Security inspection equipment control method, device, equipment and storage medium based on AI recognition

Abstract

The present application relates to the field of AI recognition technology, and discloses a security inspection equipment control method, device, equipment and storage medium based on AI recognition. The method includes: photographing a person passing through the security inspection equipment through an image acquisition terminal to obtain an original face image; performing adaptive Gaussian smoothing and standardization processing on the original face image to obtain a pre-processed face image; using a multi-stage repair network to extract features from the pre-processed face image to obtain multi-scale face features; inputting the multi-scale face features into a face recognition model, iteratively matching the similarity from data to class to obtain a candidate identity list; combining the biometric data collected by the security inspection equipment, performing multi-modal fusion and Bayesian reasoning on the candidate identity list to obtain a comprehensive identity recognition result, and controlling the opening or closing of the security inspection equipment. The present application improves the accuracy of personnel identification matching, thereby improving the accuracy of security inspection equipment control.

Description

Technical Field

[0001] The present application relates to the field of AI recognition technology, and in particular to a security inspection equipment control method, device, equipment and storage medium based on AI recognition. Background Art

[0002] With the growing demand for social security, the application of efficient and accurate security inspection equipment in public places has become increasingly important. Traditional security inspection methods often rely on manual operations, which are not only inefficient but also prone to human negligence. In recent years, the rapid development of artificial intelligence technology has brought new opportunities to the security inspection field. Security inspection equipment based on AI recognition has begun to be gradually applied in public places such as airports and subway stations, but it still faces some challenges.

[0003] The recognition accuracy of existing AI recognition security inspection systems in complex environments still needs to be improved. For example, in conditions of insufficient light, face obstruction, or poor angles, the system's recognition effect is often unsatisfactory. In addition, single-modality recognition methods are easily interfered by environmental factors and are difficult to cope with diverse security inspection scenarios. Another prominent issue is the balance between security inspection efficiency and safety. Overly strict recognition standards may lead to a large number of false alarms and increase security inspection time; while overly loose standards may reduce safety. Therefore, how to improve security inspection efficiency while ensuring safety has become one of the key directions of current research. Summary of the invention

[0004] The present application provides a security inspection equipment control method, device, equipment and storage medium based on AI recognition, which are used to improve the accuracy of personnel identification and matching, thereby improving the accuracy of security inspection equipment control.

[0005] In a first aspect, the present application provides a security inspection equipment control method based on AI recognition, and the security inspection equipment control method based on AI recognition includes:

[0006] The image acquisition terminal on the security inspection device takes a picture of the person passing through to obtain the original face image;

[0007] Performing adaptive Gaussian smoothing and normalization processing on the original face image to obtain a preprocessed face image;

[0008] Using a multi-stage restoration network to extract features from the pre-processed face image to obtain multi-scale face features;

[0009] Inputting the multi-scale facial features into a preset facial recognition model, iteratively matching data to class similarity, and obtaining a candidate identity list of the passing person;

[0010] Combining the biometric data collected by the security inspection equipment, multimodal fusion and Bayesian reasoning are performed on the candidate identity list to obtain a comprehensive identity recognition result;

[0011] The security inspection device is controlled to be turned on or off according to the comprehensive identity recognition result.

[0012] In a second aspect, the present application provides a security inspection equipment control device based on AI recognition, and the security inspection equipment control device based on AI recognition includes:

[0013] A shooting module is used to shoot a person passing through the security inspection device through an image acquisition terminal to obtain an original face image;

[0014] A processing module, used for performing adaptive Gaussian smoothing and normalization processing on the original face image to obtain a preprocessed face image;

[0015] An extraction module, used for extracting features from the pre-processed face image using a multi-stage restoration network to obtain multi-scale face features;

[0016] A matching module, used for inputting the multi-scale facial features into a preset facial recognition model, iteratively matching the data to the class similarity, and obtaining a candidate identity list of the passing person;

[0017] A reasoning module, used to combine the biometric data collected by the security inspection equipment, perform multimodal fusion and Bayesian reasoning on the candidate identity list, and obtain a comprehensive identity recognition result;

[0018] A control module is used to control the opening or closing of the security inspection equipment according to the comprehensive identity recognition result.

[0019] The third aspect of the present application provides a security inspection equipment control device based on AI recognition, comprising: a memory and at least one processor, wherein instructions are stored in the memory; the at least one processor calls the instructions in the memory so that the security inspection equipment control device based on AI recognition executes the above-mentioned security inspection equipment control method based on AI recognition.

[0020] The fourth aspect of the present application provides a computer-readable storage medium, which stores instructions. When the computer-readable storage medium is run on a computer, it enables the computer to execute the above-mentioned security inspection equipment control method based on AI recognition.

[0021] In the technical solution provided by the present application, the application of the multi-stage repair network improves the accuracy and robustness of feature extraction. Through multi-level feature extraction and optimization, it is possible to effectively process face images in complex environments and improve recognition accuracy. The data-to-class similarity iterative matching mechanism enhances the flexibility and adaptability of face recognition. It is possible to dynamically adjust the matching strategy to effectively respond to different recognition scenarios. The introduction of multimodal fusion and Bayesian reasoning improves the reliability of recognition results. By combining multiple biometric data, the system can more comprehensively evaluate identity information and reduce the misrecognition rate. Adaptive Gaussian smoothing and standardization processing improve the quality of the original image. This step can effectively reduce the impact of environmental noise and provide clearer image input for subsequent recognition. The use of the supervised attention module optimizes the feature extraction process. By automatically focusing on important feature areas, the system can more accurately capture identity feature information. Multi-scale cross-stage feature fusion technology enhances the expressiveness of features. It is possible to comprehensively utilize feature information of different scales and levels to improve the accuracy and robustness of recognition. The dynamic threshold screening mechanism improves the adaptability of the system. By adjusting the recognition threshold according to real-time conditions, the system can maintain high efficiency and accuracy in different security inspection environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0023] Figure 1 This is a schematic diagram of an embodiment of a security inspection equipment control method based on AI recognition in an embodiment of the present application;

[0024] Figure 2 This is a schematic diagram of an embodiment of a security inspection equipment control device based on AI recognition in an embodiment of the present application. DETAILED DESCRIPTION

[0025] The embodiments of the present application provide a method, device, equipment and storage medium for controlling security inspection equipment based on AI recognition. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0026] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of the security inspection equipment control method based on AI recognition includes:

[0027] Step S101: photographing a person passing through the security inspection device through an image acquisition terminal to obtain an original face image;

[0028] It is understandable that the execution subject of the present application can be a security inspection equipment control device based on AI recognition, or a terminal or a server, which is not limited here. The embodiment of the present application is described by taking the server as the execution subject as an example.

[0029] Specifically, a high-performance image acquisition terminal is integrated on the security inspection equipment. The terminal usually includes a high-definition camera and a dedicated image processing chip to ensure that the facial image of the person passing through can be clearly captured under various lighting conditions. When someone approaches the security inspection equipment, the camera is automatically triggered to start working through a sensor or infrared sensor. At the same time, the image processing chip in the system starts working and cooperates with the camera to perform real-time image acquisition. This process includes adaptive adjustment of ambient light to optimize exposure time and white balance to obtain the best image quality. While the image is being acquired, real-time face detection is performed. Through the integrated computer vision algorithm, the face area is quickly located and locked in the video stream. Common face detection algorithms include convolutional neural networks based on deep learning and traditional Haar cascade classifiers. After the face is detected, the area is focused on to ensure that the acquired image is clear enough and contains enough facial feature details to obtain the original face image.

