AI-based security image intelligent recognition system

Abstract

This invention specifically relates to an AI-based intelligent security inspection image recognition system, belonging to the field of image recognition technology. It includes an uncertainty quantification module, used to keep the Dropout layer in the multi-scale feature extraction network active during the inference phase, perform a second forward propagation on the fused security inspection image, calculate cognitive uncertainty and localization uncertainty based on group detection results, and obtain an uncertainty score through weighted summation. In this invention, four types of security inspection images are fused at the pixel level, enabling the fused security inspection image to simultaneously carry multi-dimensional material feature information of both metals and organic matter, thus overcoming the deficiency of insufficient information in single-modal images that makes it difficult to distinguish items with similar materials.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to an AI-based intelligent security inspection image recognition system. Background Technology

[0002] In security checks, the detection of prohibited items is a crucial step in ensuring public safety. Existing security image recognition systems typically use single-view or dual-view X-ray images as input, relying on the material characteristics carried by a single modality image for judgment. Single-modality images have limited ability to distinguish between items of similar materials but different categories, and their accuracy decreases in complex scenarios such as occlusion and overlapping.

[0003] Furthermore, existing systems typically output a binary judgment of "contraband / non-contraband," lacking a quantitative expression of the reliability of the identification results. When the model encounters item forms with insufficient training data coverage, the system cannot proactively prompt security personnel to conduct manual verification, posing a risk of missed detections. Summary of the Invention

[0004] The purpose of this invention is to propose an AI-based intelligent security inspection image recognition system to solve the above-mentioned problems.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: AI-based intelligent image recognition system for security checks includes: The multimodal fusion module is used to acquire multimodal security inspection images of the same inspected item simultaneously and input them into the multimodal fusion network for pixel-level fusion processing, and output the fused security inspection image; The multi-scale detection module is used to input the fused security inspection image into a multi-scale feature extraction network for target detection and recognition, and output multi-scale detection results including target bounding boxes, item category labels and original confidence scores; The uncertainty quantization module is used to keep the Dropout layer in the multi-scale feature extraction network active during the inference phase and to perform quantification on the fused security inspection image. Second forward propagation, based on Cognitive uncertainty is calculated by taking the variance of the class probability of the same target in the group detection results. The variance of each coordinate component of the bounding box in the group detection results is used to calculate the positioning uncertainty, and the cognitive uncertainty and the positioning uncertainty are weighted and summed to obtain the uncertainty score; The hierarchical decision module is used to jointly determine the original confidence score and the uncertainty score of each detection result, and divide the detection results into three levels: high confidence detection, medium confidence detection and low confidence detection, and generate hierarchical recognition results.

[0006] Preferably, the multimodal security inspection images include transmission X-ray images, backscattered X-ray images, terahertz images, and visible light images, and the input layer receiver size of the multimodal fusion network is [missing information]. The image tensor, where Indicates the image height. Indicates the image width. The four modal channels are represented; before being input into the multimodal fusion network, the pixel values ​​of the four types of images are respectively subjected to Z-score normalization.

[0007] Preferably, the multimodal fusion network is a convolutional neural network based on an attention mechanism, comprising an encoder and a decoder. The encoder extracts features from the four modal channels respectively and calculates the weight coefficients of each modality at different spatial locations through an attention module. The decoder reconstructs the fused image by upsampling layers and convolutional layers using the weighted feature maps.

[0008] Preferably, the multi-scale feature extraction network adopts a feature pyramid structure, which includes a backbone network and a detection head. The backbone network generates feature maps of multiple scales through multi-layer convolution and pooling operations, and the detection head performs target detection in parallel on the feature maps of each scale.

[0009] Preferably, the cognitive uncertainty The calculation formula is: ; in, For the first The target class probability obtained from the first forward propagation for The average of the class probabilities of each forward propagation; The positioning uncertainty The calculation formula is: ; in, For the first The bounding box obtained by the first forward propagation is the first Each coordinate component For the first Each coordinate component is in The average value during each forward propagation.

[0010] Preferably, before the weighted summation, the steps are performed on... and Perform mean normalization based on the range to map both uniformly to... The interval; the uncertainty score The calculation formula is: ; in, and These are the weighting coefficients. , , .

