A contraband few sample learning identification method for a security inspection machine

Abstract

The application relates to the field of contraband identification, and discloses a contraband few-sample learning identification method for a security inspection machine, which comprises the following steps: collecting asynchronous multi-angle images of luggage in a scanning area of the security inspection machine, identifying potential existence areas of liquid containers, and discretizing the areas into candidate dot arrays; performing micro-scale space disturbance mapping on the candidate dot arrays, forming an enhanced feature set by simulating few-sample changes; combining adjacent candidate dot features and enhanced features based on adaptive feature units to form potential contraband candidate structures; interactively coupling the potential contraband candidate structures with internal micro-environment information of the luggage, heat distribution features and contour micro-disturbance signals, and accumulating and nonlinearly enhancing signals in local paths; and dynamically adjusting identification model weights and image acquisition frequencies according to the intensity and distribution changes of node signals in a local response enhancement network. The application has the advantage of improving the liquid contraband identification capability.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of contraband identification, in particular to a contraband few-sample learning identification method for security inspection machines. BACKGROUND

[0002] In the existing airport security inspection scene, the detection of liquid contraband has always been a technical problem faced by security inspection machines. Due to the variety of liquid containers, capacity and packaging forms, and the continuous emergence of new liquid contraband, the existing deep learning model based on traditional image recognition or large-scale data training is difficult to accurately identify in the case of limited sample size. Especially for a small number or first-time liquid contraband, the training data is extremely scarce, which leads to false negatives or false positives of security inspection machines, affecting the efficiency of passing and public safety. Liquid in luggage may be wrapped in multiple layers or overlapped with other objects, further increasing the difficulty of image feature extraction. The existing method is difficult to realize the rapid and accurate identification of few-sample liquid contraband while ensuring the detection accuracy. Therefore, it is necessary to design a contraband few-sample learning identification method for security inspection machines to improve the identification ability of liquid contraband. SUMMARY

[0003] In view of the shortcomings of the prior art, the present application provides a contraband few-sample learning identification method for security inspection machines, which has the advantage of improving the identification ability of liquid contraband and solving the problems in the above background.

[0004] To achieve the above purpose of improving the identification ability of liquid contraband, the present application provides the following technical scheme: a contraband few-sample learning identification method for security inspection machines, comprising the following steps:

[0005] Asynchronous multi-angle image acquisition is performed on the luggage in the scanning area of the security inspection machine, and the potential existence area of the liquid container is identified by combining beam refraction interference and scattering attenuation analysis, and the area is discretized into a candidate point array;

[0006] The candidate point array is subjected to micro-scale space disturbance mapping, the liquid refraction mode, container micro-packaging geometric deviation and edge micro-vibration features are extracted, and encoded into adaptive feature units, and an enhanced feature set is formed by simulating few-sample changes;

[0007] Based on the adaptive feature unit, a hierarchical probability fusion strategy is adopted to combine the features of adjacent candidate points and enhanced features to identify liquid abnormal sub-regions, score each liquid abnormal sub-region by local light intensity changes, and screen out regions containing contraband liquids to form potential contraband candidate structures;

[0008] The potential contraband candidate structure is interactively coupled with the luggage internal micro-environment information, heat distribution characteristics and contour micro-perturbation signals, and through the cumulative effect and nonlinear enhancement of the signals in the local path, a local response enhancement network for final judgment is generated;

[0009] According to the intensity and distribution change of each node signal in the local response enhancement network, combined with historical liquid abnormal detection data and real-time optical disturbance mode, the identification model weight and image acquisition frequency are dynamically adjusted.

[0010] Preferably, the process of identifying the potential existence area of the liquid container is to discretize these areas into a candidate point array, which is:

[0011] The security inspection image is subjected to multi-dimensional optical interference analysis, combined with image definition evaluation, and through a layered processing strategy, the potential existence area of the liquid container is identified to generate a candidate area;

[0012] The candidate area is subjected to fine-grained decomposition by using a local image segmentation algorithm, and the edge features and transparency changes of the liquid container are identified and extracted to further delimit the liquid-containing area;

[0013] Based on the gray value distribution and texture features of the image, the boundary of the candidate area is optimized, and the candidate area is matched with known samples to screen the candidate point array and determine the candidate area of the potential liquid container. After multiple screening and optimization, the candidate point array of the liquid container is obtained.

[0014] Preferably, the process of micro-scale space disturbance mapping of the candidate point array is:

[0015] Based on the candidate point array, the fine morphological features of the liquid container are extracted through multi-scale filtering processing;

[0016] Combined with the reflectivity change and local texture density analysis, combined with the local pixel abnormal value, the potential liquid area in the candidate point array is positioned;

[0017] The nonlinear mapping method is applied to the positioned potential liquid area to nonlinearly expand the spatial features of the potential liquid area.

