically, the security inspection system first obtains a set of image patches. These image patches are segmented from the image area to be inspected, and each image patch contains certain texture and shape information. Then, according to the determined storage space, which stores representative texture images of the same texture attribute type as the image classification result. For example, if the image classification result indicates that the image area to be inspected may contain knives, the determined storage space will contain representative texture images of various knives. The security inspection system compares each image patch in the set of image patches with each representative texture image in this storage space. The comparison process may be completed by calculating the similarity between the image patch and the representative texture image. There are various methods for calculating similarity, such as calculating the pixel difference between two images, using feature matching algorithms, or applying deep learning models to extract features and make comparisons. Suppose a certain image patch in the set of image patches shows a part of a knife blade. The security inspection system compares this image patch one by one with all the representative texture images of knives in the storage space. If a highly similar representative texture image is found, the security inspection system will increase its confidence that the item in this image patch is a knife. Through this step, the security inspection system can more accurately identify dangerous items in the image, such as knives, guns, or other contraband, thereby improving the accuracy and efficiency of security inspection. The application of this comparison method is not limited to knives and is also applicable to the identification of other types of items, which is an essential part of the security inspection process.

[0142] Preferably,

[0143] The step of calculating the similarity between the suspicious image object and each comparison image object of the dangerous item images in the pre-built dangerous item image library one by one is as shown in the appendix Figure 8 and includes:

[0144] S3001, traverse each dangerous item image in the dangerous item image library, and calculate the object shape similarity between the object shape included in the obtained suspicious image object and the object shape included in each comparison image object of the dangerous item images respectively to obtain the object shape similarity calculation result;

[0145] S3002, if the object shape similarity calculation result is greater than a preset second similarity threshold, add this dangerous item image to the temporary set of dangerous item images until the entire dangerous item image library is traversed;

[0146] S3003, obtain the part of each comparison image object of the dangerous item images in the temporary set of dangerous item images that matches the shape of the suspicious image object to obtain the comparison and screening content of each comparison image object of the dangerous item images;

[0147] S3004, calculate the similarity between the object texture included in the suspicious image object and the texture features of the comparison and screening content in the comparison image object of each dangerous goods image in the temporary concentration of dangerous goods images;

[0148] It can be understood that the security inspection system will first start a traversal program. This program will visit each image in the dangerous goods image library one by one. When each new dangerous goods image is visited, the security inspection system will extract the comparison image object in the current image, which is usually achieved through image processing and computer vision technologies. Then, the security inspection system will obtain the suspicious image objects obtained in the previous step and extract the shape information included in these objects. The extraction of shape information may involve image processing technologies such as edge detection and contour extraction to accurately capture the shape characteristics of the object. Then, the security inspection system compares the shape in the suspicious image object with the shape in the comparison image object of the currently visited dangerous goods image. This comparison is based on the calculation of shape similarity, and algorithms such as contour matching and shape context can be used to measure the similarity between the two shapes. For example, in the airport security inspection scenario, if the security inspection system detects a suspicious image object whose shape is similar to a knife, the security inspection system will compare this shape with the knife image stored in the dangerous goods image library. By calculating the shape similarity, the security inspection system can determine whether this suspicious image object matches the knife image in the library. Finally, each shape comparison will obtain an object shape comparison result, which will tell the security inspection system the similarity between the suspicious image object and the current dangerous goods image. If this similarity exceeds the preset threshold, then the security inspection system will consider that a potential dangerous good has been found and can trigger further inspections or alarms. In this way, the security inspection system quickly and accurately identifies suspicious items similar in shape to dangerous goods, thereby improving the efficiency and security of security inspections;