[0011] Preferably, the hierarchical rule for the joint determination is: When the original confidence score is higher than the first threshold and the uncertainty score is lower than the second threshold, the detection result is identified as a high-confidence detection. When the original confidence score is between the first threshold and the third threshold, or when the original confidence score is higher than the first threshold and the uncertainty score is not lower than the second threshold, the detection result is considered as a moderately confident detection. When the original confidence score is lower than the third threshold, the detection result is identified as a low-confidence detection. Wherein, the first threshold is greater than the third threshold, and the second threshold is less than the fourth threshold.

[0012] Preferably, for detection results with an uncertainty score higher than the fourth threshold, even if their original confidence score is higher than the first threshold, the detection result is still classified into the medium confidence detection level.

[0013] Preferably, the method further includes: acquiring feedback data generated by security personnel after manually reviewing medium-confidence and low-confidence detection results, wherein the feedback data includes positive samples, corrected samples, and negative samples; and performing incremental training on the multimodal fusion network and the multi-scale feature extraction network based on the dataset obtained by merging the feedback data with the original training dataset, thereby updating the network parameters; wherein the learning rate of the incremental training is set to one-tenth to one-fifth of the initial training learning rate.

[0014] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention fuses four types of security inspection images at the pixel level, enabling the fused security inspection images to simultaneously carry multi-dimensional material feature information of metals and organic matter, thus overcoming the deficiency of single-modal image information that makes it difficult to distinguish items with similar materials.

[0015] 2. This invention performs incremental training by merging manual review feedback data with the original training data, and performs targeted data augmentation for easily confused categories in the difficult example library. The multimodal fusion network and multi-scale feature extraction network continuously improve their ability to distinguish new forms of contraband and easily confused items as the system runs. Attached Figure Description

[0016] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0017] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0018] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0019] Example 1

[0020] Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.

[0021] In this embodiment, it includes: Currently, intelligent recognition of security inspection images mainly relies on the analysis of images collected by single-type security inspection equipment, such as using only single-view or dual-view X-ray images as input. Single-modal images contain limited material feature information, making it difficult to distinguish items with similar materials but different categories. Furthermore, the output of existing recognition systems is typically a binary judgment of "contraband / non-contraband," lacking a quantitative expression of the reliability of the recognition result. This leads to a risk of missed detections when the model's judgment capability is insufficient in certain complex scenarios, failing to effectively prompt security personnel for verification.

[0022] To address the aforementioned issues, this embodiment proposes a security inspection image recognition method. Through the collaborative processing of multimodal image fusion, multi-scale feature extraction, uncertainty quantification assessment, and hierarchical decision output, it achieves quantitative expression and hierarchical handling of the credibility of security inspection image recognition results.

[0023] In this embodiment, the security checkpoint is equipped with multiple types of imaging devices, including transmission X-ray imaging devices, backscatter X-ray imaging devices, terahertz imaging devices, and visible light imaging devices, and each device simultaneously collects data on the same inspected item.

[0024] A security inspection image recognition method according to an embodiment of the present invention includes the following modules: Module 1: Acquire multimodal security inspection images and perform fusion processing to generate fused security inspection images; Spatial registration and temporal synchronization were performed on four types of images simultaneously acquired from the same inspected item to obtain aligned multimodal security inspection images. The four types of images are: transmitted X-ray images, backscattered X-ray images, terahertz images, and visible light images. Specifically, transmitted X-ray images carry information on material density and atomic number, backscattered X-ray images carry information on material surface features and effective atomic number, terahertz images carry contour information of non-metallic items, and visible light images carry information on the appearance of the package.

[0025] The aligned images of the four types are input into a multimodal fusion network for pixel-level fusion processing, and the fused security inspection image is output.

[0026] It should be noted that the aforementioned multimodal fusion network is a convolutional neural network based on an attention mechanism, and its input layer receiver size is [missing information]. The image tensor, where Indicates the image height (in pixels). Indicates the image width (pixels). This represents four modal channels. Before inputting into the multimodal fusion network, the pixel values ​​of the four image classes are Z-score normalized to eliminate the influence of differences in the units and numerical ranges of pixel values ​​output by different imaging devices on the fusion operation. The multimodal fusion network consists of an encoder and a decoder: the encoder is composed of multiple convolutional and pooling layers, extracting features from each of the four modal channels and then calculating the weight coefficients of each modality at different spatial locations through an attention module; the decoder reconstructs the fused image from the weighted feature maps through upsampling and convolutional layers. The output layer of the multimodal fusion network has an output size of [size missing]. Fusion of security inspection images.

[0027] The multimodal fusion network is trained using supervised learning, employing a labeled dataset containing multiple sets of four-modal images and their corresponding manually annotated target category and location information. The loss function consists of two parts: a structural similarity loss between the fused image and the four input modal images, and a detection loss based on the fused image for target detection. The Adam algorithm is used for optimization.