[0018] Preferably, the process of extracting the liquid refraction pattern, container micro-packaging geometric deviation and edge micro-vibration features and encoding them into the adaptive feature unit is:

[0019] Based on the positioned potential liquid area, the liquid refraction features are extracted and the container surface optical changes are subjected to in-depth analysis to obtain the liquid refraction pattern;

[0020] Combined with the container micro-packaging geometric deviation and local pixel vibration response extracted in the micro-scale space disturbance mapping process, the image deformation and micro-dynamic analysis algorithm is used to correct the morphological distortion and vibration interference;

[0021] The extracted optical, geometric and vibration features are adaptively coded, and a high-dimensional feature representation is constructed through a dynamic feature embedding mechanism to generate adaptive feature units.

[0022] Preferably, the process of identifying the liquid abnormal sub-region is:

[0023] Taking the adaptive feature units as the input basis, the feature vectors between adjacent candidate points are analyzed by hierarchical probability fusion to strengthen the correlation between optical, geometric and vibration features.

[0024] In combination with historical detection data, the fused feature set is updated with dynamic weights to construct a self-learning abnormal scoring system, and the feature combination is optimized through multiple iterations.

[0025] In the optimized feature space, a multi-evaluation mechanism is used to identify high-risk regions containing prohibited liquids to generate liquid abnormal sub-regions.

[0026] Preferably, the process of scoring each liquid abnormal sub-region by local light intensity change is:

[0027] The light intensity fluctuation of each liquid abnormal sub-region is analyzed in real time, and the optical response features of the liquid are extracted based on the change of image pixel values.

[0028] A weighted average algorithm is used to combine the local light intensity change and texture fluctuation to generate a light intensity score for each candidate region, which is compared with the optical response features of the liquid.

[0029] Through a multi-screening mechanism, the scoring results are filtered and corrected to eliminate candidate regions that do not meet the liquid features.

[0030] Preferably, the process of forming a potential contraband candidate structure is:

[0031] Based on the light intensity score results, in combination with the spatial distribution and arrangement relationship inside the luggage, candidate regions that may contain contraband are selected;

[0032] A clustering algorithm is used to perform clustering analysis on these candidate regions, and regions with similar features are merged into a single candidate structure.

[0033] According to the feature morphology and position of the candidate structure, the potential contraband candidate structure is determined through further verification.

[0034] Preferably, the process of coupling the potential contraband candidate structure with the internal micro-environment information, thermal distribution features and contour perturbation signals of the luggage is:

[0035] Fusing the geometric information of the potential contraband candidate structure with the temperature and humidity inside the luggage, a comprehensive feature matrix is generated;

[0036] Through the synchronous analysis of thermal induction imaging and optical signals, the potential contraband candidate structure is further confirmed, and the boundary and position of the candidate region are corrected in combination with the perturbation signal;

[0037] The interactive coupling mechanism of local signals is constructed, so that the temperature, humidity and profile perturbation characteristics in the candidate structure are cooperatively processed under multi-source data fusion.

[0038] Preferably, the process of generating a local response enhancement network for final judgment is as follows:

[0039] According to the interactive coupling of local signals, multiple algorithms are used to extract signal features and organize time sequence and spatial relationship;

[0040] The processed signal features are embedded and combined through the network structure to form a multi-channel representation;

[0041] The embedding results of each channel are integrated into a representation structure of network nodes and edges to construct a local response enhancement network.

[0042] Preferably, the process of dynamically adjusting the recognition model weight and image acquisition frequency is as follows:

[0043] Based on the embedding features of each node and edge in the local response enhancement network and the output signal of the network, the reliability and potential deviation of the current recognition result are evaluated;

[0044] According to the feedback information of each node representation and edge weight in the network, the weight distribution of each channel feature and the node related parameters in the recognition model are dynamically adjusted;

[0045] In combination with real-time scanning data and historical liquid contraband detection data, the image acquisition frequency and scanning strategy are optimized.

[0046] Compared with the prior art, the present application provides a contraband few sample learning recognition method for security check machines, which has the following beneficial effects:

[0047] The application realizes high-precision preliminary positioning of the potential existing area of the liquid container by performing asynchronous multi-angle image acquisition in the scanning area of the security inspection machine and combining beam refraction interference and scattering attenuation analysis, so that effective candidate point arrays can also be obtained under the condition of few samples. Through micro-scale space disturbance mapping and extraction of liquid refraction mode, container micro-packaging geometric deviation and edge micro-vibration characteristics, and encoding them into adaptive feature units, the feature expression capability can be enhanced under the condition of limited samples, and the generalization performance of few-sample learning can be significantly improved. The hierarchical probability fusion strategy and light intensity scoring mechanism can effectively identify liquid abnormal sub-regions, distinguish real liquid features from background interference, and realize accurate screening of potential contraband candidate structures. Through the interactive coupling of the micro-environment information, thermal distribution characteristics and contour micro-perturbation signals inside the luggage, and the construction of the local response enhancement network, the synergistic effect of multi-source signals can be fully tapped, so that the spatial features, optical properties and thermal response information of potential contraband can be comprehensively evaluated in a high-dimensional network, and the detection accuracy can be significantly improved. Combined with the node signal intensity, distribution change, historical liquid abnormal detection data and real-time optical disturbance mode, the recognition model weight and image acquisition frequency are dynamically adjusted to realize adaptive optimization of the system, so that the detection process can ensure high recognition rate while taking into account efficiency and resource utilization, especially for low-sample liquid contraband identification in complex luggage environment. The application breaks through the problem of low recognition accuracy and easy interference under the condition of few samples in traditional security inspection liquid identification, and can effectively improve the security inspection efficiency, safety and automation level. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 A schematic diagram of the method of the application is shown. DETAILED DESCRIPTION