[0149] It is worth emphasizing that when the object shape comparison result shows that the suspicious image object matches the object shape of a certain dangerous goods image, the security inspection system adds this matching dangerous goods image to the temporary set of dangerous goods images. Specifically, when the security inspection system determines through the shape comparison algorithm that the similarity between the object in the suspicious image and the shape of an image in the dangerous goods image library exceeds the preset threshold, it is considered that the shape matching is successful. At this time, the security inspection system will perform an incorporation operation, which means adding this dangerous goods image that matches the shape of the suspicious image to a temporary storage area, which is called the temporary set of dangerous goods images. For example, in the airport security inspection scenario, if the security inspection system detects that the shape of an item in a luggage is highly similar to the shape of a prohibited knife in the library, the security inspection system will determine that this item may be a contraband. At this time, the security inspection system will add the image of this knife to the temporary set of dangerous goods images for further processing and confirmation later. The role of the temporary set of dangerous goods images is to temporarily store all dangerous goods images that match the shape of the suspicious image, so that security inspectors can conveniently view and compare these images to make accurate judgments. At the same time, this temporary set can also be used as a basis for further analysis and processing later, such as for generating reports, statistical data, etc. It should be noted that the images in the temporary set of dangerous goods images are not the basis for finally determining contraband, but provide a reference and warning for security inspectors to help them more quickly identify and confirm potential dangerous goods. Whether an item is finally determined to be contraband still needs to be determined in combination with other security inspection means and the professional judgment of security inspectors.After traversing the dangerous goods image library, compare the object texture contained in the obtained suspicious image object with the comparison and screening content in the comparison image object of each dangerous goods image in the temporary concentration of dangerous goods images to obtain a comparison result; after the security inspection system finishes traversing the entire dangerous goods image library, it starts to compare the object texture contained in the suspicious image object with the comparison and screening content of each image in the temporary concentration of dangerous goods images. Specifically, the security inspection system first extracts the object texture features in the suspicious image object. Texture features can describe the arrangement rules and periodicity between pixels or regions in an image. For example, the surface of an object may be rough, delicate, directional, etc.; these features can be extracted through image processing techniques such as gray-level co-occurrence matrix, Gabor filter, etc.; next, the security inspection system traverses each image in the temporary concentration of dangerous goods images. For each image in the temporary concentration, the security inspection system extracts the texture features of its comparison and screening content. Here, the comparison and screening content refers to the part determined to match the shape of the suspicious image in the previous step; then, the security inspection system compares the texture features of the suspicious image object with the texture features of the comparison and screening content of each image in the temporary concentration of dangerous goods images one by one. During the comparison process, various texture similarity measurement methods, such as histogram intersection, structural similarity index (SSIM), etc., can be used to measure the similarity between two textures. After the comparison is completed, the security inspection system will obtain a comparison result. This result is a quantitative index used to represent the similarity degree of the texture between the suspicious image and the dangerous goods image. If the similarity exceeds a preset threshold, then the security inspection system can determine that there is also a high degree of match in texture between the suspicious image and the dangerous goods image; for example, in the airport security inspection scenario, if the security inspection system detects that an item in a suspicious luggage is similar in shape to a knife and the image of this knife has been added to the temporary concentration of dangerous goods images, the security inspection system will further check whether the texture of this suspicious knife matches the texture of the knife image in the temporary concentration. If the match is successful, then the security inspection system will issue a higher-level alarm to prompt the security inspection personnel that this suspicious item is very likely to be a real knife.

[0150] Preferably,

[0151] Traverse each dangerous goods image in the dangerous goods image library, and calculate the object shape similarity between the object shape contained in the obtained suspicious image object and the object shape contained in the comparison image object of each dangerous goods image respectively to obtain the object shape similarity calculation result, as shown in the appendix Figure 9 shown, including:

[0152] S30011, group the semantics of the object shapes included in each dangerous item image in the dangerous item image library to obtain the grouping of each dangerous item image group, and obtain the centroid of each dangerous item image group;

[0153] S30012, obtain the spatial distance between the object shape included in the suspicious image object and the centroid of each group, and sort the spatial distances to determine the order of traversing each dangerous item image group;

[0154] S30013, according to the traversal order of each dangerous item image group, respectively obtain the similarity between the object shape included in the comparison image object of the dangerous item image in each dangerous item image group and the object shape included in the suspicious image object;

[0155] S30014, if the similarity between the object shape included in the suspicious image object and the object shape included in the comparison image object of any dangerous item image in a certain dangerous item image group is greater than the preset second similarity threshold, after traversing this dangerous item image group, do not continue to traverse the next dangerous item image group;