[0028] Module 2: Input the fused security inspection images into a multi-scale feature extraction network to obtain multi-scale detection results; The fused security inspection images are input into a multi-scale feature extraction network for target detection and recognition, outputting multi-scale detection results. Each detection result in the multi-scale detection results includes: target bounding box, item category label, and the original confidence score for that category. Item category labels include categories such as knives, firearms, explosives, flammable liquids, and controlled implements.

[0029] It should be noted that the multi-scale feature extraction network described above employs a convolutional neural network with a feature pyramid structure. The input layer of the multi-scale feature extraction network receives the fused security inspection image, and the output layer outputs multiple detection results, each containing bounding box coordinates. Category tag index and original confidence score ,in , The coordinates of the center point of the bounding box. , These represent the bounding box width and height, respectively. The multi-scale feature extraction network consists of two parts: a backbone network and a detection head. The backbone network generates feature maps at multiple scales through multiple layers of convolution and pooling operations. The detection head performs object detection in parallel on each scale feature map. Large-scale feature maps are used to detect large objects, medium-scale feature maps are used to detect medium-sized objects, and small-scale feature maps are used to detect small objects.

[0030] The multi-scale feature extraction network is trained using supervised learning, employing a labeled dataset containing fused security inspection images and their corresponding target bounding boxes and category labels. The loss function includes classification loss and bounding box regression loss, with cross-entropy loss for classification and IoU loss for bounding box regression. The Adam optimization algorithm is used.

[0031] Module 3: Quantitatively evaluate the uncertainty of multi-scale detection results and obtain the uncertainty score of each detection result; For each detection result, the Monte Carlo Dropout method is used to quantify uncertainty. This includes the following sub-modules: Module 3.1: During the inference phase, the Dropout layer in the multi-scale feature extraction network remains active, and the fused security inspection images are processed. Each forward propagation randomly discards different neurons while acquiring the same input. Group test results, among which This represents the preset number of forward propagations.

[0032] Module 3.2: Statistics The class probability distribution and bounding box position distribution of the same target in the group detection results are used to calculate cognitive uncertainty and localization uncertainty, respectively. Cognitive uncertainty characterizes the degree of uncertainty of the model parameters, through... The variance of the class probability of the forward propagation result is obtained by calculation. Positioning uncertainty characterizes the degree of dispersion of the result caused by noise in the input data itself, through... The variance of the bounding box coordinates in the forward propagation results is calculated.

[0033] Computational cognitive uncertainty The formula is: ; in, For the preset number of forward propagations, For the first The target class probability obtained from the first forward propagation for The average of the class probabilities of each forward propagation, i.e. .

[0034] Calculate positioning uncertainty The formula is: ; in, For the preset number of forward propagations, For the first The bounding box obtained by the first forward propagation is the first Each coordinate component For the first Each coordinate component is in The average value during each forward propagation, i.e. , Corresponding to bounding box coordinates The four components.

[0035] It should be noted that, It is calculated from the variance of the category probability, and its numerical range is: ; The variance of the bounding box coordinate components is accumulated, and its numerical range is affected by the range of coordinate values. To ensure that the two types of uncertainties have comparable numerical dimensions, before substituting them into the weighted summation formula in module 3.3, the variances of the bounding box coordinate components are calculated separately. and Perform mean normalization based on the range to map both uniformly to... Intervals are used to eliminate the influence of dimensional differences on weighted calculations.

[0036] Module 3.3: Combines the normalized cognitive uncertainty and localization uncertainty into an uncertainty score for the detection result. : ; in, This represents the cognitive uncertainty after normalization. The uncertainty in positioning after normalization and These are weighting coefficients used to balance the contributions of the two types of uncertainty. , , .

[0037] Module 4: Based on the original confidence score and uncertainty score, a joint decision is made to generate a hierarchical identification result; The original confidence scores and uncertainty scores of each detection result are jointly judged to classify the detection results into three confidence levels, generating a graded identification result.

[0038] Specifically, the grading rules are as follows: When the original confidence score is higher than the first threshold and the uncertainty score is lower than the second threshold, the detection result is identified as a high-confidence detection, and a contraband alarm is directly output and the target location is automatically marked on the image.

[0039] When the original confidence score is between the first and third thresholds, or when the original confidence score is higher than the first threshold and the uncertainty score is not lower than the second threshold, the detection result is considered as a moderately reliable detection, and a candidate alarm is output with a suggestion for manual review.