[0030] Step S102, performing adaptive Gaussian smoothing and normalization processing on the original face image to obtain a preprocessed face image;

[0031] Specifically, the brightness histogram of the original face image is analyzed, and the image brightness distribution characteristics are obtained by calculating the number of pixels of different gray levels in the image. By analyzing the brightness histogram, the areas with too high or too low brightness in the image can be identified. According to the image brightness distribution characteristics, the original face image is adaptively gamma corrected to adjust the overall brightness and contrast of the image. Adaptive gamma correction is to adjust the gamma value of the image so that the darker parts of the image become brighter and the brighter parts become appropriately darker, thereby obtaining a brightness-equalized image. Edge detection is performed on the brightness-equalized image to identify the intensity and position of the edge of the object in the image. Edge detection usually uses algorithms such as the Sobel operator and the Canny operator to generate an image edge intensity map by calculating the gradient of grayscale changes in the image. The image edge intensity map highlights the areas with large changes in the image, which usually correspond to important details in the image, such as the contour and feature points of the face. According to the image edge intensity map, the variance of the local area is calculated to obtain the adaptive Gaussian kernel parameters. The calculation of variance reflects the grayscale change of the image in the local area. By adaptively adjusting the Gaussian kernel parameters, the important edge details are retained while smoothing the image. Based on the adaptive Gaussian kernel parameters, the brightness equalized image is convolved to obtain a smoothed image after noise reduction. Gaussian smoothing is to remove high-frequency noise in the image by convolving the image with the Gaussian kernel, while retaining the low-frequency signal, making the image smoother and more natural. The smoothed image is subjected to geometric transformation and affine transformation to correct the tilt and distortion in the image to obtain a corrected face image. According to the preset face key point model, the feature points of the corrected face image are located to obtain the coordinates of the face key points. The face key point model is usually based on a deep learning algorithm. By training a large number of face images, it can accurately locate the important feature points on the face, such as the corners of the eyes, the tip of the nose, the corners of the mouth, etc. Based on the coordinates of the face key points, the corrected face image is cropped and scaled to obtain a face image of uniform size. By cropping and scaling, face images of different sizes and proportions are standardized to the same size, which is convenient for subsequent feature extraction and recognition. Normalize the pixel values ​​of face images of uniform size to obtain normalized face images. Scale the pixel values ​​in the image to a fixed range, such as between 0 and 1, to eliminate the brightness and contrast differences between different images and enhance the robustness of the model. Perform data enhancement on the normalized face images to increase the diversity of training data. Data enhancement includes random rotation, translation, scaling, mirroring and other operations. By introducing more changes in the training data, the model can be more robust to different postures, lighting and expression changes, and a preprocessed face image can be obtained.

[0032] Step S103, using a multi-stage restoration network to extract features from the pre-processed face image to obtain multi-scale face features;

[0033] Specifically, a multi-stage restoration network is designed, which consists of multiple convolutional layers and pooling layers. Through the step-by-step processing of these layers, feature information is extracted from the pre-processed face image. In the first stage, the pre-processed face image is processed by multi-layer convolution and pooling operations to obtain a low-level feature map. Upsampling and deconvolution operations are performed on the low-level feature map of the first stage to restore the resolution of the image while maintaining the integrity of the feature information, thereby obtaining the first-stage preliminary feature representation. In order to optimize the preliminary feature representation, the first supervised attention module is introduced to optimize the preliminary feature representation through supervised learning. The supervised attention module can adjust the feature representation weights according to the pre-defined goals, highlight important features, suppress irrelevant information, and obtain the optimized first-stage features. The optimized first-stage features are input into the second-stage encoder-decoder structure, and more advanced semantic features are extracted through more complex convolution and deconvolution operations. The second-stage advanced semantic features can capture deeper image information, including the shape and structure of the object. After the advanced semantic features are optimized by the second supervised attention module, the feature expression ability can be improved, and the optimized second-stage features can be obtained. The optimized second-stage features are input into a network based on the original input resolution. The network can refine the image while maintaining the original resolution to obtain a spatially accurate feature map. Multi-scale cross-stage feature fusion is performed on the optimized first-stage features, the optimized second-stage features, and the spatially accurate feature map. Based on the preliminary fusion features, the feature weights of different scales and layers are calculated to obtain an adaptive feature weight matrix. The adaptive feature weight matrix can automatically adjust the weights of each feature according to the importance of each layer and each scale feature, making the feature fusion more reasonable and effective. Based on the adaptive feature weight matrix, the preliminary fusion features are weighted and combined to obtain multi-scale comprehensive features. The multi-scale comprehensive features are subjected to nonlinear transformation and normalization to obtain multi-scale face features. Nonlinear transformation can process features through activation functions (such as ReLU, Sigmoid, etc.) to make the features more expressive. Normalization can eliminate the dimensional differences between features, making the features comparable at the same scale.

[0034] The low-level feature map of the first stage is upsampled by bilinear interpolation, and the low-level feature map is enlarged to twice the resolution of the original image to obtain the first upsampled feature map. The upsampling process performs bilinear interpolation calculations on each pixel in the image, so that the upsampled feature map can maintain high image details and feature continuity. The first upsampled feature map is transposed convolution operation, the upsampled feature map is further enlarged, and the details and edge information of the feature map are enhanced in the deconvolution process to obtain the preliminary deconvolution feature map. The transposed convolution operation can gradually restore high-level features to the original resolution through back propagation, while effectively retaining the spatial structure and texture information of the image. The preliminary deconvolution feature map is channel-joined with the jump connection features of the original resolution. The jump connection features are high-resolution features extracted from the early layers of the network. Through the channel joint operation, the preliminary deconvolution feature map is combined with these high-resolution features to form a fused feature map. The fused feature map is subjected to a 1x1 convolution operation, and the number of channels in the fused feature map is compressed to a reasonable range through 1x1 convolution, and the first-stage preliminary feature representation after channel compression is obtained. The 1x1 convolution operation can not only reduce the computational complexity, but also recombine the feature channels through linear transformation, thereby improving the compactness and expressiveness of the feature representation. The first supervised attention module calculates the channel attention weights of the first-stage preliminary feature representation to obtain the channel attention matrix. The channel attention weights reflect the importance of each channel to the feature representation. By weighting the feature channels, important features are enhanced and unimportant features are suppressed to obtain channel-enhanced features. The spatial attention weights of the channel-enhanced features are calculated to obtain the spatial attention matrix. By weighting the spatial positions of the feature maps, the feature representations of important areas are enhanced and the features of unimportant areas are suppressed to obtain dual attention features. The dual attention features are residually connected to the first-stage preliminary feature representation to obtain the attention-optimized features. The residual connection operation can retain the original information in the input features by directly adding the input features to the output features, while enhancing the local details and global semantic information of the feature representation through the attention mechanism. Batch normalization and ReLU activation function operations are performed on the attention optimization features. Batch normalization makes the distribution of features in the network more stable, accelerating the network training process; ReLU activation function makes the features have stronger nonlinear expression capabilities, and the optimized first-stage features are obtained.

[0035] The optimized first-stage features are input into the second-stage encoder-decoder structure. This structure usually includes multiple convolutional layers and pooling layers, which can extract deeper features in the encoding stage. Through multi-layer convolution and pooling operations, the optimized first-stage features are processed layer by layer, and high-level semantic information is gradually extracted to obtain the second-stage encoded features. The second-stage encoded features are subjected to dilated convolution. The dilated convolution kernel of the dilated convolution is used to expand the receptive field of the convolution without increasing the number of parameters, so as to capture more contextual information and obtain the dilated receptive field features. The dilated receptive field features are fused with the first-stage skip connection features. Through cross-stage feature fusion, low-level detail information and high-level semantic information are combined to obtain multi-scale semantic features. The multi-scale semantic features are deconvolved. Through deconvolution, the feature map is enlarged to a higher resolution to obtain the second upsampled feature map. The deconvolution operation can not only restore the spatial size of the feature map, but also enhance the fineness of the feature representation by gradually restoring the detail information. The second upsampled feature map is processed by the self-attention mechanism. Through the self-attention mechanism, the feature representation of each position is adaptively adjusted according to the relationship within the feature map to obtain the self-attention enhanced feature. The self-attention enhanced feature is residually connected with the corresponding scale feature in the encoding stage. Through the residual connection, the original information in the feature of the encoding stage is retained. At the same time, the local details and global semantic information of the feature representation are enhanced through the self-attention mechanism to obtain the residual enhanced feature. The channel weight matrix of the residual enhanced feature is calculated through the second supervised attention module. The channel weight matrix reflects the importance of each feature channel. By weighting, important feature channels are highlighted and unimportant feature channels are suppressed to obtain the channel enhanced feature. Spatial attention analysis is performed on the channel enhanced feature. By calculating the weight of each spatial position, the spatial weight matrix is ​​obtained. Spatial attention analysis can adaptively adjust the feature representation of each position according to the spatial distribution in the feature map, thereby highlighting the important regional information in the image. Based on the spatial weight matrix, the channel enhanced feature is weighted to optimize the spatial information in the feature map. The weighted feature maps are batch normalized and nonlinearly activated. Batch normalization makes the distribution of feature maps in the network more stable, reducing the internal covariate shift during training. Nonlinear activation functions (such as ReLU) are used to enhance the nonlinear expression ability of feature maps and obtain optimized second-stage features.