[0156] It is understandable that the security inspection system first groups the images in the dangerous goods image library and determines the centroid of each group. This process is mainly based on the semantics of the object shapes in the images to ensure that dangerous goods images with similar shapes are grouped together. Specifically, the security inspection system can adopt a clustering algorithm, such as K-means clustering, to group the images in the dangerous goods image library. During this process, the security inspection system extracts the shape features of the objects in each image. These features may include contours, edges, corner points, etc., and converts these features into a mathematical vector representation. These vectors capture the key information of the object shape and are the basis for clustering. Then, the security inspection system applies the clustering algorithm to divide these shape feature vectors into different groups, also known as clustering clusters. Each clustering cluster represents a type of dangerous goods image with a similar shape. For example, in the scenario of airport security inspection, all images of knife shapes can be grouped into one clustering cluster, while all images of gun shapes may be grouped into another clustering cluster. After clustering, the security inspection system calculates the centroid of each clustering cluster. The centroid is a central point representing the shape features of the clustering cluster, which is obtained by calculating the average value of all shape feature vectors within the cluster. This centroid plays a key role in the subsequent comparison process because it can quickly locate the group of dangerous goods images that are most likely to be similar to the suspicious image. For example, if the security inspection system detects a suspicious image object with a shape similar to a knife, the security inspection system will first calculate the similarity between this shape and the centroids of each clustering cluster. By comparing the similarities, the security inspection system can quickly find the clustering cluster that is most similar to the knife shape and conduct a more detailed comparison within this clustering cluster, thereby improving the efficiency and accuracy of security inspection.The security inspection system determines the spatial distance between the object shape contained in the suspicious image object and the centroid of each group of dangerous goods image groups. This distance can be understood here as the semantic distance, that is, the similarity between shape features. Then, based on these distances, the security inspection system checks each group of dangerous goods image groups where the centroid is located one by one in an increasing order. Specifically, the security inspection system first extracts the object shape features contained in the suspicious image object. These features may include contour information, corner points, area, etc. Then, these features are converted into a feature vector. For example, a feature vector may be represented as [0.5, 0.8, 0.3], where each value represents a different shape feature value. Next, the security inspection system calculates the spatial distance between this feature vector and the feature vectors of each group centroid. This calculation usually uses a mathematical distance formula, such as the Euclidean distance, to quantify the similarity between two shapes. For example, if the feature vector of the suspicious image is [0.5, 0.8, 0.3], and the feature vector of a certain group centroid is [0.4, 0.7, 0.5], then the Euclidean distance between them can be used as a measure of similarity. After calculating all the spatial distances, the security inspection system sorts these distances to determine an increasing order. This order indicates the similarity degree between the suspicious image and each group of dangerous goods image groups. The closer the distance, the higher the similarity. Then, based on this increasing order, the security inspection system traverses each group of dangerous goods image groups where the centroid is located one by one. Here, traversing means that the security inspection system will first check the group of dangerous goods images that is most similar to the suspicious image, that is, the group with the smallest spatial distance, and then sequentially check the groups with lower similarity; for example, in airport security, if the shape of the suspicious image object is most similar to the centroid of the knife shape group, the security inspection system will first check the images in this group to see if there is an item that matches the suspicious image. If no matching item is found, the security inspection system will continue to check the next most similar group until a matching dangerous good is found or all groups are checked. For each traversed group of dangerous goods image groups, the object shape contained in the suspicious image object obtained is compared with the object shape contained in the comparison image object of the currently traversed dangerous goods image one by one to obtain the object shape comparison result; the security inspection system continues its refined comparison process and checks each group of dangerous goods image groups determined to need to be traversed in the previous steps one by one. This process is to ensure that it can accurately identify whether there is a match between the suspicious image object and the images in the dangerous goods image library;Specifically, when the security inspection system moves to a specific group of dangerous item images, it will start to check each dangerous item image in the group one by one. In this step, the security inspection system will extract the object shape features contained in the suspicious image object again. These features may involve the outline, size, proportion of the object, and other key shape attributes. Subsequently, the security inspection system will compare the shape features of the extracted suspicious image object with the object shape features in the dangerous item image being currently inspected. This comparison process is achieved by calculating the similarity between the two shape features. The calculation of similarity may be based on various shape descriptors and matching algorithms, such as using shape descriptors like Hu moments, Zernike moments, or more complex shape matching algorithms. For example, if the suspicious image object is in the shape of a knife, and the currently traversed group of dangerous item images contains multiple images of knives in different styles, then the security inspection system will compare the shapes of these knife images with the suspicious image one by one. During the comparison process, the security inspection system will generate an object shape comparison result, which is a numerical value or a judgment criterion used to quantify the similarity between the two shapes. This object shape comparison result is crucial for subsequent decisions. If the comparison result shows that the suspicious image is highly similar to a certain dangerous item image, then the security inspection system can trigger an alarm to instruct the security personnel to conduct further inspections. On the contrary, if the comparison result is lower than a preset threshold, then the security inspection system will continue to check the next image in the group until all images have been inspected. If an object shape comparison result indicating a match to the object shape is obtained for the object shape comparison of the traversed group of dangerous item images, after traversing the group of dangerous item images where the centroid of each grouping is located, the traversal of the group of dangerous item images where the centroid of each grouping is located will end. When the security inspection system has completed the object shape comparison of all images in the currently traversed group of dangerous item images, the security inspection system will check these comparison results. If, during this process, the security inspection system obtains an object shape comparison result indicating a match to the object shape, that is, a dangerous item image similar to the shape of the suspicious image object is found, then the security inspection system will end the traversal of the current and subsequent groups of dangerous item images;Specifically, assume that the security inspection system is inspecting a suspicious image object similar in shape to a knife. The security inspection system has walked through the dangerous item image group containing various knife images through the previous steps and compared the shape of the object with each image in the group. If during this process, the security inspection system finds that a certain knife image highly matches the shape of the suspicious image object, that is, the object shape comparison result exceeds the preset matching threshold, then the security inspection system will determine that a dangerous item matching the suspicious image has been found. At this time, since a matching item has been found, it is no longer necessary for the security inspection system to continue walking through other dangerous item image groups. Therefore, after confirming the matching result, the security inspection system will immediately end the walk through the current and subsequent dangerous item image groups and can trigger the corresponding alarm or notification mechanism so that the security inspection personnel can intervene and handle it in a timely manner. For example, if the security inspection system finds a suspicious knife shape during the baggage inspection at an airport, the security inspection system will perform image comparison according to the above steps. When the security inspection system walks through the dangerous item image group containing knife images and finds an image that matches the suspicious image, the security inspection system will stop further walking and immediately notify the security inspection personnel so that they can quickly respond to ensure the safety of the airport.;