[0040] When the original confidence score is lower than the third threshold, the detection result is identified as a low-confidence detection, no alarm is output, but the suspected area is marked on the image with a weak hint.

[0041] In this embodiment, to avoid missed detections due to model overconfidence, it should be noted that for detection results with an uncertainty score higher than the fourth threshold and an original confidence score higher than the first threshold, the detection result is classified into the medium-confidence detection level according to the aforementioned medium-confidence detection clause. This process ensures that the model can proactively submit the results for manual review when facing unseen image shapes or severely occluded scenes, thus avoiding missed detections due to model overconfidence.

[0042] It should be noted that the first to fourth thresholds in the above grading rules satisfy the following constraints: the first threshold is greater than the third threshold, meaning the lower confidence limit corresponding to high-confidence detection is higher than the upper confidence limit corresponding to low-confidence detection, and the interval between the two thresholds corresponds to a medium confidence range; the second threshold is less than the fourth threshold, meaning the upper uncertainty limit corresponding to high-confidence detection is lower than the lower uncertainty limit corresponding to low-confidence detection, and the interval between the two thresholds corresponds to a medium confidence range. Due to uncertainty scoring... The range of values ​​after normalization is: Both the second and fourth thresholds are within Values ​​are taken within a range; the original confidence score is also within the range. The values ​​are taken within the range, and both the first and third thresholds are within the range. Values ​​can be taken within a range.

[0043] Module 5: Incrementally train the multimodal fusion network and multiscale feature extraction network based on manual review feedback data, and update the network parameters; This process acquires feedback data generated by security personnel after manually reviewing medium-confidence and low-confidence detection results. The feedback data includes three types of samples: corrected labels confirmed by the review are positive samples, labels corrected by the review are correction samples, and false alarm results rejected by the review are negative samples.

[0044] Based on the three types of samples mentioned above, the multimodal fusion network and the multi-scale feature extraction network are incrementally trained periodically to update the network parameters. Supervised learning is used for incremental training, merging newly added feedback data with the original training dataset. The merged dataset is then used to fine-tune the multimodal fusion network and the multi-scale feature extraction network. The loss function remains consistent with the initial training phase, and the Adam optimization algorithm is used. The learning rate is set to one-tenth to one-fifth of the initial training learning rate to maintain the stability of the learned knowledge.

[0045] In this embodiment, during manual review, security personnel also mark easily confused item pairs, for example, marking items that are similar in appearance but different in category as easily confused pairs. The information on easily confused item pairs is stored in a difficult example library. During incremental training, targeted data augmentation and training are performed on the easily confused categories in the difficult example library to improve the discrimination ability of the multimodal fusion network and the multi-scale feature extraction network for these easily confused categories.

[0046] Example 2

[0047] A security checkpoint at a transportation hub conducted a security check on a piece of carry-on luggage in a certain year. The luggage had a complex internal structure, containing multiple daily necessities and electronic devices. The security checkpoint was simultaneously equipped with transmission X-ray imaging, backscattered X-ray imaging, terahertz imaging, and visible light imaging equipment, all four types of equipment simultaneously capturing data from the luggage. In the initial system state, both the multimodal fusion network and the multi-scale feature extraction network had completed initial training, and the network parameters were usable. The objective of this security check task was to identify potential contraband in the luggage and output a graded identification result with a confidence level.

[0048] Example of Module 1: Multimodal security inspection image acquisition and fusion processing; After the four types of imaging devices simultaneously acquired images of the luggage, the system performed spatial registration and temporal synchronization processing on the four types of images. Transmission X-ray images showed a high-density area inside the luggage, with an outline similar to a knife shape; backscattered X-ray images showed strong surface scattering signals at the same location; terahertz images showed clear edges in this area; and visible light images showed no abnormalities in the luggage's appearance. The four aligned images were uniformly sized to 480 pixels high and 640 pixels wide, and after Z-score normalization, were formatted based on these dimensions. The image tensor is input to a multimodal fusion network. The encoder extracts features for each modal channel separately, and the attention module assigns higher weights to the transmitted X-ray channel and the backscattered X-ray channel in high-density regions. The decoder output size is [size missing]. The fused security inspection images are shown in Table 1 below.