[0049] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0050] Embodiment 1: Please refer to Figure 1 A contraband few-sample learning and identification method for a security inspection machine in the embodiments of the application includes the following steps:

[0051] S1: Asynchronous multi-angle image acquisition is performed on the luggage in the scanning area of the security inspection machine, and at the same time, beam refraction interference and scattering attenuation analysis are combined to identify the potential existing area of the liquid container, and these areas are discretized into candidate point arrays.

[0052] The process of identifying the potential existing area of the liquid container in S1 and discretizing these areas into candidate point arrays is as follows:

[0053] The security image is subjected to multi-dimensional optical interference analysis, combined with image definition evaluation, and the potential presence area of the liquid container is identified through a layered processing strategy to generate a candidate area; the light refraction, transmission and scattering characteristics in the visible light, near-infrared and X-ray channels are processed in parallel, the areas where liquid medium may exist are identified by comparing the changes in edge profile, brightness gradient and texture continuity of multi-band images, the optical consistency index (such as blur coefficient, edge sharpness change rate) in the area is calculated by using the image definition evaluation module, and the area with significant definition difference is regarded as the potential liquid container presence area to form an initial candidate area set, and the layered processing process combines the real-time light source parameters of the security equipment and the environmental brightness information to keep the correspondence between the identified area and the physical lighting conditions;

[0054] The candidate area is subjected to fine-grained decomposition by using a local image segmentation algorithm, the edge features and transparency changes of the liquid container are identified and extracted, and the area containing liquid is further delimited; the local image segmentation operation is performed on the obtained candidate area, the region growing algorithm based on gradient enhancement and the edge detection combination strategy are adopted to gradually separate the structural features inside the candidate area, the boundary transition zone between the liquid medium and the non-liquid substance is identified by calculating the brightness fluctuation and transparency change curve of the local pixel neighborhood, the dynamic threshold adjustment mechanism is adopted to correct the edge extraction error in view of the common reflection interference of transparent bottle or plastic container surface in the security image, and the spatial consistency of the transparency change area is judged by using the region connectivity detection method to further determine the spatial range of the liquid container boundary;

[0055] Based on the gray value distribution and texture features of the image, the boundary of the candidate area is optimized, the candidate area is matched with the known samples, the candidate point array is screened, and the candidate area of the potential liquid container is determined, and after multiple screening and optimization, the candidate point array of the liquid container is obtained; the gray histogram of the candidate area is smoothed and standardized, the gray peak distance and variance change rate are calculated to judge the difference in substance density and light absorption rate inside the area, the surface microstructure parameters including direction consistency, energy distribution and contrast index are extracted by using the local texture analysis algorithm (such as gray level co-occurrence matrix method) to judge whether there is regular reflection or liquid ripple signal on the container surface, and the features are compared with the liquid container standard sample feature template stored in the system, after multiple levels of matching, optimization and area overlap analysis, the system outputs the candidate point array of the liquid container, which is composed of multiple spatial coordinate points, each point contains three types of feature parameters of brightness, transparency and texture.

[0056] S2: Micro-scale spatial perturbation mapping is performed on the candidate point array to extract liquid refraction patterns, container micro-packaging geometric deviations, and edge micro-vibration characteristics, and encode them into adaptive feature units to form an enhanced feature set through simulation of few-sample changes.

[0057] The process of micro-scale spatial perturbation mapping on the candidate point array in S2 is as follows:

[0058] Based on the candidate point array, fine morphological features of the liquid container are extracted through multi-scale filtering processing. Multi-scale filtering is performed on the candidate point array obtained through preliminary target detection in the security inspection image, including Gaussian filtering, bilateral filtering, directional guide filtering, and other different scale spatial smoothing and edge preservation processing. Through response calculation of light intensity changes of the candidate point array under different spatial scales, noise interference can be suppressed and the stable features of the liquid container edge and morphological boundary can be enhanced. In the filtering process, a scale response function is established to represent the geometric consistency and local energy distribution characteristics of the candidate region under each scale, so as to extract the fine morphological feature information of the liquid container, including contour curvature, boundary thickness variation, shape transition of the bottle mouth and bottom.