[0036] Step S104: input the multi-scale facial features into a preset face recognition model, iteratively match the data to class similarity, and obtain a candidate identity list of the passed personnel;

[0037] Specifically, the multi-scale facial features are input into a preset face recognition model, and the multi-scale facial features are subjected to dimensionality reduction processing. The high-dimensional feature vector is reduced to a low-dimensional compressed feature vector by methods such as principal component analysis or linear discriminant analysis. An initial similarity threshold is constructed according to a preset initial query class to obtain an initial matching parameter. The initial query class is usually composed of prior knowledge or pre-defined samples. The initial similarity threshold is calculated by the similarity distribution of the initial samples as the starting point of the matching process. A first round of retrieval is performed in the face database based on the compressed feature vector and the initial matching parameter, and a preliminary candidate list is obtained by comparing the similarity between the compressed feature vector and the sample in the database. The samples in the preliminary candidate list are a group of facial features that are most similar to the input features. Through preliminary retrieval, possible candidate identities can be preliminarily screened out. The cosine distance and Mahalanobis distance are calculated for the facial features in the preliminary candidate list and the compressed feature vector to obtain a comprehensive distance matrix. The cosine distance can measure the angle between two vectors, thereby evaluating the similarity between feature vectors; the Mahalanobis distance can consider the covariance of feature distribution and evaluate the distance between samples. The preliminary candidate list is sorted according to the comprehensive distance matrix to obtain an updated candidate list. Top-K samples are selected from the updated candidate list to construct an extended query class to obtain an updated query class. The Top-K selection method is to select the first K most similar samples from the updated candidate list, and construct a new query class based on the features of these samples for use in the next round of retrieval. A new similarity threshold is calculated based on the updated query class to obtain an updated matching parameter. The new similarity threshold is calculated based on the distribution of sample features in the updated query class. Compared with the initial threshold, it is closer to the actual situation and can improve the accuracy of matching. The next round of iterative retrieval is performed according to the updated matching parameters. A new round of sample retrieval is performed in the face database with the updated matching parameters to obtain a new candidate list. The new candidate list contains the most similar samples of the updated query class features in the database, and gradually approaches the real candidate identity through iterative retrieval. Repeat the above steps and gradually optimize the candidate list through multiple rounds of iterative retrieval. Each round of iteration continuously adjusts and optimizes the query class and similarity threshold based on the previous retrieval results and the updated matching parameters to gradually approach the most real candidate identity. When the preset number of iterations is reached, the final candidate list is obtained, and the final candidate list is subjected to adaptive threshold screening. Through the adaptive threshold, the screening criteria are flexibly adjusted according to the actual situation to ensure that the final screened candidate identity list has high accuracy and credibility.

[0038] Step S105: Combine the biometric data collected by the security inspection equipment, perform multimodal fusion and Bayesian reasoning on the candidate identity list, and obtain a comprehensive identity recognition result;

[0039] Specifically, the gait data, body shape data and infrared thermal imaging data collected by the security inspection equipment are preprocessed. The preprocessing steps include denoising, normalization and alignment operations. The denoising process eliminates random noise in the data, and the normalization process scales the data to the same scale range to ensure the comparability of data of different modalities in subsequent processing. The alignment operation enables data of different modalities to be compared and fused on the same time axis. Through the preprocessing step, standardized biometric data is obtained, and features are extracted. Deep learning methods such as convolutional neural networks or recurrent neural networks are used to extract multimodal biometric feature vectors from gait data, body shape data and infrared thermal imaging data. Feature alignment is performed on the multimodal biometric feature vector and the facial features in the candidate identity list. The feature alignment step aligns the multimodal biometric feature vector and the facial features based on time synchronization and spatial alignment to obtain an alignment feature matrix. The alignment feature matrix places the features of different modalities in a unified representation space to ensure consistency in subsequent processing. The aligned feature matrix is ​​modeled for intra-modal relations. Methods such as graph neural networks or self-attention mechanisms are used to establish relations between features within each modality to obtain intra-modal enhanced features. Intra-modal relation modeling can capture the intrinsic connections between features of the same modality and enhance the expressiveness of features. Cross-modal attention analysis is performed on the intra-modal enhanced features. Through the self-attention mechanism, the mutual influence between features of different modalities is analyzed to obtain inter-modal interaction features. Cross-modal attention analysis can effectively fuse information from different modalities, highlight important features, suppress irrelevant features, and obtain more accurate and comprehensive feature representations. Inter-modal interaction features are fused. Through weighted averaging or multi-layer perceptron methods, the interaction features of different modalities are fused together to obtain a fused feature vector. A prior probability distribution is constructed based on historical recognition records and current security level settings. The prior probability of each identity is obtained through statistical analysis and empirical modeling. The identity prior probability reflects the probability distribution of each identity under specific scenarios and conditions. Based on the fusion feature vector and identity prior probability, a Bayesian network model is constructed. Through Bayesian reasoning, the prior probability and the current observation features are combined to obtain the posterior probability distribution of each candidate identity. The maximum a posteriori probability is estimated for the posterior probability distribution, and the most likely identity is determined by solving the maximum value in the posterior probability distribution. The maximum a posteriori probability estimation ensures that the most credible identity is selected based on a comprehensive consideration of prior information and current observation data. The most likely identity is compared with the preset confidence threshold. If the posterior probability exceeds the confidence threshold, the identity is determined as the final comprehensive identity recognition result. Through multimodal fusion and Bayesian reasoning methods, high-accuracy identity recognition is achieved in complex environments.

[0040] Step S106: Control the opening or closing of the security inspection equipment according to the comprehensive identity recognition result.

[0041] Specifically, the comprehensive identity recognition result is combined with the control logic of the security inspection equipment. When the comprehensive identity recognition result is generated, the result is analyzed and verified to ensure its accuracy and reliability. The comprehensive identity recognition result usually includes the most likely identity identification and its corresponding confidence score. This recognition result is compared with the pre-set security inspection authority and security policy to determine whether the currently identified identity has the authority to pass the security inspection. The confidence score in the comprehensive identity recognition result is compared with the preset confidence threshold. The confidence threshold is set by the system administrator according to specific security requirements, usually determined when the system is first deployed, and adjusted according to actual conditions during use. If the confidence score of the recognition result is higher than the confidence threshold, it indicates that the recognition result has a high reliability. At this time, the system will regard the result as a valid identity recognition, trigger the corresponding control logic, and command the security inspection equipment to start. For example, the access control system will open the gate to allow people to pass; the baggage security inspection system will start the conveyor belt to continue to process subsequent baggage; if it is a scene that requires a higher level of inspection, the system will issue a prompt sound or light signal to notify the security inspector to conduct further inspection. If the confidence score of the recognition result is lower than the confidence threshold, the system will consider that the current recognition result is not reliable enough to determine the identity of the person. At this point, the system will trigger another set of control logic to command the security equipment to close. For example, the access control system will keep the gate closed to prevent people from passing; the baggage security system will suspend the operation of the conveyor belt and wait for manual intervention; in scenarios where a higher level of inspection is required, the system will issue an alarm signal to alert security personnel. When implementing control actions, the system considers equipment status and fault handling. For example, when the security equipment is performing an opening or closing action, the system needs to monitor the equipment status in real time to ensure the smooth execution of the action; if the equipment fails during execution, the system needs to issue an alarm in time and start the backup equipment or manual processing process to ensure the continuity and safety of the security process.

