A method and system for fast non-invasive screening of smoking packages based on millimeter wave radar imaging

Abstract

The present application belongs to the technical field of tobacco detection, and particularly relates to a kind of fast non-invasive screening method of tobacco-related package based on millimeter wave radar imaging;The stacking structure topology reasoning module provided by the present application introduces a dynamic edge update mechanism guided by physics, solves the feature breakage problem caused by occlusion and chaotic stacking when traditional static graph convolution is used to process millimeter wave tobacco screening, and the dynamic edge weight prediction subnetwork can remove false connections that do not conform to the physical properties of the cigarette according to physical prior information in real time;The embedding of physical constraints significantly reduces the false positive rate of the model, and through the hard constraints of non-penetrability and coplanarity, the model eliminates the low-level error of misjudging different objects in front and back as single-layer stacked cigarettes;At the same time, due to the introduction of inductive bias based on physical laws, the stacking structure topology reasoning module does not need to rely on massive data for pure statistical learning, and can grasp the essential law of cigarette stacking with only a small amount of labeled samples.

Description

Technical Field

[0001] This invention belongs to the field of tobacco detection technology, specifically relating to a method and system for rapid, non-invasive screening of tobacco-related packages by relying on millimeter-wave radar imaging data and combining artificial intelligence image recognition technology. Background Technology

[0002] While existing millimeter-wave radar screening solutions have solved the portability and radiation issues, their accompanying AI recognition models mostly employ general target detection architectures. The standard YOLO, SSD, and other general target detector AI recognition models used in existing millimeter-wave screening systems suffer from high false positive rates and still require unpacking and verification. These general models typically project millimeter-wave 3D images directly into a single 2D grayscale image, discarding crucial depth dimension information. This results in an inability to distinguish between a single layer of metal foil tightly adhering to the inner wall of the package and the strong reflections of multiple stacked cigarette packs, leading to a persistently high false positive rate. Standard YOLO, SSD, and other general target detectors excel at extracting texture and color features, but millimeter-wave images are colorless, odorless, and lack rich texture. The core characteristics of cigarette packs lie in their parallelepiped geometry and the topological regularity of multiple packs stacked periodically. Existing models lack the ability to reason about this high-order geometry, making them prone to missed detection when cigarette packs are entangled in clothing or placed at an angle.

[0003] Meanwhile, due to the sparse and noisy nature of millimeter-wave images, millimeter-wave imaging usually presents a sparse point distribution and is easily interfered with by highly reflective debris such as metal zippers and buttons, resulting in a low detection rate when faced with severely squeezed or partially damaged cigarette packs. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a rapid, non-invasive screening method and system for smoke-contaminated packages based on millimeter-wave radar imaging, which solves the problems existing in the prior art.

[0005] This invention provides a rapid, non-invasive screening method for smoke-contaminated packages based on millimeter-wave radar imaging, the method comprising the following steps:

[0006] S1: Acquire millimeter-wave radar 3D imaging data of the package to be inspected; S2: Perform data preprocessing and multi-view projection on the three-dimensional imaging data to form a feature map; S3: Input the feature map into the adaptive feature enhancement recognition model that integrates prior knowledge of stacked morphology to realize cigarette feature recognition; the overall architecture of the adaptive feature enhancement recognition model that integrates prior knowledge of stacked morphology consists of four parts: backbone feature extraction network, deep hierarchical attention module, stacked structure topology inference module and detection head; Specifically, S3 is: S3.1: Multi-scale feature map extraction is performed on the feature map using a backbone feature extraction network; S3.2: Input the multi-scale feature map into the deep hierarchical attention module to perform deep dimension feature weighting to form a fused feature map; S3.3: Input the fused feature map into the stacked structure topology reasoning module to model spatial topology relationships and generate a structured feature map; S4: Input the structured feature map into the detection head and output the screening result.

[0007] Preferably, in step S3.1, the backbone feature extraction network uses the CSPDarknet structure as the backbone network, the feature map is copied twice, one copy is processed by dense convolutional blocks, and the other copy skips the connection directly, and finally the two copies are spliced ​​at the output end. A channel compression convolutional layer is added at the beginning of the backbone feature extraction network. This layer uses a 1×1 image convolution kernel to compress the 1536-dimensional channels to 64 dimensions. After the channel compression convolutional layer outputs the compressed feature map, the Stem layer of the CSPDarknet structure receives the compressed feature map and performs preliminary downsampling and feature abstraction through two 3×3 convolutional layers. Then, it passes through four CSP modules in sequence. Each CSP module contains a residual convolutional block and a cross-stage local fusion structure. Finally, it outputs three feature maps of different scales, namely the large-size feature map F. P3 Mesoscale feature map F P4 and small-scale feature map F P5 .

[0008] Preferably, in step S3.2, the deep hierarchical attention module receives the feature map output by the backbone feature extraction network, first restores the channel dimension to a depth layer, and then performs inter-layer correlation modeling, that is, introduces a deep sequence modeling sub-network using a Transformer image encoder structure. The deep sequence modeling sub-network treats the depth layer as a time series, and then performs self-attention calculation on the time series to capture the self-attention score between any two depth layers. Finally, a physical bias term based on a Gaussian distribution is added to the self-attention score. Then, attention mask generation and application are performed, outputting attention weights for 512 depth layers. The attention weights are extended to match the shape of the multi-scale feature map and multiplied element-wise with the multi-scale feature map. Then, upsampling and stitching are performed to form a fused feature map.

[0009] Preferably, the expression for the Gaussian-based physical bias term is: ; In the formula, The modified attention weights are defined by adding spatial bias, where λ is the bias intensity coefficient, i and j represent the spatial indices of the nodes, and σ represents the spatial scale parameter.