[0059] In combination with reflectivity changes and local texture density analysis, in combination with local pixel outliers, the potential liquid region in the candidate point array is located. The reflectivity changes of each pixel point in the candidate point array are modeled, the light intensity fluctuation curve under different incident angles and imaging conditions is calculated, and the surface uniformity and internal texture complexity of each region are determined in combination with local texture density analysis (such as density estimation based on gray level co-occurrence matrix or local binary pattern LBP). The internal region of the liquid container usually shows high light intensity reflectivity gradient change and low texture density characteristics. By setting an adaptive threshold, pixel clusters with high reflection gradient but low structure texture are detected. Through local outlier analysis of pixels (including light intensity deviation, gradient variance, and abnormal noise point detection), false responses caused by metal, glass edge, and other non-liquid regions are further removed, so as to accurately locate the potential liquid region boundary in the candidate point array.

[0060] The nonlinear mapping method is applied to the positioned potential liquid region, and the spatial features of the potential liquid region are nonlinearly expanded; a nonlinear mapping model is constructed by using a radial basis function or a hyperbolic tangent kernel function, and the original two-dimensional pixel coordinates and light intensity features are mapped to a high-dimensional manifold space, in which the spatial distribution of the liquid region can be smoothly interpolated and the local curvature is enhanced to enlarge the separability of the liquid region in the high-dimensional space. Through principal component analysis or manifold compression processing on the spatial features after nonlinear mapping, more discriminative perturbation mapping results can be generated, so that the liquid region shows high feature consistency and anti-interference ability in the subsequent classification and recognition stage, thereby realizing the micro-scale spatial characterization and enhancement of the features inside the liquid container.

[0061] The process of extracting the liquid refraction pattern, the container micro-packaging geometric deviation, and the edge micro-vibration features in S2 and encoding them into the adaptive feature unit is as follows:

[0062] Based on the positioned potential liquid region, the liquid refraction features are extracted and the optical changes of the container surface are analyzed in depth to obtain the liquid refraction pattern; the liquid region is optically analyzed by using a high-resolution light intensity gradient map and polarized light information to construct a liquid refraction model. By analyzing the light propagation path inside the liquid, the refraction angle change, and the edge spot diffusion characteristics, and combining the refractive index, curvature, and thickness change of the container surface material, the refraction pattern generated by the liquid under different viewing angles is extracted. This process also includes spatial filtering and texture enhancement of optical changes to distinguish the small optical shifts caused by liquid refraction from noise interference, thereby generating a refraction feature matrix reflecting the internal structure of the liquid and the optical characteristics of the container;

[0063] The image deformation and micro-dynamic analysis algorithm is used to correct the morphological distortion and vibration interference in combination with the container micro-packaging geometric deviation and local pixel vibration response extracted in the micro-scale spatial perturbation mapping process; for the possible micro-packaging geometric deviation (such as slight bending of the bottle body and irregular sealing) and local pixel vibration response (such as edge micro-vibration caused by transportation or scanning jitter) of the container in the actual imaging process, the micro-dynamic analysis algorithm is used to compare the continuous frames or multi-view images of the candidate region, detect and quantify the geometric deviation and vibration amplitude, and through the image deformation correction method (such as thin plate spline deformation or affine correction model), the morphological correction of the liquid region and the container edge is performed to minimize the feature shift caused by distortion and vibration. This process ensures that the extracted geometric and vibration features truly reflect the microstructure and dynamic response of the container;

[0064] The extracted optical, geometric and vibration features are adaptively coded, and a high-dimensional feature representation is constructed through a dynamic feature embedding mechanism to generate an adaptive feature unit; the liquid refraction mode, container geometric deviation and edge micro-vibration features are taken as inputs, the feature normalization and weight distribution are performed through the adaptive coding module, different types of features are mapped to a unified high-dimensional representation space, the adaptive coding adopts a dynamic feature embedding mechanism, the features are weighted and fused and nonlinearly mapped in combination with the temporal and spatial context information, the expression ability of key optical, geometric and vibration information is enhanced, and the generated adaptive feature unit can maintain the discriminability of the unique optical, structural and dynamic features of the liquid container in the high-dimensional space, providing a directly usable high-information input for few-shot learning, and realizing fine feature description of the micro liquid container.

[0065] S3: Based on the adaptive feature unit, a hierarchical probability fusion strategy is adopted to combine the features of adjacent candidate points and enhanced features to identify liquid abnormal sub-regions, score each liquid abnormal sub-region based on local light intensity changes, filter out the regions containing prohibited liquids, and form potential prohibited candidate structures.

[0066] The process of identifying liquid abnormal sub-regions in S3 is as follows:

[0067] The adaptive feature unit is taken as an input basis to perform hierarchical probability fusion analysis on the feature vectors between adjacent candidate points, and the correlation between optical, geometric and vibration features is strengthened; the generated adaptive feature unit is taken as an input, and the optical, geometric and vibration feature vectors of adjacent pixels or point array units in the candidate point array are subjected to hierarchical probability fusion processing, including calculating the conditional probability distribution in a multi-layer neighborhood structure, quantifying the similarity and spatial correlation of adjacent feature vectors, and enhancing the correlation between different features layer by layer through a hierarchical fusion algorithm, which can cooperatively map the optical refraction, micro geometric deviation and edge micro-vibration features in space, so that the joint effect of liquid internal abnormal patterns and container micro-disturbances can be fully expressed;