[0042] In the embodiment of the present application, the application of the multi-stage repair network improves the accuracy and robustness of feature extraction. Through multi-level feature extraction and optimization, face images in complex environments can be effectively processed and the recognition accuracy can be improved. The data-to-class similarity iterative matching mechanism enhances the flexibility and adaptability of face recognition. The matching strategy can be adjusted dynamically to effectively respond to different recognition scenarios. The introduction of multimodal fusion and Bayesian reasoning improves the reliability of recognition results. By combining multiple biometric data, the system can more comprehensively evaluate identity information and reduce the misrecognition rate. Adaptive Gaussian smoothing and standardization processing improve the quality of the original image. This step can effectively reduce the impact of environmental noise and provide clearer image input for subsequent recognition. The use of the supervised attention module optimizes the feature extraction process. By automatically focusing on important feature areas, the system can more accurately capture identity feature information. Multi-scale cross-stage feature fusion technology enhances the expressiveness of features. Feature information of different scales and levels can be comprehensively utilized to improve the accuracy and robustness of recognition. The dynamic threshold screening mechanism improves the adaptability of the system. By adjusting the recognition threshold according to real-time conditions, the system can maintain high efficiency and accuracy in different security inspection environments.

[0043] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0044] (1) Perform brightness histogram analysis on the original face image to obtain the image brightness distribution characteristics, and perform adaptive gamma correction on the original face image based on the image brightness distribution characteristics to obtain a brightness-balanced image;

[0045] (2) Perform edge detection on the brightness equalized image to obtain an image edge intensity map, and calculate the variance of the local area based on the image edge intensity map to obtain the adaptive Gaussian kernel parameters;

[0046] (3) Perform convolution operation on the brightness equalized image based on the adaptive Gaussian kernel parameters to obtain a smoothed image after noise reduction, and perform geometric transformation and affine transformation on the smoothed image to obtain a corrected face image;

[0047] (4) Locating feature points of the corrected face image according to a preset face key point model to obtain the coordinates of the face key points, and cropping and scaling the corrected face image based on the coordinates of the face key points to obtain a face image of uniform size;

[0048] (5) Perform pixel value normalization on the face image of uniform size to obtain a normalized face image, and perform data enhancement on the normalized face image to obtain a preprocessed face image.

[0049] Specifically, the original face image is analyzed by brightness histogram, which reflects the number of pixels at each gray level in the image. By calculating the histogram, the image brightness distribution features are obtained, which represent the brightness distribution of each area in the image. is the original image, represents its brightness histogram, then The calculation formula is:

[0050] ;

[0051] in, is the Dirac function, and are the width and height of the image, is the image at position The brightness value, is the brightness level. Adaptive gamma correction is performed on the original face image according to the image brightness distribution characteristics to adjust the image contrast and make the image brightness distribution more uniform. The formula for gamma correction is:

[0052] ;

[0053] in, is the rectified image, is the gamma value, which is adjusted adaptively according to the distribution characteristics of the brightness histogram. When the image is dark overall, use a smaller When the image is brighter overall, use a larger value. After obtaining the brightness equalized image, edge detection is performed. Edge detection can use the Sobel operator or the Canny operator to detect the edge by calculating the gradient of the pixel grayscale value in the image. Assume and Respectively represent the gradient of the image in the horizontal and vertical directions, and the image edge intensity The calculation formula is:

[0054] ;

[0055] in, is the image edge intensity map. The variance of the local area is calculated based on the image edge intensity map to obtain the adaptive Gaussian kernel parameters. The variance calculation formula is:

[0056] ;

[0057] in, is the variance, is the first The value of pixels, is the average value of the local area, is the number of pixels in the local area. Adaptive Gaussian kernel parameter Adjust according to the variance. The larger the variance, The larger the value, the better. Based on the adaptive Gaussian kernel parameters, the brightness equalized image is convolved and the Gaussian kernel is used for convolution to achieve the noise reduction effect. The convolution calculation formula is:

[0058] ;

[0059] in, is the smoothed image, is the Gaussian kernel, is the size of the kernel. The denoised smoothed image undergoes geometric transformation and affine transformation to correct the geometric distortion in the image. Geometric transformation and affine transformation can be achieved through matrix operations. The affine transformation matrix is:

[0060] ;

[0061] in, are the rotation and scaling parameters, is the translation parameter. , correct the smoothed image to the standard position to obtain the corrected face image. According to the preset face key point model, locate the feature points of the corrected face image. Assume that the model contains 68 key points, locate these key points through the convolutional neural network (CNN) to obtain the coordinates of the face key points. Assume that the key point coordinates are ,in Based on the coordinates of the facial key points, the corrected face image is cropped and scaled to make the face image uniform in size. Assume that the target size is , crop the face area and scale it to the target size through affine transformation to obtain a face image of uniform size. Normalize the pixel values ​​of the face image of uniform size and scale the pixel values ​​to the range of [0, 1] to eliminate the brightness and contrast differences between different images to obtain a normalized face image. Perform data enhancement on the normalized face image to generate multiple variants through operations such as rotation, translation, scaling, and mirroring to increase the diversity of training data and obtain a preprocessed face image. Data enhancement can effectively improve the robustness and generalization ability of the model, so that it has better performance in practical applications.

[0062] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0063] (1) A multi-stage restoration network is used to perform multi-layer convolution and pooling operations on the pre-processed face image to obtain the first-stage low-level feature map;

[0064] (2) Perform upsampling and deconvolution operations on the low-level feature map of the first stage to obtain the preliminary feature representation of the first stage, and optimize the preliminary feature representation of the first stage through the first supervised attention module to obtain the optimized first stage features;

[0065] (3) Inputting the optimized first-stage features into the second-stage encoder-decoder structure to obtain the second-stage high-level semantic features, and optimizing the second-stage high-level semantic features through the second supervised attention module to obtain the optimized second-stage features;

[0066] (4) Input the optimized second-stage features into the network based on the original input resolution to obtain a spatially accurate feature map;

[0067] (5) Perform multi-scale cross-stage feature fusion on the optimized first-stage features, the optimized second-stage features, and the spatially precise feature map to obtain preliminary fusion features;

[0068] (6) Calculate the feature weights of different scales and levels based on the preliminary fusion features to obtain an adaptive feature weight matrix, and perform weighted combination of the preliminary fusion features based on the adaptive feature weight matrix to obtain multi-scale comprehensive features;

[0069] (7) Perform nonlinear transformation and normalization on the multi-scale comprehensive features to obtain multi-scale facial features.