[0157] Preferably, it further includes:

[0158] If there is no comparison image object in the dangerous item image library that matches the suspicious image object, obtain the expert dangerous item recognition result obtained by expert inspection of the image area to be investigated;

[0159] Integrate the suspicious image object and the suspicious image content in the image area to be investigated according to the expert dangerous item recognition result to obtain a dangerous item image;

[0160] Add the obtained dangerous item image to the dangerous item image library for iterative update of the dangerous item image library;

[0161] It is understandable that when the security inspection system discovers through comparison that there is no matching comparison image object in the dangerous goods image library for the obtained suspicious image object, the security inspection system seeks the help of experts (i.e., manual) for further investigation. This means that although the security inspection system has powerful image recognition and processing capabilities, in some cases, the professional knowledge and experience of human experts are still required for judgment. The expert carefully examines the image area to be investigated and gives the dangerous goods recognition result based on their professional knowledge and experience. If the expert confirms that the suspicious image content in this area is indeed a dangerous good, the security inspection system will take the next step. First, the security inspection system integrates the suspicious image object and the suspicious image content in the image area to be investigated to generate a new dangerous goods image. This image will be added to the dangerous goods image library for reference during future security inspections. In this way, the dangerous goods image library of the security inspection system will be continuously iterated and improved, which not only helps to improve the recognition ability of the security inspection system for known dangerous goods but also enables it to gradually adapt to newly emerging types of dangerous goods. For example, in the airport security inspection scenario, if the security inspection system detects that an item in a luggage does not match any item in the existing dangerous goods image library in terms of shape and texture, the security inspection system will request an expert to conduct an investigation. If the expert confirms that this is a new type of dangerous good, such as a new type of explosive device, then the security inspection system will generate a new dangerous goods image and add it to the image library. In this way, during future security inspections, if a similar item is encountered again, the security inspection system will be able to quickly and accurately identify it. By introducing the expert investigation and image library iteration mechanism, the recognition ability of the security inspection system for unknown dangerous goods is enhanced, and the accuracy and security of security inspection are improved.