[0068] In combination with historical detection data, the dynamic weight of the fused feature set is updated to construct a self-learning abnormal scoring system, and the feature combination is optimized through multiple iterations; after the feature fusion is completed, historical detection data is introduced, including the abnormal labeling and statistical distribution information of the liquid containers in the scanning samples, the dynamic weight of the fused feature set is adjusted, the weight update adopts a self-learning mechanism, the important features are given higher weights by calculating their discriminability and frequency in the historical abnormal samples, while the low-contribution features are suppressed, the feature combination is continuously adjusted through multiple iterations, so that the fused features can reflect the abnormal signals while taking into account the distinguishability of normal liquids or containers, forming an abnormal scoring system based on the fused features, each candidate sub-region can calculate the corresponding abnormal score to provide a quantitative basis for risk judgment;

[0069] In the optimized feature space, a multi-evaluation mechanism is used to identify high-risk areas containing prohibited liquids to generate liquid anomaly sub-regions; in the optimized high-dimensional feature space, the multi-evaluation mechanism is used for anomaly region identification, including local abnormal distribution detection, neighborhood aggregation evaluation, and global score threshold judgment. The local abnormal distribution detection is used to identify micro-regions deviating from the normal mode. The neighborhood aggregation evaluation enhances the reliability of identifying continuous abnormal sub-regions through spatial context correlation. The global score threshold judgment ensures that only high-risk abnormal regions are marked as liquid anomaly sub-regions. This mechanism can comprehensively consider the joint performance of optical refraction, micro-geometric deviation, and edge micro-vibration features in space and time to accurately locate the sub-regions containing prohibited liquids.

[0070] The process of scoring each liquid anomaly sub-region by local light intensity change in S3 is as follows:

[0071] Real-time analysis of light intensity fluctuations in each liquid anomaly sub-region is performed based on changes in image pixel values to extract optical response features of the liquid. Pixel-level light intensity fluctuation monitoring is performed on each identified liquid anomaly sub-region. The gray value change and gradient change of each pixel are calculated through consecutive frames or multi-view images to obtain local optical response features, including analyzing the brightness deviation, spot movement, and edge reflectivity fluctuation caused by liquid surface refraction. The light intensity dynamic behavior is quantified by combining the statistical features of the pixels in the region (such as mean, variance, skewness, and kurtosis). By tracking and analyzing the optical response of the liquid region in real time, the system can distinguish between liquid features and background noise or light intensity changes caused by non-liquid objects, providing accurate optical input data for light intensity scoring.

[0072] A weighted average algorithm is used to combine local light intensity changes and texture fluctuations to generate light intensity scores for each candidate region and compare them with the optical response features of the liquid. After extracting the optical response features, the light intensity data of the candidate liquid anomaly sub-regions is processed using a weighted average algorithm. The weight settings consider the importance of pixel position, light intensity change amplitude, and local texture consistency. Local texture fluctuations are obtained through gray level co-occurrence matrix or local binary pattern analysis to assist in determining whether the light intensity change is from the internal structure of the liquid or the reflection of the container surface. The calculated light intensity scores are used to quantify the significance of the liquid features of each candidate region and are compared with the optical response features extracted in the previous step to enhance the accuracy of the scores, making high-score regions correspond to regions with obvious liquid features and stable optical response.

[0073] The score results are filtered and corrected through a multiple screening mechanism to eliminate candidate regions that do not meet the liquid characteristics. After obtaining the preliminary light intensity score, a multiple screening mechanism is introduced for result optimization, including: threshold filtering to exclude regions with a light intensity score lower than a preset threshold; spatial consistency check to eliminate isolated abnormal points by analyzing the continuity of adjacent sub-region scores; and time dynamic verification to eliminate false positive regions caused by transient light intensity fluctuations by analyzing the light intensity change trend of consecutive frames. Through the above screening and correction steps, high-confidence abnormal sub-regions matching the liquid characteristics are retained, while false positives caused by optical interference, noise or non-liquid objects are eliminated.

[0074] The process of forming a potential contraband candidate structure in S3 is:

[0075] Based on the light intensity score results, the spatial distribution and arrangement relationship inside the luggage are combined to screen out candidate regions that may contain contraband. The light intensity score results of the liquid abnormal sub-regions are jointly analyzed with the spatial structure information inside the luggage. By mapping the high-score regions to the luggage spatial coordinates in the three-dimensional scan or two-dimensional X-ray image, and combining the arrangement relationship, relative position and density distribution of the items, candidate regions that may contain contraband are screened out. Not only the single high light intensity score region is considered, but also the surrounding neighborhood region is expanded to ensure that the complete region of the potential contraband can be covered, thereby generating a preliminary candidate set with spatial consistency and logical association;

[0076] A clustering algorithm is used to analyze the candidate regions, and regions with similar characteristics are merged into a single candidate structure. Based on the preliminary screening of the candidate regions, a clustering algorithm (such as hierarchical clustering based on feature vector similarity) is applied to analyze the region characteristics. The clustering process combines light intensity, geometric morphology, edge micro-vibration and other multi-dimensional features to merge spatially adjacent and similar candidate regions into a single candidate structure. Through clustering, duplicate segmentation and isolated points are eliminated, the integrity and consistency of the candidate structure are improved, and a more refined potential contraband region is formed.