[0070] Specifically, first, a multi-stage restoration network is used to perform multi-layer convolution and pooling operations on the pre-processed face image to obtain the first-stage low-level feature map. The multi-stage restoration network usually includes multiple convolution layers and pooling layers. Through the step-by-step processing of these layers, rich feature information is extracted from the pre-processed face image. The convolution operation extracts local features in the image by convolving the convolution kernel with the input image. The formula is as follows:

[0071] ;

[0072] in, Indicates After the convolution layer, at position The feature map of It is The weights of the convolution kernels of the layers, It is Layer in position The feature map of It is The bias term of the layer, is an activation function (such as ReLU). The pooling operation downsamples the convolutional feature map to reduce the size of the feature map while retaining important feature information. Upsampling and deconvolution operations are performed based on the low-level feature map of the first stage to obtain the preliminary feature representation of the first stage. The upsampling operation enlarges the size of the feature map through interpolation or other methods, while the deconvolution operation restores the details of the original image through inverse convolution. The formula for deconvolution is as follows:

[0073] ;

[0074] in, Indicates After layer deconvolution at position The feature map of the first stage is similar to the convolution formula. The first supervised attention module is used to optimize the first-stage preliminary feature representation. The attention module calculates the importance of different positions in the feature map, weights the features, highlights important features, and suppresses unimportant features to obtain the optimized first-stage features. The optimized first-stage features are input into the second-stage encoder-decoder structure. In the second stage, more advanced semantic features are extracted through more complex convolution and deconvolution operations. After these advanced semantic features are optimized by the second supervised attention module, the expressive power of the features can be improved to obtain the optimized second-stage features. The optimized second-stage features are input into the network based on the original input resolution. By maintaining the original resolution, the image is refined to obtain a spatially accurate feature map. Multi-scale cross-stage feature fusion is performed on the optimized first-stage features, the optimized second-stage features, and the spatially accurate feature maps. By combining features of different stages and scales, the feature information of each stage and scale is fully utilized to obtain a richer and more comprehensive preliminary fusion feature. The feature weights of different scales and levels are calculated based on the preliminary fusion features to obtain an adaptive feature weight matrix. The adaptive feature weight matrix can automatically adjust the weight of each feature according to the importance of each layer and each scale feature, making the feature fusion more reasonable and effective. Based on the adaptive feature weight matrix, the preliminary fusion features are weighted and combined to obtain multi-scale comprehensive features. The multi-scale comprehensive features are subjected to nonlinear transformation and normalization. Nonlinear transformation can process features through activation functions (such as ReLU, Sigmoid, etc.), making the features more expressive. Normalization can eliminate the dimensional differences between features, making the features comparable at the same scale, and finally obtaining multi-scale face features.

[0075] In a specific embodiment, the execution step performs upsampling and deconvolution operations according to the first stage low-level feature map to obtain the first stage preliminary feature representation, and optimizes the first stage preliminary feature representation through the first supervised attention module to obtain the optimized first stage feature. The process can specifically include the following steps:

[0076] (1) Perform bilinear interpolation upsampling on the low-level feature map of the first stage to obtain a first upsampled feature map, and perform a transposed convolution operation on the first upsampled feature map to obtain a preliminary deconvolution feature map;

[0077] (2) Channel-wise concatenation of the preliminary deconvolution feature map and the jump connection feature of the original resolution is performed to obtain a fused feature map, and a 1x1 convolution operation is performed on the fused feature map to obtain the first-stage preliminary feature representation after channel compression;

[0078] (3) The first supervised attention module calculates the channel attention weights of the preliminary feature representation of the first stage to obtain the channel attention matrix, and the preliminary feature representation of the first stage is weighted based on the channel attention matrix to obtain the channel enhanced features;

[0079] (4) Calculate the spatial attention weight of the channel enhancement feature to obtain the spatial attention matrix, and weight the channel enhancement feature based on the spatial attention matrix to obtain the dual attention feature;

[0080] (5) Perform a residual connection between the dual attention features and the preliminary feature representation of the first stage to obtain the attention optimized features, and perform batch normalization and ReLU activation function operations on the attention optimized features to obtain the optimized first stage features.

[0081] Specifically, the low-level feature map of the first stage is upsampled by bilinear interpolation. Bilinear interpolation is a common image magnification method. It performs linear interpolation calculation on the four neighboring pixels of each pixel to obtain the interpolated pixel value and expand the size of the feature map. Assume that the original feature map is , whose size is , through bilinear interpolation, the size is The first upsampled feature map The calculation formula for bilinear interpolation is:

[0082] ;

[0083] in, is the interpolation weight, Denotes a round-down operation. Perform a transposed convolution operation on the first upsampled feature map. The transposed convolution further expands the spatial size of the feature map by means of inverse convolution while retaining high-resolution detail information. The calculation formula for the transposed convolution is:

[0084] ;

[0085] in, is the initial deconvolution feature map, is the weight of the transposed convolution kernel, is the size of the convolution kernel. The preliminary deconvolution feature map is concatenated with the original resolution skip connection features by channel. The skip connection features are high-resolution features extracted from the early layers of the network. These features contain low-level detail information of the image. By concatenating the preliminary deconvolution feature map with the skip connection features in the channel dimension, the feature information at different levels is fused to obtain the fused feature map. . Perform a 1x1 convolution operation on the fused feature map to reduce the number of channels and obtain the first stage preliminary feature representation after channel compression. 1x1 convolution can effectively compress the number of channels of the feature map while keeping the spatial resolution unchanged by performing a linear transformation at each position. The calculation formula of 1x1 convolution is:

[0086] ;

[0087] in, is the feature map after channel compression, is the weight of the 1x1 convolution kernel, is the number of channels. The first supervised attention module calculates the channel attention weights of the first stage preliminary feature representation. The attention module weights the features by calculating the importance weight of each channel to highlight the important features. The calculation formula of the channel attention weight is:

[0088] ;

[0089] in, is the channel attention weight, is an activation function (such as Sigmoid). The channel-enhanced features are obtained by weighting the preliminary feature representation of the first stage based on the channel attention matrix. The weighted operation is performed by applying the channel attention weight to the feature map of each channel, and the formula is:

[0090] ;

[0091] Calculate the spatial attention weights of the channel enhancement features. Spatial attention weights the features by calculating the importance weights of each position in the feature map to highlight important areas. The calculation formula for spatial attention weights is:

[0092] ;

[0093] in, is the spatial attention weight, is a weight matrix. The channel-enhanced features are weighted based on the spatial attention matrix to obtain dual attention features. The weighted operation is performed by applying the spatial attention weight to the feature map at each position, and the formula is:

[0094] ;

[0095] The channel-enhanced features are weighted based on the spatial attention matrix to obtain dual attention features. The weighted operation applies the spatial attention weight to the feature map at each position, and the formula is:

[0096] ;

[0097] The dual attention features are residually connected to the preliminary feature representation of the first stage to obtain the attention optimized features. The residual connection retains the original information in the input features by adding the input features to the output features, and at the same time enhances the feature representation through the attention mechanism. The formula is:

[0098] ;

[0099] Perform batch normalization and ReLU activation function operations on the attention optimization features to obtain the optimized first-stage features. Batch normalization reduces internal covariate shift and accelerates the training process by normalizing the features. The formula is:

[0100] ;

[0101] in, and Channel The mean and variance of is a small constant to prevent division by zero. The ReLU activation function improves the expressiveness of the model by performing a nonlinear transformation on the normalized features. The formula is:

[0102] ;

[0103] In a specific embodiment, the execution step inputs the optimized first-stage features into the second-stage encoder-decoder structure to obtain the second-stage high-level semantic features, and optimizes the second-stage high-level semantic features through the second supervised attention module. The process of obtaining the optimized second-stage features can specifically include the following steps:

[0104] (1) Input the optimized first-stage features into the second-stage encoder-decoder structure, perform multi-layer convolution and pooling operations on the optimized first-stage features to obtain the second-stage encoding features;

[0105] (2) Perform dilated convolution on the second-stage encoded features to obtain expanded receptive field features, and then fuse the expanded receptive field features with the skip connection features of the first stage to obtain multi-scale semantic features;

[0106] (3) Perform a deconvolution operation on the multi-scale semantic features to obtain a second upsampled feature map, and perform a self-attention mechanism on the second upsampled feature map to obtain a self-attention enhanced feature;

[0107] (4) Perform a residual connection between the self-attention enhanced feature and the corresponding scale feature in the encoding stage to obtain the residual enhanced feature, and calculate the channel weight matrix of the residual enhanced feature through the second supervised attention module;

[0108] (5) Weighting the residual enhancement features based on the channel weight matrix to obtain the channel enhancement features, and performing spatial attention analysis on the channel enhancement features to obtain the spatial weight matrix;

[0109] (6) The channel enhancement features are weighted based on the spatial weight matrix, and batch normalization and nonlinear activation are performed to obtain the optimized second-stage features.