[0162] The acquisition of the dangerous goods images in the dangerous goods image library includes:

[0163] Extract multiple comparison image regions from the comparison images in the pre-built comparison image library;

[0164] Extract the suspicious image content and the suspicious image object of each comparison image region respectively;

[0165] Take the suspicious image content and the suspicious image object of the same comparison image region as candidate suspicious merged data, and determine the occurrence density corresponding to each candidate suspicious merged data;

[0166] Take the candidate suspicious merged data with an occurrence density greater than the preset density threshold as the dangerous goods image;

[0167] It is understandable that the first task of the security inspection system is to obtain a comparison image library and extract multiple comparison image regions from the comparison images included in the comparison image library. This step establishes a basic data set for comparison and analysis. First, the security inspection system collects images from different sources to build a rich comparison image library. These images can come from historical security inspection records, public data sets, or other reliable image resources. For example, in the airport security inspection scenario, the comparison image library may contain X-ray images of luggage in various shapes and sizes, which capture different arrangements and combinations of items inside the luggage, providing rich comparison materials for the security inspection system. Next, the security inspection system uses image processing techniques to perform region segmentation on each image in the comparison image library and extracts multiple comparison image regions. Region segmentation can be achieved through edge detection, threshold segmentation, or more advanced deep learning segmentation algorithms (such as the U-Net neural network). Each comparison image region represents an independent object or a part of an object in the image. Taking the X-ray luggage image as an example, through region segmentation, the security inspection system can decompose a complex luggage image into several simple regions, such as the electronic device region, the clothing region, the liquid bottle region, etc. These regions will serve as the basic units for subsequent analysis and comparison. The security inspection system conducts in-depth analysis on each comparison image region to extract suspicious image content and objects. First, the security inspection system selects a comparison image region for analysis. This region may be automatically determined based on the previous segmentation algorithm or manually selected by security inspectors. The selected region usually contains some abnormal or suspicious features that may indicate the presence of dangerous items. Then, the security inspection system uses image processing techniques and machine learning algorithms to extract the suspicious image content and objects within this region (refer to the steps mentioned above). For example, the security inspection system can identify regions with colors different from common items through color analysis or detect items with abnormal shapes through shape analysis. In addition, features such as density and texture may also be used to identify dangerous items. In a specific security inspection scenario, such as airport luggage security inspection, the security inspection system can detect a high-density region, which is usually associated with metal or other potential dangerous materials. The security inspection system will mark this region as suspicious and further analyze the surrounding items and features. To more accurately identify suspicious items, the security inspection system can adopt deep learning models, such as convolutional neural networks, to perform more advanced analysis on the images. These models can learn to identify the features of various dangerous items through training and automatically detect these features in new images. For example, if the security inspection system detects an object shaped like a gun, it can use a pre-trained model to confirm whether this object is really a gun. This process includes extracting the feature vector of the object and then comparing it with the known feature vectors in the model;Next, the main task of the security inspection system is to determine candidate suspicious merged data based on the suspicious image content and suspicious image objects originating from the same comparison image area, and further determine the occurrence density corresponding to these data respectively. First, based on the extracted suspicious image content and objects, which may include abnormal shapes, uneven densities, or other characteristics similar to dangerous goods, the security inspection system will conduct a detailed analysis of these suspicious images to find the correlations and similarities between them. Then, the security inspection system merges the suspicious image content and objects originating from the same comparison image area to form candidate suspicious merged data. For example, if similar metallic block-shaped objects are detected in the X-ray images of multiple pieces of luggage, the security inspection system will merge this data to form a candidate suspicious merged data regarding the metallic block-shaped object. Then, the security inspection system calculates the occurrence density of each candidate suspicious merged data. The occurrence density refers to the frequency and distribution of these data in the comparison image library. Through statistics and analysis, the security inspection system can determine which candidate suspicious merged data occur frequently and are thus more likely to be characteristics of dangerous goods. For example, in the airport security inspection scenario, if the security inspection system detects suspicious images similar to knives in multiple pieces of luggage and these images appear in similar positions in the luggage, then the security inspection system will merge these images into candidate suspicious merged data. Then, by calculating the occurrence density of these data in the luggage images, the security inspection system can determine whether images similar to knives occur frequently and thus determine whether they are dangerous goods. The security inspection system determines which data truly represent dangerous goods based on the previously calculated occurrence density of the candidate suspicious merged data and determines the dangerous goods images accordingly. First, the security inspection system sets a preset occurrence density threshold, which is set based on historical data, safety standards, and actual requirements and is used to determine which candidate suspicious merged data are frequent enough to be regarded as potential dangerous goods characteristics. Next, the security inspection system compares the occurrence density of each candidate suspicious merged data with this preset threshold. Only when the occurrence density of the candidate suspicious merged data is greater than the preset threshold will the security inspection system regard it as a characteristic representing dangerous goods. For example, in the airport security inspection scenario, if the security inspection system finds that the occurrence density of candidate suspicious merged data similar to knives is significantly higher than the preset threshold, then the security inspection system determines that the probability of such images being dangerous goods images is very high and marks them as dangerous goods images. Once the dangerous goods images are determined, the security inspection system will add them to the dangerous goods image library, and the images in this library will be used in subsequent security inspection processes to help automatically and quickly and accurately identify potential dangerous goods.;