[0077] According to the characteristics and position of the candidate structure, the potential contraband candidate structure is determined through further verification. The geometric characteristics (such as size, contour, volume ratio), spatial position and relative relationship with the overall layout of the luggage are combined for further verification. The verification methods include rule screening, shape comparison based on template matching and abnormal pattern detection to exclude candidate structures that are similar to regular items but do not meet the characteristics of contraband. The potential contraband candidate structure is determined, and its characteristic information, position and spatial relationship all meet the judgment criteria for contraband identification.

[0078] S4: Interactively coupling the potential contraband candidate structure with the luggage internal micro-environment information, heat distribution characteristics and contour perturbation signals, and generating a local response enhancement network for final judgment through the cumulative effect and nonlinear enhancement of signals in the local path.

[0079] The process of interactively coupling the potential contraband candidate structure with the luggage internal micro-environment information, heat distribution characteristics and contour perturbation signals in S4 is as follows:

[0080] Fusing the geometric information of the potential contraband candidate structure with the temperature and humidity inside the luggage to generate a comprehensive feature matrix; obtaining the geometric information of the aforementioned potential contraband candidate structure, including volume, contour shape and spatial position coordinates, and collecting micro-environment data inside the luggage, such as temperature, humidity and their spatial distribution characteristics, and through the construction of a unified multi-dimensional feature matrix, the geometric information of the candidate structure is correspondingly mapped with the micro-environment parameters, so that each candidate structure can be associated with the temperature and humidity state of the region where it is located. The comprehensive feature matrix not only records the spatial characteristics of the candidate structure, but also contains dynamic change information of the micro-environment;

[0081] Further confirming the potential contraband candidate structure through synchronous analysis of thermal imaging and optical signals, and correcting the boundary and position of the candidate region in combination with the perturbation signal; after generating the comprehensive feature matrix, synchronous analysis of infrared thermal imaging data and optical imaging data is used to jointly evaluate the heat distribution characteristics and optical response of the candidate structure. Thermal imaging can reveal temperature abnormal areas inside or on the surface of the object, while optical signals reflect liquid refraction, edge texture and micro-vibration characteristics. In combination with the perturbation signal (such as local light intensity fluctuation or contour micro-vibration), the boundary of the candidate structure is dynamically corrected, and its spatial position is adjusted, so that the shape and positioning of the potential contraband region are more accurate, and the authenticity and consistency of the candidate structure under multi-source information are ensured;

[0082] Building an interactive coupling mechanism of local signals to enable temperature, humidity and contour perturbation characteristics inside the candidate structure to be collaboratively processed under multi-source data fusion; building a local signal interactive coupling mechanism inside the candidate structure to jointly model temperature, humidity and contour perturbation characteristics in a multi-dimensional space. Through the establishment of signal weight distribution, correlation matrix and collaborative update rules, various characteristics can influence and strengthen each other, while suppressing environmental noise or isolated interference. The interactive coupling mechanism can perform all-around feature analysis on the candidate structure under multi-source data fusion, improve the accuracy of contraband detection, and realize deep collaborative processing of geometry, thermal environment and perturbation signals.

[0083] The process of generating a local response enhancement network for final judgment in S4 is as follows:

[0084] According to the local signals after interaction coupling, multiple algorithms are used to extract signal features and organize time sequence and spatial relationship; the local signals of the aforementioned potential contraband candidate structure after multi-source interaction coupling are obtained, including temperature, humidity, optical response and profile perturbation characteristics, for these signals, multiple algorithms are applied for feature extraction, such as extracting frequency domain features based on Fourier transform or wavelet analysis, extracting spatial structure features based on gradient and texture operator, and extracting dynamic change pattern through time sequence sliding window or recursive difference method, the extracted features not only contain signal amplitude and distribution information, but also organize its time sequence and spatial neighborhood relationship, forming a complete local feature set;

[0085] The processed signal features are embedded and combined through the network structure to form a multi-channel representation; after feature extraction is completed, a network structure is constructed to map the local feature vector to a high-dimensional embedding space, including using graph neural network, convolutional neural network or self-attention mechanism to encode the features, inputting different types of features (optical, thermal environment, perturbation signal) as independent channels into the network, the network performs nonlinear transformation and combination on each channel signal through embedding function, and fuses the time sequence and spatial relationship information to generate a multi-channel feature representation, so that each candidate region has complete and multi-dimensional response representation in the network, realizing the collaborative reinforcement between features;

[0086] The embedding results of each channel are integrated into the representation structure of network nodes and edges to construct a local response enhancement network; the multi-channel embedding results are mapped into network nodes according to the spatial neighborhood relationship and feature correlation of the candidate structure, each node corresponds to a candidate region or a sub-feature unit; the edges between nodes represent feature interaction, spatial proximity or time correlation, through this node and edge representation structure, a complete local response enhancement network is constructed, the network can capture the spatial distribution of local abnormal signals and the collaborative effect of multi-source features on the graph structure, realizing high-precision judgment of potential contraband candidate structure.