[0110] Specifically, the optimized first-stage features are input into the second-stage encoder-decoder structure. Through multi-layer convolution and pooling operations, the optimized first-stage features are processed layer by layer to extract higher-level semantic information. The convolution operation extracts local features in the image by convolving the convolution kernel with the input feature map. The pooling operation downsamples the convolved feature map to reduce the size of the feature map while retaining important feature information. The calculation formula for the convolution operation is:

[0111] ;

[0112] in, Indicates After the convolution layer, at position The feature map of It is The weights of the convolution kernels of the layers, It is Layer in position The feature map of It is The bias term of the layer, is an activation function (such as ReLU). After multiple layers of convolution and pooling operations, the second-stage encoding features are obtained. The second-stage encoding features are subjected to dilated convolution. Through the dilated convolution kernel of the dilated convolution, the receptive field of the convolution is expanded without increasing the number of parameters, thereby capturing more contextual information and obtaining the dilated receptive field feature. The calculation formula of the dilated convolution is:

[0113] ;

[0114] in, is the feature of expanding the receptive field, is the weight of the convolution kernel, is the dilation rate of the dilated convolution, is the input feature map at position The expanded receptive field feature is fused with the jump connection feature of the first stage, and the multi-scale semantic feature is obtained by fusing the feature information of different levels. The multi-scale semantic feature is deconvolved to expand the spatial size of the feature map through deconvolution while retaining the high-resolution detail information to obtain the second up-sampled feature map. The calculation formula of deconvolution is:

[0115] ;

[0116] in, is the feature map after deconvolution, is the weight of the deconvolution kernel, is the size of the convolution kernel. The second upsampled feature map is processed by the self-attention mechanism. By calculating the importance weights of different positions in the feature map, the features are weighted to obtain the self-attention enhanced features. The self-attention enhanced features are residually connected with the corresponding scale features in the encoding stage to obtain the residual enhanced features. The residual connection retains the original information in the input features by adding the input features to the output features, and at the same time enhances the feature representation through the attention mechanism. The formula for the residual connection is:

[0117] ;

[0118] in, is the residual enhancement feature, is the self-attention enhancement feature, is the corresponding scale feature in the encoding stage. The channel weight matrix of the residual enhanced feature is calculated by the second supervised attention module. The attention module weights the features by calculating the importance weight of each channel to highlight the important features. The calculation formula of the channel attention weight is:

[0119] ;

[0120] in, is the channel attention weight, is the weight matrix, is an activation function (such as Sigmoid). The residual enhanced features are weighted based on the channel weight matrix to obtain channel enhanced features. The weighted operation is performed by applying the channel attention weight to the feature map of each channel, and the formula is:

[0121] ;

[0122] Perform spatial attention analysis on the channel enhancement features, calculate the importance weight of each position in the feature map, and obtain the spatial weight matrix. The calculation formula for spatial attention weight is:

[0123] ;

[0124] in, is the spatial attention weight, is a weight matrix. The channel-enhanced features are weighted based on the spatial weight matrix to obtain dual attention features. The weighted operation applies the spatial attention weight to the feature map at each position, and the formula is:

[0125] ;

[0126] Batch normalization and nonlinear activation are performed on the dual attention features. Batch normalization reduces internal covariate shift and accelerates the training process by normalizing the features. The formula is:

[0127] ;

[0128] in, and Channel The mean and variance of Is a small constant to prevent division by zero. Nonlinear activation uses the ReLU activation function to perform nonlinear transformation on the normalized features. The formula is:

[0129] ;

[0130] The optimized second-stage features are obtained.

[0131] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0132] S1: Input the multi-scale face features into the preset face recognition model, and perform dimensionality reduction processing on the multi-scale face features to obtain a compressed feature vector;

[0133] S2: construct an initial similarity threshold according to the preset initial query class to obtain initial matching parameters;

[0134] S3: Perform the first round of search in the face database based on the compressed feature vector and initial matching parameters to obtain a preliminary candidate list;

[0135] S4: Calculate the cosine distance and Mahalanobis distance between the facial features and the compressed feature vectors in the preliminary candidate list to obtain a comprehensive distance matrix;

[0136] S5: sorting the preliminary candidate list according to the comprehensive distance matrix to obtain an updated candidate list;

[0137] S6: Select the top-K samples from the updated candidate list, construct the extended query class, and obtain the updated query class;

[0138] S7: Calculate a new similarity threshold based on the updated query class to obtain an updated matching parameter;

[0139] S8: Perform the next round of iterative retrieval according to the updated matching parameters to obtain a new candidate list;

[0140] S9: Repeat steps S4 to S8 until a preset number of iterations is reached to obtain a final candidate list, and perform adaptive threshold screening on the final candidate list to obtain a candidate identity list of the passed personnel.

[0141] Specifically, the multi-scale facial features are input into a preset face recognition model, and the multi-scale facial features are subjected to dimensionality reduction processing to reduce the feature dimensions and retain the main information. Dimensionality reduction usually uses methods such as principal component analysis or linear discriminant analysis. Assume that the multi-scale facial features are a matrix , and obtain the compressed feature vector by dimensionality reduction The dimensionality reduction formula of PCA is:

[0142] ;

[0143] in, is the projection matrix containing the principal components, is the feature vector after dimensionality reduction. The initial similarity threshold is constructed according to the preset initial query class to obtain the initial matching parameter. The initial query class is composed of predefined samples, and the feature distribution of these samples is used to calculate the initial similarity threshold. The similarity threshold can be determined by calculating the average similarity and standard deviation between the samples of the initial query class, and the formula is:

[0144] ;

[0145] in, is the initial similarity threshold, is the average similarity between the initial query class samples, is the standard deviation, is an adjustment parameter. The first round of search is performed in the face database based on the compressed feature vector and the initial matching parameters. By comparing the similarity between the compressed feature vector and the samples in the database, a preliminary candidate list is obtained. The similarity calculation usually uses cosine similarity or Euclidean distance. The calculation formula of cosine similarity is:

[0146] ;

[0147] in, and are two eigenvectors, is the norm of the vector. The cosine distance and Mahalanobis distance are calculated for the face features in the preliminary candidate list and the compressed feature vector. The cosine distance is used to measure the angle between two vectors, while the Mahalanobis distance takes into account the covariance of the feature distribution. The calculation formula of the Mahalanobis distance is:

[0148] ;

[0149] in, is the covariance matrix of the feature distribution. The preliminary candidate list is sorted according to the comprehensive distance matrix to obtain an updated candidate list. The comprehensive distance matrix is ​​a weighted combination of cosine distance and Mahalanobis distance, and the formula is:

[0150] ;

[0151] in, is the comprehensive distance matrix, is the cosine distance matrix, and is a weight parameter. Select the top-K samples from the updated candidate list, construct an extended query class, and obtain an updated query class. The top-K selection method is to select the top K most similar samples from the updated candidate list, and the features of these samples are used to update the query class. Calculate a new similarity threshold based on the updated query class to obtain an updated matching parameter. The new similarity threshold is determined by calculating the average similarity and standard deviation between the updated query class samples, and the formula is the same as the initial similarity threshold. Perform the next round of iterative retrieval according to the updated matching parameters, and obtain a new candidate list by comparing the similarity between the compressed feature vector and the sample in the database. Repeat steps S4 to S8 until the preset number of iterations is reached to obtain the final candidate list. After multiple rounds of iterative retrieval and screening, the final candidate list gradually approaches the most realistic candidate identity. The final candidate list is screened by adaptive threshold, and finally a candidate identity list of the person who passed the test is obtained. Adaptive threshold screening adjusts the similarity threshold to flexibly screen the candidate identity according to the actual situation to ensure the accuracy and reliability of the recognition result.