[0168] The step of separately identifying suspicious image content for the one or more image regions to obtain one or more suspicious images further includes:

[0169] Based on the matching of the comparison suspicious image content in the dangerous goods image where the comparison image object is located, label the suspicious image content in the image area to obtain a labeled image;

[0170] Display the labeled image and give an alarm;

[0171] It can be understood that in the airport security inspection scenario, in addition to accurately identifying dangerous goods in luggage, the security inspection system also needs to be able to quickly and effectively convey the identification results to the security inspection personnel so that timely countermeasures can be taken. This derivative solution is designed for this need. When the security inspection system identifies a risk of dangerous goods in the X-ray image of the luggage to be identified through the method described above, the security inspection system will further perform a series of operations to ensure that the risk is handled in a timely and effective manner; First, based on the matching of the comparison suspicious image content in the dangerous goods image where the comparison image object is located, label the corresponding suspicious image content in the X-ray image of the luggage to be identified. This labeling can be in the form of color highlighting, border marking, text description, etc., so that the security inspection personnel can quickly locate and identify the dangerous goods in the image. For example, if the security inspection system identifies an item suspected to be a knife, it may frame the item with a red border on the X-ray image and add a text description of suspected knife beside it. After the labeling is completed, the security inspection system will generate a labeled image, which not only contains the original X-ray image information but also superimposes the labeling content of the dangerous goods by the security inspection system; Such a labeled image can intuitively display the potential risk points in the luggage, greatly improving the work efficiency and accuracy of the security inspection personnel; Next, the security inspection system displays this labeled image, which is usually achieved through the display beside the security inspection equipment, so that the security inspection personnel can view and handle potential risks in real time. At the same time, the security inspection system will also control the airport security inspection equipment to give an alarm, such as emitting a sound or a light prompt to attract the attention of the security inspection personnel; For example, if a luggage bag shows an item similar to a knife hidden inside after X-ray scanning, after the security inspection system identifies this risk, it will display an image with a red border labeled as suspected knife on the display and emit an alarm sound. After seeing this prompt, the security inspection personnel can immediately conduct a further inspection of the luggage to ensure safety. Through the labeling and display technology, the perception and processing speed of potential risks by the security inspection personnel are effectively improved, providing a more efficient and safe work process for airport security inspection.

[0172] Embodiment 2:

[0173] Figure 10 It is a schematic diagram of the device of an automatic identification system for dangerous goods in luggage security inspection based on artificial intelligence shown according to another exemplary embodiment. The system is applied to the security inspection system 1000 and includes:

[0174] A processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can run on the processor 1001, and when the processor 1001 executes the program, the steps in the method of any of the above embodiments are implemented;

[0175] The memory 1002 stores a computer program that can run on the processor. The memory 1002 is configured to store instructions and applications executable by the processor 1001, and can also cache data to be processed or already processed by the processor 1001 and each module in the system 1000 (for example, image data, audio data, voice communication data, and video communication data). It can be implemented by a flash memory (FLASH) or a random access memory (RAM). When the processor 1001 executes the program, the steps of the method for automatically identifying dangerous items in baggage security inspection based on artificial intelligence in any of the above items are implemented; the processor 1001 generally controls the overall operation of the security inspection system 1000.

[0176] This embodiment provides a storage medium that stores a computer program, and when the computer program is executed by the main controller, each step in the above method is implemented;

[0177] It can be understood that the above-mentioned storage medium can be a read-only memory, a magnetic disk, an optical disc, etc.

[0178] It can be understood that the same or similar parts in the above embodiments can be referred to each other, and the content not detailed in some embodiments can be referred to the same or similar content in other embodiments.

[0179] It should be noted that in the description of the present invention, terms such as "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. In addition, in the description of the present invention, unless otherwise specified, the meaning of "a plurality" refers to at least two.