[0087] S5: According to the intensity and distribution change of each node signal in the local response enhancement network, combined with historical liquid anomaly detection data and real-time optical disturbance pattern, dynamically adjust the recognition model weight and image acquisition frequency.

[0088] The process of dynamically adjusting the recognition model weight and image acquisition frequency in S5 is as follows:

[0089] Based on the embedding features of each node and edge in the local response enhancement network and the network output signal, the reliability and potential bias of the current recognition result are evaluated; the high-dimensional embedding features of each node and edge in the local response enhancement network are obtained, and the recognition result of the candidate structure is analyzed in combination with the final output signal of the network, the evaluation method includes calculating the variance of the node feature, the embedding vector similarity and the edge weight change trend, and combining the output classification probability and the historical model performance to quantify the reliability of the current recognition result, and by identifying the potential bias such as optical refraction anomaly, thermal environment interference or micro-vibration signal anomaly, the performance stability of the network under different candidate structures and scanning environments can be judged, and feedback basis is provided for dynamic optimization;

[0090] Based on the feedback information of each node representation and edge weight in the network, the weight distribution of each channel feature and the node related parameters in the recognition model are dynamically adjusted; after obtaining the feedback information of the nodes and edges, the system dynamically adjusts the feature weight of different channels in the network according to the contribution of each feature channel in the recognition result, for example, the importance of optical response, thermal distribution and micro-disturbance signal features is weighted and optimized, and the connection weight between nodes is adjusted to strengthen the influence of key features in the network, and the node embedding parameters and activation function response are fine-tuned, so that the network can adapt to the local characteristics of different candidate structures, reduce the misrecognition rate, improve the discrimination accuracy of the model for liquid prohibited articles, and realize real-time adaptive update of the recognition model;

[0091] In combination with real-time scanning data and historical liquid prohibited article detection data, the image acquisition frequency and scanning strategy are optimized; the candidate structure information and local response data obtained in the real-time scanning process are compared and analyzed with the historical liquid prohibited article detection records, the detection difficulty and error distribution of different regions and different types of liquids are evaluated, based on this evaluation, the image acquisition frequency is dynamically adjusted, including increasing the scanning frame rate of key high-risk areas, reducing the acquisition frequency of low-risk areas, and optimizing the scanning path and angle strategy, so as to ensure that high information content areas are fully captured, while reducing the overall scanning redundancy, through this strategy optimization based on real-time and historical data, the balance between scanning efficiency and recognition accuracy is realized, and intelligent and reliable liquid prohibited article detection support is provided for security inspection systems.

[0092] It should be noted that, in this article, relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0093] While embodiments of the application have been shown and described, it is to be understood that the application is not limited to the details of the embodiments described, since numerous changes, modifications, substitutions and variations can be made thereto without departing from the spirit and scope of the application as defined by the appended claims and their equivalents.

Claims

1. A contraband few-sample learning identification method for a security inspection machine, characterized in that, The method comprises the following steps: Asynchronous multi-angle image acquisition is performed on the luggage in the scanning area of the security inspection machine, and potential liquid container regions are identified by combining beam refraction interference and scattering attenuation analysis, and the regions are discretized into a candidate point array; Micro-scale spatial disturbance mapping is performed on the candidate point array to extract liquid refraction patterns, container micro-packaging geometric deviations, and edge micro-vibration features, which are coded into adaptive feature units, and enhanced feature sets are formed by simulating a few sample changes; The process of micro-scale spatial disturbance mapping on the candidate point array is as follows: Based on the candidate point array, fine morphological features of the liquid container are extracted through multi-scale filtering processing; Potential liquid regions in the candidate point array are located by combining reflectivity changes and local texture density analysis, and combining local pixel outliers; A nonlinear mapping method is applied to the located potential liquid regions to nonlinearly expand the spatial features of the potential liquid regions; Based on the adaptive feature units, a hierarchical probability fusion strategy is adopted to combine adjacent candidate point features and enhanced features to identify liquid abnormal sub-regions, score each liquid abnormal sub-region based on local light intensity changes, filter out regions containing prohibited liquids, and form potential prohibited item candidate structures; The process of the hierarchical probability fusion strategy is as follows: The optical, geometric, and vibration feature vectors of adjacent pixels or point array units in the candidate point array are subjected to hierarchical probability fusion processing, including calculating conditional probability distribution in a multi-layer neighborhood structure, quantifying the similarity and spatial correlation of adjacent feature vectors, and enhancing the correlation between different features through hierarchical fusion algorithm layer by layer, which can cooperatively map the optical refraction, micro-geometric deviation, and edge micro-vibration features in space; The potential prohibited item candidate structures are interactively coupled with the internal micro-environment information, thermal distribution features, and contour micro-perturbation signals of the luggage, and a local response enhancement network for final judgment is generated through the cumulative effect and nonlinear enhancement of signals in the local path; According to the intensity and distribution changes of the signals of each node in the local response enhancement network, the historical liquid abnormal detection data and real-time optical disturbance patterns are combined to dynamically adjust the recognition model weight and image acquisition frequency. 2.The contraband few-sample learning identification method for a security inspection machine according to claim 1, characterized in that, The process of identifying potential liquid container regions and discretizing the regions into a candidate point array is as follows: Multi-dimensional optical interference analysis is performed on the security inspection image, and image sharpness evaluation is combined to identify potential liquid container regions through a hierarchical processing strategy to generate candidate regions; A local image segmentation algorithm is used to finely decompose the candidate regions to identify and extract the edge features and transparency changes of the liquid containers, and further define the liquid-containing regions; Based on the gray value distribution and texture features of the image, the boundaries of the candidate regions are optimized, the candidate regions are matched with known samples, the candidate point array is screened, and the candidate regions of the potential liquid containers are determined. After multiple layers of screening and optimization, the candidate point array of the liquid container is obtained.