[0049] Table 1. Image Acquisition and Registration Information for Four Modalities

[0050] Example of Module 2: Multi-scale feature extraction and detection result acquisition; After fusing security inspection images into a multi-scale feature extraction network, the backbone network generates feature maps at three scales. The large-scale feature map did not detect large items. The medium-scale feature map shows the image coordinates... A target was detected nearby, with the bounding box as follows: The category label is "knife," and the original confidence score is 0.82. Small-scale feature maps are located in the image coordinates... A target was detected nearby, with the bounding box as follows: The category label is "controlled equipment," and the original confidence score is 0.51. Both test results were then processed through a subsequent uncertainty quantification process, as shown in Table 2 below.

[0051] Table 2 Multiscale detection results

[0052] Example of Module 3: Quantitative Assessment of Uncertainty; The system performs Monte Carlo Dropout uncertainty quantification on the two detection results separately, with a preset number of forward propagations. It is 30 times.

[0053] For detection 1 (knife), in 30 forward propagations, the probability distribution of each category was relatively concentrated, with a high mean and small variance, indicating cognitive uncertainty. The initial values ​​are too low; the bounding box coordinates are stable across 30 propagations, indicating positioning uncertainty. The original values ​​were also too low. After normalization, both types of uncertainty were at a low level.

[0054] For detection 2 (controlled equipment), in 30 forward propagations, the category probability distribution was relatively dispersed with large variance, indicating cognitive uncertainty. The initial values ​​were too high; the bounding box coordinates fluctuated significantly in each propagation, indicating positioning uncertainty. The original values ​​were also too high. After normalization, both types of uncertainty remained at a high level.

[0055] Take weighting coefficients , According to the formula Calculate the final uncertainty score ,in To normalize cognitive uncertainty, The normalized positioning uncertainty is shown in Table 3 below.

[0056] Table 3. Quantitative Assessment Results of Uncertainty

[0057] Example of Module 4: Joint Decision Making and Generation of Hierarchical Identification Results; The system combines the original confidence scores and uncertainty scores of each detection result with the grading rules for joint judgment. In this scenario, the first threshold is set to 0.75, the third threshold to 0.45, the second threshold to 0.15, and the fourth threshold to 0.50.

[0058] The original confidence score of Detection 1 (knife) is 0.82, which is higher than the first threshold of 0.75, and the uncertainty score is 0.08, which is lower than the second threshold of 0.15. This meets the high-confidence detection criteria, and the system directly outputs a contraband alarm, which is then integrated into the security inspection image coordinates. The target location is automatically marked nearby.

[0059] The original confidence score of Detection 2 (controlled instrument) is 0.51, which is between the third threshold of 0.45 and the first threshold of 0.75, and the uncertainty score of 0.61 is higher than the fourth threshold of 0.50. According to the grading rules, the original confidence score of this detection result of 0.51 is between the first threshold of 0.75 and the third threshold of 0.45, and is classified into the medium confidence detection level. A candidate alarm is output with a suggestion for manual review, as shown in Table 4 below.

[0060] Table 4. Results of Joint Decision-Making Hierarchy Identification

[0061] Example of Module 5: Manual review and feedback and incremental training; Security personnel manually verified item 2 (controlled implement). Upon manual confirmation, the item was found to be a metal hairpin, a non-prohibited item, indicating a false alarm, and was classified as a negative sample. The security personnel also labeled "metal hairpin" and "controlled implement" as easily confused items and added this pair to the difficult case database.

[0062] During this incremental training cycle, the system collected a number of feedback data points, including positive samples, corrected samples, and negative samples. After merging the feedback data with the original training dataset, the multimodal fusion network and the multi-scale feature extraction network were fine-tuned, with the learning rate set to one-tenth of the initial training learning rate. For the easily confused pair "metal hairpins and controlled implements" in the difficult example library, targeted data augmentation was performed during incremental training to improve the network's ability to distinguish this easily confused category, as shown in Table 5 below.

[0063] Table 5 Summary of Manual Review Feedback Data

[0064] Throughout the implementation process, data starts from four modalities of images, undergoes spatial registration and Z-score normalization, and then enters the multimodal fusion network as image tensors, outputting a fused security inspection image. The fused security inspection image enters a multi-scale feature extraction network, generating detection results containing bounding box coordinates, category labels, and original confidence scores. The detection results undergo 30 forward propagations using the Monte Carlo Dropout method to calculate cognitive uncertainty and localization uncertainty, which are then normalized and weighted to synthesize an uncertainty score. The original confidence score and the uncertainty score are used together in a hierarchical decision-making process. Detection 1 is directly alerted because both indicators meet the high confidence condition, while Detection 2 is guided to the manual review stage because its uncertainty score exceeds the fourth threshold. Negative samples and easily confused pairs generated during manual review are fed back into the incremental training process, updating network parameters and forming a complete data loop.