[0152] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0153] (1) Preprocessing the gait data, body shape data and infrared thermal imaging data collected by the security inspection equipment to obtain standardized biometric data, and extracting features from the standardized biometric data to obtain a multimodal biometric vector;

[0154] (2) Align the multimodal biometric feature vector and the facial features in the candidate identity list to obtain an aligned feature matrix, and model the intra-modal relationship of the aligned feature matrix to obtain intra-modal enhanced features;

[0155] (3) Perform cross-modal attention analysis on the intra-modal enhanced features to obtain inter-modal interaction features, and perform feature fusion on the inter-modal interaction features to obtain a fused feature vector;

[0156] (4) Construct a prior probability distribution based on historical identification records and current security level settings to obtain identity prior probability, and construct a Bayesian network model based on the fused feature vector and identity prior probability to obtain the posterior probability distribution;

[0157] (5) Perform maximum a posteriori probability estimation on the posterior probability distribution to obtain the most likely identity, and compare the most likely identity with the preset confidence threshold to obtain a comprehensive identity recognition result.

[0158] Specifically, the gait data, body shape data and infrared thermal imaging data collected by the security inspection equipment are preprocessed. The preprocessing steps include denoising, normalization and alignment operations to ensure the consistency and comparability of data of different modalities in subsequent processing. Denoising can use filtering technology, such as Gaussian filtering or median filtering, to remove random noise in the data. Normalization scales the data to the same scale range. Common methods include minimum-maximum normalization and Z-score normalization. Alignment operations align data of different modalities in the same time axis and spatial coordinate system through time synchronization and spatial alignment methods. Feature extraction is performed on the standardized biometric data. Deep learning methods such as convolutional neural networks or recursive neural networks are used. Recursive neural networks can be used for feature extraction of gait data because gait data has time series characteristics. Convolutional neural networks can be used for feature extraction of body shape data and infrared thermal imaging data because these two types of data have spatial structures. Through feature extraction, a multimodal biometric feature vector is obtained. Feature alignment is performed on the multimodal biometric feature vector and the facial features in the candidate identity list. Feature alignment uses a method based on time synchronization and spatial alignment to align multimodal biometric feature vectors and facial features to obtain an aligned feature matrix. Intra-modal relationship modeling is performed on the aligned feature matrix, and a graph neural network or self-attention mechanism is used to establish the relationship between features within each modality to obtain intra-modal enhanced features. Intra-modal relationship modeling can capture the intrinsic connection between features of the same modality and enhance the expressiveness of features. Cross-modal attention analysis is performed on the intra-modal enhanced features, and the mutual influence between features of different modalities is analyzed using the self-attention mechanism to obtain inter-modal interaction features. Cross-modal attention analysis can effectively fuse information from different modalities, highlight important features, suppress irrelevant features, and obtain more accurate and comprehensive feature representations. Inter-modal interaction features are feature fused, and the interaction features of different modalities are fused together through methods such as weighted averaging or multi-layer perceptron to obtain a fused feature vector. A prior probability distribution is constructed based on historical recognition records and current security level settings. The prior probability distribution reflects the probability distribution of each identity under specific scenarios and conditions. Assume that the prior probability is ,in is the identity variable. Through statistical analysis and empirical modeling, the identity prior probability is obtained. Based on the fusion feature vector and identity prior probability, a Bayesian network model is constructed. Through Bayesian reasoning, the prior probability and the current observation characteristics are combined to obtain the posterior probability distribution of each candidate identity. The Bayesian network model can effectively integrate prior information and observation data to obtain more accurate probability estimates. The formula of Bayes' theorem is:

[0159] ;

[0160] in, is the likelihood function, which indicates the probability of observing the current feature under the condition of a given identity; is the prior probability; is a normalization constant. The maximum a posteriori probability (MAP) estimation is performed on the posterior probability distribution. The most likely identity is determined by solving the maximum value in the posterior probability distribution. The formula for MAP estimation is:

[0161] ;

[0162] The most likely identity is compared with the preset confidence threshold. If the posterior probability exceeds the confidence threshold, the identity is determined to be the final comprehensive identity recognition result.

[0163] The above describes the security inspection equipment control method based on AI recognition in the embodiment of the present application. The following describes the security inspection equipment control device based on AI recognition in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the security inspection equipment control device based on AI recognition includes:

[0164] The shooting module 201 is used to shoot the passing person through the image acquisition terminal on the security inspection device to obtain the original face image;

[0165] The processing module 202 is used to perform adaptive Gaussian smoothing and normalization processing on the original face image to obtain a pre-processed face image;

[0166] Extraction module 203, used to extract features from the pre-processed face image using a multi-stage restoration network to obtain multi-scale face features;

[0167] A matching module 204 is used to input the multi-scale facial features into a preset facial recognition model, and iteratively match the similarity between data and classes to obtain a candidate identity list of the passed personnel;

[0168] The reasoning module 205 is used to combine the biometric data collected by the security inspection equipment, perform multimodal fusion and Bayesian reasoning on the candidate identity list, and obtain a comprehensive identity recognition result;

[0169] The control module 206 is used to control the opening or closing of the security inspection equipment according to the comprehensive identity recognition result.

[0170] Through the synergy of the above components, the application of multi-stage inpainting network improves the accuracy and robustness of feature extraction. Through multi-level feature extraction and optimization, it can effectively process face images in complex environments and improve recognition accuracy. The data-to-class similarity iterative matching mechanism enhances the flexibility and adaptability of face recognition. It can dynamically adjust the matching strategy to effectively cope with different recognition scenarios. The introduction of multimodal fusion and Bayesian reasoning improves the reliability of recognition results. By combining multiple biometric data, the system can more comprehensively evaluate identity information and reduce the false recognition rate. Adaptive Gaussian smoothing and normalization processing improve the quality of the original image. This step can effectively reduce the impact of environmental noise and provide clearer image input for subsequent recognition. The use of supervised attention module optimizes the feature extraction process. By automatically focusing on important feature areas, the system can more accurately capture identity feature information. Multi-scale cross-stage feature fusion technology enhances the expressiveness of features. It can comprehensively utilize feature information of different scales and levels to improve the accuracy and robustness of recognition. The dynamic threshold screening mechanism improves the adaptability of the system. By adjusting the recognition threshold according to real-time conditions, the system can maintain efficiency and accuracy in different security inspection environments.

[0171] The present application also provides a security inspection equipment control device based on AI recognition, and the security inspection equipment control device based on AI recognition includes a memory and a processor, and the memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor executes the steps of the security inspection equipment control method based on AI recognition in the above-mentioned embodiments.

[0172] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the security inspection equipment control method based on AI recognition.

[0173] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0174] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

[0175] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A security inspection equipment control method based on AI recognition, characterized in that: The method comprises: The image acquisition terminal on the security inspection device takes a picture of the person passing through to obtain the original face image; Performing adaptive Gaussian smoothing and normalization processing on the original face image to obtain a preprocessed face image; The method adopts a multi-stage restoration network to extract features from the pre-processed face image to obtain multi-scale face features; specifically includes: adopting a multi-stage restoration network to perform multi-layer convolution and pooling operations on the pre-processed face image to obtain a first-stage low-level feature map; performing upsampling and deconvolution operations on the first-stage low-level feature map to obtain a first-stage preliminary feature representation, and optimizing the first-stage preliminary feature representation through a first supervised attention module to obtain an optimized first-stage feature; inputting the optimized first-stage feature into a second-stage encoder-decoder structure to obtain a second-stage high-level semantic feature, and optimizing the second-stage high-level semantic feature through a second supervised attention module. The optimized features are optimized to obtain optimized second-stage features; the optimized second-stage features are input into a network based on the original input resolution to obtain a spatially accurate feature map; multi-scale cross-stage feature fusion is performed on the optimized first-stage features, the optimized second-stage features and the spatially accurate feature map to obtain a preliminary fusion feature; feature weights of different scales and levels are calculated according to the preliminary fusion feature to obtain an adaptive feature weight matrix, and the preliminary fusion features are weighted combined based on the adaptive feature weight matrix to obtain a multi-scale comprehensive feature; nonlinear transformation and normalization are performed on the multi-scale comprehensive feature to obtain a multi-scale face feature; Inputting the multi-scale facial features into a preset face recognition model, iteratively matching the data to class similarity, and obtaining a candidate identity list of the passing person; Combining the biometric data collected by the security inspection equipment, multimodal fusion and Bayesian reasoning are performed on the candidate identity list to obtain a comprehensive identity recognition result; The security inspection device is controlled to be turned on or off according to the comprehensive identity recognition result.