3. The method according to claim 1, wherein The process of extracting liquid refraction patterns, container micro-packaging geometric deviations, and edge micro-vibration features and coding them into adaptive feature units is as follows: Based on the positioning of the potential liquid area, the liquid refraction characteristics are extracted and the optical changes of the container surface are deeply analyzed to obtain the liquid refraction pattern; Combined with the container micro-packaging geometric deviation and local pixel vibration response extracted in the micro-scale space disturbance mapping process, image deformation and micro-dynamic analysis algorithm is used to correct the morphological distortion and vibration interference; The extracted optical, geometric and vibration characteristics are adaptively coded, and a high-dimensional feature representation is constructed through a dynamic feature embedding mechanism to generate adaptive feature units.

4. The method according to claim 3, wherein, The process of identifying liquid abnormal sub-regions is as follows: Taking the adaptive feature unit as the input basis, the feature vectors between adjacent candidate points are analyzed by hierarchical probability fusion, and the correlation between optical, geometric and vibration features is strengthened; Combined with historical detection data, the dynamic weight of the fused feature set is updated to construct a self-learning abnormal scoring system, and the feature combination is optimized through multiple iterations; In the optimized feature space, a multi-evaluation mechanism is used to identify high-risk areas containing prohibited liquids and generate liquid abnormal sub-regions.

5. The method according to claim 4, wherein, The process of scoring each liquid abnormal sub-region by local light intensity change is as follows: Real-time analysis of the light intensity fluctuation of each liquid abnormal sub-region is performed, and the optical response characteristics of the liquid are extracted based on the change of image pixel value; Using a weighted average algorithm, the local light intensity change and texture fluctuation are combined to generate a light intensity score for each candidate region, which is compared with the optical response characteristics of the liquid; Through a multi-screening mechanism, the scoring results are filtered and corrected to eliminate candidate regions that do not meet the liquid characteristics.

6. The method according to claim 5, wherein The process of forming potential prohibited item candidate structures is as follows: Based on the light intensity score results, combined with the spatial distribution and arrangement relationship inside the luggage, candidate regions that may contain prohibited items are selected; A clustering algorithm is used to cluster and analyze these candidate regions, and regions with similar features are merged into a single candidate structure; According to the feature morphology and position of the candidate structure, further verification is performed to determine the potential prohibited item candidate structure.

7. The method according to claim 6, wherein, The process of coupling the potential prohibited item candidate structure with the internal micro-environment information, thermal distribution characteristics and contour perturbation signals of the luggage is as follows: Fuse the geometric information of the potential prohibited item candidate structure with the temperature and humidity inside the luggage to generate a comprehensive feature matrix; Through synchronous analysis of thermal imaging and optical signals, the potential prohibited item candidate structure is further confirmed, and the boundary and position of the candidate region are corrected combined with the perturbation signal; Construct an interactive coupling mechanism for local signals to enable temperature, humidity and contour perturbation characteristics in the candidate structure to be processed cooperatively under multi-source data fusion.

8. The method according to claim 7, wherein The process of generating a local response enhancement network for final judgment is as follows: According to the interactive coupling of local signals, multiple algorithms are used to extract signal features and organize temporal and spatial relationships; Through network structure, the processed signal features are embedded and combined to form a multi-channel representation; Integrate each channel embedding result into a network node and edge representation structure to construct a local response enhancement network.

9. The method according to claim 8, wherein, The process of dynamically adjusting the recognition model weight and image acquisition frequency is as follows: Based on the embedding features of nodes and edges in the local response enhancement network and the output signal of the network, the reliability and potential bias of the current recognition result are evaluated; According to the feedback information of the representation of each node and the weight of each edge in the network, the weight distribution of each channel feature and the node-related parameters in the recognition model are dynamically adjusted; Combined with real-time scanning data and historical liquid prohibited item detection data, the image acquisition frequency and scanning strategy are optimized.

Images