[0065] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0066] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0067] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An AI-based security image intelligent identification system, characterized in that, include: The multimodal fusion module is used to acquire multimodal security inspection images of the same inspected item simultaneously and input them into the multimodal fusion network for pixel-level fusion processing, and output the fused security inspection image; The multi-scale detection module is used to input the fused security inspection image into a multi-scale feature extraction network for target detection and recognition, and output multi-scale detection results including target bounding boxes, item category labels and original confidence scores; an uncertainty quantification module configured to keep Dropout layers in the multi-scale feature extraction network active during an inference phase, perform a next forward propagation based on calculate a cognitive uncertainty based on a variance of class probabilities of a same target in group detection results, and calculate a positioning uncertainty based on a variance of each coordinate component of a bounding box in the group detection results, and obtain an uncertainty score by weighted sum of the cognitive uncertainty and the positioning uncertainty. The hierarchical decision module is used to jointly determine the original confidence score and the uncertainty score of each detection result, and divide the detection results into three levels: high confidence detection, medium confidence detection and low confidence detection, and generate hierarchical recognition results.

2. The AI-based intelligent security inspection image recognition system according to claim 1, characterized in that, The multi-modal security inspection image includes a transmission X-ray image, a backscattering X-ray image, a terahertz image and a visible light image, an input layer of the multi-modal fusion network receives an image tensor with a size of , wherein represents an image height, represents an image width, represents four modal channels; the pixel values of the four types of images are respectively subjected to Z-score standardization processing before the multi-modal fusion network is input. 3.The AI-based security check image intelligent identification system according to claim 1, characterized in that, The multimodal fusion network is a convolutional neural network based on an attention mechanism, which includes an encoder and a decoder. The encoder extracts features from the four modal channels respectively and then calculates the weight coefficients of each modality at different spatial locations through the attention module. The decoder reconstructs the fused image by upsampling layers and convolutional layers from the weighted feature maps.

4. The AI-based intelligent security inspection image recognition system according to claim 1, characterized in that, The multi-scale feature extraction network adopts a feature pyramid structure, which includes a backbone network and a detection head. The backbone network generates feature maps of multiple scales through multi-layer convolution and pooling operations, and the detection head performs object detection in parallel on the feature maps of each scale.

5. The AI-based intelligent security inspection image recognition system according to claim 1, characterized in that, Cognitive uncertainty The calculation formula is: ； in, For the first The target class probability obtained from the first forward propagation for The average of the class probabilities of each forward propagation; The positioning uncertainty The calculation formula is: ； in, For the first The bounding box obtained by the first forward propagation is the first Each coordinate component For the first Each coordinate component is in The average value during each forward propagation.

6. The AI-based intelligent security inspection image recognition system according to claim 1, characterized in that, Before weighted summation, respectively... and Perform mean normalization based on the range to map both uniformly to... The interval; the uncertainty score The calculation formula is: ； in, and These are the weighting coefficients. , , .

7. The AI-based intelligent security inspection image recognition system according to claim 1, characterized in that, The grading rules for joint determination are as follows: When the original confidence score is higher than the first threshold and the uncertainty score is lower than the second threshold, the detection result is identified as a high-confidence detection. When the original confidence score is between the first threshold and the third threshold, or when the original confidence score is higher than the first threshold and the uncertainty score is not lower than the second threshold, the detection result is considered as a moderately confident detection. When the original confidence score is lower than the third threshold, the detection result is identified as a low-confidence detection. Wherein, the first threshold is greater than the third threshold, and the second threshold is less than the fourth threshold.

8. The AI-based intelligent security inspection image recognition system according to claim 7, characterized in that, For detection results with an uncertainty score higher than the fourth threshold, even if their original confidence score is higher than the first threshold, the detection result is still classified as a medium-confidence detection level.

9. The AI-based intelligent security inspection image recognition system according to claim 1, characterized in that, Also includes: The system obtains feedback data generated by security personnel after manually reviewing medium-confidence and low-confidence detection results. The feedback data includes positive samples, corrected samples, and negative samples. Based on the dataset obtained by merging the feedback data with the original training dataset, the multimodal fusion network and the multiscale feature extraction network are incrementally trained to update the network parameters; wherein, the learning rate of the incremental training is set to one-tenth to one-fifth of the initial training learning rate.

Images