[0180] Any process or method description shown in the flowchart or described in other ways herein can be understood as representing a module, segment, or part of code including one or more executable instructions for implementing a specific logical function or process. The scope of the preferred embodiments of the present invention includes additional implementations, where the functions can be executed in a substantially simultaneous manner or in the reverse order according to the functions involved, rather than in the order shown or discussed, and this should be understood by those skilled in the technical field to which the embodiments of the present invention belong.

[0181] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits with logic gate circuits for implementing logical functions on data signals, application specific integrated circuits with appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0182] Those of ordinary skill in the art can understand that all or part of the steps carried by the methods of the above embodiments can be completed by instructing relevant hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiments.

[0183] In addition, in each embodiment of the present invention, the functional units can be integrated into a processing module, or each unit can exist physically alone, or two or more units can be integrated into one module. The above integrated module can be implemented in the form of hardware or in the form of a software functional module. When the above integrated module is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0184] The above-mentioned storage medium can be a read-only memory, a magnetic disk, an optical disk, etc.

[0185] In the description of this specification, the descriptions with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0186] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. An automatic identification method for dangerous items in baggage security inspection based on artificial intelligence, characterized in that, The method includes: Obtaining an X-ray image of the luggage to be recognized, extracting one or more image regions from the X-ray image, and respectively performing suspicious image content recognition on the one or more image regions to obtain one or more suspicious images; Respectively obtaining the confidence levels of the suspicious image objects included in the one or more suspicious images, and screening the one or more suspicious images based on a preset confidence level threshold to obtain one or more image regions to be investigated; Respectively extracting the suspicious image objects in the one or more image regions to be investigated, and calculating the similarity between the suspicious image objects and the comparison image objects of each dangerous item image in a pre-established dangerous item image library one by one; The respectively extracting the suspicious image objects in the one or more image regions to be investigated includes: Determining the specific recognized item types of the one or more image regions to be investigated; Respectively performing object detection under the specific recognized item types on the one or more image regions to be investigated; Extracting the texture features of the target object based on the object detection results; Fusing the object shapes and the corresponding texture features extracted from the one or more image regions to be investigated to obtain one or more suspicious image objects; The respectively performing object detection under the specific recognized item types on the one or more image regions to be investigated includes: Obtaining a pre-constructed object detection transfer network for the specific recognized item types; the object detection transfer network is obtained by performing transfer learning based on an object detection overall network for the specific recognized item types; the object detection overall network is debugged based on image examples carrying prior markings of the object shapes under the specific recognized item types; Respectively inputting the one or more image regions to be investigated into the pre-constructed object detection transfer network, and performing object detection on the image regions to be investigated through the object detection transfer network; The object detection transfer network is obtained by performing transfer learning based on an object detection overall network for the specific recognized item types; The object detection overall network is debugged based on image examples carrying prior markings of the object shapes under the specific recognized item types; The extracting the texture features of the target object based on the object detection results includes: Obtaining a pre-constructed representative texture image set for the specific recognized item types, where the representative texture image set is a set of typical texture samples for the specific recognized item types; Performing an image block splitting operation on the image regions to be investigated respectively to obtain an image block set; Calculating the similarity between each image block in the image block set and each representative texture image in the representative texture image set respectively; Selecting the image blocks with a similarity to the representative texture image greater than a preset first similarity threshold as representative image blocks; Extracting the texture of the representative image blocks as the texture features of the target object; The calculating the similarity between each image block in the image block set and each representative texture image in the representative texture image set respectively includes: Save the representative texture images belonging to the same troubleshooting attribute type in the representative texture image set in the same storage space, and each storage space corresponds to a different texture attribute type; Obtain a pre-built image recognition network; the image recognition network is debugged according to an image sample set; Classify the image area to be troubleshot through the image recognition network to obtain an image classification result; Determine the storage space of the texture attribute type to which the corresponding representative texture image belongs according to the image classification result; Calculate the similarity between each image block in the image block set and each representative texture image stored in the determined storage space; Obtain a comparison image object that matches the suspicious image object in the pre-built dangerous goods image library according to the similarity calculation result, extract the comparison suspicious image content corresponding to the suspicious image object in the matched comparison image object, and verify the image area to be troubleshot where the suspicious image object is extracted through the comparison suspicious image content to obtain the verification result of the image area to be troubleshot; Determine the dangerous goods recognition result of the X-ray image of the luggage to be recognized according to the verification results of the one or more image areas to be troubleshot; If there is no comparison image object that matches the suspicious image object in the dangerous goods image library, obtain the expert dangerous goods recognition result obtained by expert troubleshooting of the image area to be troubleshot; Integrate the suspicious image object and the suspicious image content in the image area to be troubleshot according to the expert dangerous goods recognition result to obtain a dangerous goods image; Add the integrated dangerous goods image to the dangerous goods image library for iterative update of the dangerous goods image library; The acquisition of the dangerous goods images in the dangerous goods image library includes: Extract a plurality of comparison image areas from the comparison images in the pre-built comparison image library; Extract the suspicious image content and the suspicious image object of each comparison image area respectively; Use the suspicious image content and the suspicious image object in the same comparison image area as candidate suspicious merge data, and determine the occurrence density corresponding to each candidate suspicious merge data; Use the candidate suspicious merge data with an occurrence density greater than a preset density threshold as a dangerous goods image; The step of separately performing suspicious image content recognition on the one or more image areas to obtain one or more suspicious images further includes: Based on the comparison suspicious image content in the dangerous goods image where the matched comparison image object is located, label the suspicious image content in the image area to obtain a labeled image; Display and alarm the labeled image.