2. The security inspection equipment control method based on AI recognition according to claim 1 is characterized in that: The step of performing adaptive Gaussian smoothing and normalization processing on the original face image to obtain a preprocessed face image includes: Performing brightness histogram analysis on the original face image to obtain image brightness distribution characteristics, and performing adaptive gamma correction on the original face image according to the image brightness distribution characteristics to obtain a brightness-balanced image; Performing edge detection on the brightness equalized image to obtain an image edge intensity map, and calculating the variance of a local area according to the image edge intensity map to obtain an adaptive Gaussian kernel parameter; Performing a convolution operation on the brightness equalized image based on the adaptive Gaussian kernel parameters to obtain a denoised smoothed image, and performing a geometric transformation and an affine transformation on the smoothed image to obtain a corrected face image; Positioning feature points of the corrected face image according to a preset face key point model to obtain face key point coordinates, and cropping and scaling the corrected face image based on the face key point coordinates to obtain a face image of uniform size; The face image of the uniform size is subjected to pixel value normalization processing to obtain a normalized face image, and the normalized face image is subjected to data enhancement to obtain a preprocessed face image.

3. The security inspection equipment control method based on AI recognition according to claim 1 is characterized in that: The upsampling and deconvolution operations are performed according to the first-stage low-level feature map to obtain the first-stage preliminary feature representation, and the first-stage preliminary feature representation is optimized by the first supervised attention module to obtain the optimized first-stage features, including: Performing bilinear interpolation upsampling on the first-stage low-level feature map to obtain a first upsampled feature map, and performing a transposed convolution operation on the first upsampled feature map to obtain a preliminary deconvolution feature map; Channel-joining the preliminary deconvolution feature map with the jump connection feature of the original resolution to obtain a fused feature map, and performing a 1x1 convolution operation on the fused feature map to obtain a first-stage preliminary feature representation after channel compression; Calculating the channel attention weights of the preliminary feature representation of the first stage through the first supervised attention module to obtain a channel attention matrix, and weighting the preliminary feature representation of the first stage based on the channel attention matrix to obtain channel enhanced features; Calculating the spatial attention weight of the channel enhancement feature to obtain a spatial attention matrix, and weighting the channel enhancement feature based on the spatial attention matrix to obtain a dual attention feature; The dual attention feature is residually connected with the first-stage preliminary feature representation to obtain an attention-optimized feature, and the attention-optimized feature is batch normalized and ReLU activation function operated to obtain an optimized first-stage feature.

4. The security inspection equipment control method based on AI recognition according to claim 3 is characterized in that: The step of inputting the optimized first-stage features into the second-stage encoder-decoder structure to obtain the second-stage high-level semantic features, and optimizing the second-stage high-level semantic features through the second supervised attention module to obtain the optimized second-stage features includes: Inputting the optimized first-stage features into the second-stage encoder-decoder structure, performing multi-layer convolution and pooling operations on the optimized first-stage features to obtain second-stage encoding features; Performing a dilated convolution on the second-stage encoding features to obtain an expanded receptive field feature, and performing feature fusion on the expanded receptive field feature and the skip connection feature of the first stage to obtain a multi-scale semantic feature; Performing a deconvolution operation on the multi-scale semantic features to obtain a second up-sampled feature map, and performing a self-attention mechanism processing on the second up-sampled feature map to obtain a self-attention enhanced feature; Residual connection is performed between the self-attention enhancement feature and the corresponding scale feature in the encoding stage to obtain a residual enhancement feature, and a channel weight matrix of the residual enhancement feature is calculated through a second supervised attention module; The residual enhancement feature is weighted based on the channel weight matrix to obtain a channel enhancement feature, and a spatial attention analysis is performed on the channel enhancement feature to obtain a spatial weight matrix; The channel enhancement features are weighted based on the spatial weight matrix, and batch normalization and nonlinear activation are performed to obtain optimized second-stage features.

5. The security inspection equipment control method based on AI recognition according to claim 4 is characterized in that: The step of inputting the multi-scale facial features into a preset facial recognition model and iteratively matching the similarity between data and classes to obtain a candidate identity list of the passing person includes: S1: inputting the multi-scale facial features into a preset face recognition model, and performing dimensionality reduction processing on the multi-scale facial features to obtain a compressed feature vector; S2: construct an initial similarity threshold according to the preset initial query class to obtain initial matching parameters; S3: performing a first round of search in the face database based on the compressed feature vector and the initial matching parameter to obtain a preliminary candidate list; S4: calculating the cosine distance and the Mahalanobis distance between the facial features in the preliminary candidate list and the compressed feature vector to obtain a comprehensive distance matrix; S5: sorting the preliminary candidate list according to the comprehensive distance matrix to obtain an updated candidate list; S6: Select Top-K samples from the updated candidate list, construct an extended query class, and obtain an updated query class; S7: Calculate a new similarity threshold based on the updated query class to obtain updated matching parameters; S8: performing the next round of iterative search according to the updated matching parameters to obtain a new candidate list; S9: Repeat steps S4 to S8 until a preset number of iterations is reached to obtain a final candidate list, and perform adaptive threshold screening on the final candidate list to obtain a candidate identity list of the passed persons.

6. The security inspection equipment control method based on AI recognition according to claim 5 is characterized in that: The biometric data collected by the security inspection equipment is combined to perform multimodal fusion and Bayesian reasoning on the candidate identity list to obtain a comprehensive identity recognition result, including: Preprocessing the gait data, body shape data and infrared thermal imaging data collected by the security inspection equipment to obtain standardized biometric data, and performing feature extraction on the standardized biometric data to obtain a multimodal biometric vector; Performing feature alignment on the multimodal biometric feature vector and the facial features in the candidate identity list to obtain an alignment feature matrix, and performing intra-modal relationship modeling on the alignment feature matrix to obtain intra-modal enhanced features; Performing cross-modal attention analysis on the intra-modal enhanced features to obtain inter-modal interaction features, and performing feature fusion on the inter-modal interaction features to obtain a fused feature vector; Constructing a priori probability distribution according to historical identification records and current security level settings to obtain identity prior probability, and constructing a Bayesian network model based on the fused feature vector and the identity prior probability to obtain a posterior probability distribution; The maximum a posteriori probability is estimated for the posterior probability distribution to obtain the most likely identity, and the most likely identity is compared with a preset confidence threshold to obtain a comprehensive identity recognition result.

7. A security inspection equipment control device based on AI recognition, characterized in that: Used to execute the security inspection equipment control method based on AI recognition as described in any one of claims 1 to 6, the device comprises: A shooting module is used to shoot a person passing through the security inspection device through an image acquisition terminal to obtain an original face image; A processing module, used for performing adaptive Gaussian smoothing and normalization processing on the original face image to obtain a preprocessed face image; An extraction module, used for extracting features from the pre-processed face image using a multi-stage restoration network to obtain multi-scale face features; A matching module, used for inputting the multi-scale facial features into a preset facial recognition model, iteratively matching the data to the class similarity, and obtaining a candidate identity list of the passing person; A reasoning module, used to combine the biometric data collected by the security inspection equipment, perform multimodal fusion and Bayesian reasoning on the candidate identity list, and obtain a comprehensive identity recognition result; A control module is used to control the opening or closing of the security inspection equipment according to the comprehensive identity recognition result.

8. A security inspection equipment control device based on AI recognition, characterized in that: The security inspection equipment control device based on AI recognition includes: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instruction in the memory so that the AI-recognition-based security inspection equipment control device executes the AI-recognition-based security inspection equipment control method as described in any one of claims 1-6.

9. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instruction is executed by the processor, the security inspection equipment control method based on AI recognition as described in any one of claims 1 to 6 is implemented.

Images