2. The method according to claim 1, wherein The step of separately performing suspicious image content recognition on the one or more image areas to obtain one or more suspicious images includes: Coarsely classify the one or more image areas according to the edge contours of the one or more image areas; Obtain the target detection algorithm corresponding to the one or more image areas according to the coarse classification results of the one or more image areas; Performing suspicious image content detection on the image region through the target detection algorithm. If target content is detected, the image region where the target content is detected is used as a suspicious image, and one or more suspicious images are obtained.

3. The method according to claim 2, wherein the performing suspicious image content detection on the image region through the target detection algorithm includes: obtaining the feature type of the image region; if the feature type of the image region is a color feature, performing color feature comparison on the image region through a corresponding target detection algorithm to determine the target color; performing supplementary feature detection on the image region through the target detection algorithm, and fusing the target color and the supplementary feature based on the distribution conditions of the target color and the supplementary feature in the image region respectively, and determining whether it is target content based on the fusion result.

4. The method according to claim 1, wherein the calculating the similarity between the suspicious image object and the comparison image objects of each dangerous goods image in the pre-built dangerous goods image library one by one includes: traversing each dangerous goods image in the dangerous goods image library, and calculating the object shape similarity between the object shape included in the obtained suspicious image object and the object shape included in the comparison image object of each dangerous goods image respectively, to obtain the object shape similarity calculation result; if the object shape similarity calculation result is greater than a preset second similarity threshold, adding the dangerous goods image to the temporary set of dangerous goods images until the entire dangerous goods image library is traversed; obtaining the part of the comparison image object of each dangerous goods image in the temporary set of dangerous goods images that matches the shape of the suspicious image object, to obtain the comparison and screening content of the comparison image object of each dangerous goods image; calculating the similarity between the object texture included in the suspicious image object and the texture features of the comparison and screening content in the comparison image object of each dangerous goods image in the temporary set of dangerous goods images.

5. The method according to claim 4, wherein the traversing each dangerous goods image in the dangerous goods image library, and calculating the object shape similarity between the object shape included in the obtained suspicious image object and the object shape included in the comparison image object of each dangerous goods image respectively, to obtain the object shape similarity calculation result, includes: grouping the semantics of the object shapes included in each dangerous goods image in the dangerous goods image library to obtain the grouping of each dangerous goods image group, and obtaining the centroid of each dangerous goods image group; obtaining the spatial distance between the object shape included in the suspicious image object and the centroid of each group, and sorting the spatial distances to determine the order of traversing each dangerous goods image group; according to the traversing order of each dangerous goods image group, respectively obtaining the similarity between the object shape included in the comparison image object of each dangerous goods image in each dangerous goods image group and the object shape included in the suspicious image object. If, within a group of dangerous item images, the similarity between the object shape included in the suspicious image object and the object shape included in the comparison image object of any dangerous item image in this group is greater than a preset second similarity threshold, then after traversing this group of dangerous item images, the system will not continue to traverse the next group of dangerous item images.

6. An automatic dangerous item recognition system for luggage security inspection based on artificial intelligence, characterized in that, The system includes a memory and a processor; The memory stores a computer program that can run on the processor, and when the processor executes the program, it implements the steps in the method described in any one of claims 1 to 5.

Images