[0010] Preferably, S3.3 specifically includes: Dynamic graph construction based on physical priors; the fused feature map is divided into an S×S grid to generate an initial node set V0, where each node in the initial node set V0 contains a position coordinate image and a feature vector f. i ; Physically constrained edge candidate set generation: Construct a physically guided candidate edge set. For each node, only if candidate node v j A potential connection is established only if the following conditions are met: Non-penetrating constraint: Based on the depth information provided by the depth layer attention module, if the difference between the depth values ​​of two nodes is greater than the difference threshold, they are determined to be objects in the preceding and following layers, and no edge is established. Coplanarity constraint: Calculate the angle between the normal vectors of the planes containing the two nodes. If the angle is less than the angle threshold, they are determined to be coplanar, and edges are established first. Minimum Spacing Constraint: There are usually tiny gaps (2-5mm) between cigarette packs. If the distance between two nodes is less than the preset threshold for the thickness error of the cigarette pack, they are determined to be different sampling points of the same object, and no edge is established to prevent a single cigarette pack from being identified as multiple. Dynamic edge weight update introduces an edge weight prediction subnetwork, which uses the feature vector f of each node. i As input, the reliability score of the predicted edge is predicted in real time. During the training process, the edge weight prediction subnetwork learns that parallel edges with uniform spacing have high weights and non-parallel edges with non-uniform spacing have low weights. During the inference phase, the adjacency matrix is ​​dynamically adjusted at the same time, and finally, multi-granularity convolution is performed on each node according to the dynamic edge weight. Topological confidence field generation and feature reprojection: After multi-granularity convolution, each node is endowed with rich hierarchical semantic information. Through an MLP layer, a topological confidence is output for each node. Then, graph feature reprojection is performed to map the learning results of the graph space back to the image space. Using bilinear interpolation, the feature vector and topological confidence of each node are mapped back to the feature map generated in step S2 to generate a structured feature map.

[0011] Preferably, in step S4, the detection head is a Path Aggregation Network.

[0012] Preferably, in step S1, a portable millimeter-wave radar scanning device is used to acquire millimeter-wave radar three-dimensional imaging data of the package to be inspected; the portable millimeter-wave radar scanning device includes a radio frequency front-end, an antenna array, and a motion mechanism, wherein the radio frequency front-end adopts frequency-modulated continuous wave, operates in the frequency band of 77-81GHz, has a sweep bandwidth of 4GHz, and achieves a distance resolution of 3.75cm.

[0013] Preferably, the scanning mode is selected according to the size of the package; for small packages, i.e. packages with a size of <30cm×30cm×30cm, the portable millimeter-wave radar scanning device starts near-field static scanning, the radar array is close to the surface of the package, and the entire package area is covered by electronic beam scanning, with a scanning time of ≤5 seconds.

[0014] Preferably, in step S2, the preprocessing includes adaptive noise filtering based on density clustering and region of interest clipping based on wrapping contours.

[0015] According to another aspect of the present invention, a rapid non-invasive screening system for smoke-contaminated packages based on millimeter-wave radar imaging is provided. The system employs the aforementioned rapid non-invasive screening method for smoke-contaminated packages based on millimeter-wave radar imaging. The system includes: The data acquisition module is used to acquire millimeter-wave radar three-dimensional imaging data of the package to be inspected; The feature extraction module is used to perform data preprocessing and multi-view projection on the three-dimensional imaging data to form a feature map; The recognition module is used to input the feature map into an adaptive feature enhancement recognition model that integrates prior knowledge of stacking morphology to achieve cigarette feature recognition; The result output module is used to input the structured feature map into the detection head and output the screening result.

[0016] The embodiments of the present invention have the following technical effects: The stacked structure topology inference module provided by this invention solves the feature breakage problem caused by occlusion and stacking disorder in traditional static graph convolution when processing millimeter-wave smoke-related screening by introducing a physically guided dynamic edge update mechanism. The dynamic edge weight prediction subnetwork can eliminate erroneous connections that do not conform to the physical properties of cigarette packs in real time based on physical priors. The embedding of physical constraints significantly reduces the false alarm rate of the model. By hard-constraining non-penetration and coplanarity, the model eliminates the low-level error of misjudging two different layers of items as a single-layer stack of cigarette packs. At the same time, due to the introduction of inductive bias based on physical laws, the stacked structure topology inference module does not need to rely on massive amounts of data for pure statistical learning. It can grasp the essential rules of cigarette pack stacking with only a small number of labeled samples, which is particularly important in the actual law enforcement environment where samples of smoke-related cases are scarce.

[0017] In the Gaussian distribution-based physical bias term provided by this invention, the introduction of the Gaussian kernel function enables spatially adjacent nodes to be distinguished even if their semantic feature similarity is low due to occlusion, through a very small spatial distance image (i). j) The image triggers a high-intensity bias bonus, thereby prioritizing the enhancement of the continuous contour lines of the cigarette packs during the feature aggregation stage. This effectively overcomes the shortcomings of traditional attention mechanisms that rely solely on semantic similarity and are prone to incorrectly severing broken edges. Simultaneously, it exhibits extremely high adaptability in suppressing background noise. For isolated high-reflectivity artifacts, due to the lack of spatially adjacent effective nodes, the Gaussian term imposes an exponentially decaying penalty, ensuring they are almost entirely excluded from the effective topological structure in subsequent graph convolutions, significantly reducing the false alarm rate. Secondly, this function greatly enhances the model's implicit perception of the physical arrangement of cigarette packs. As regular rigid bodies, the stacking of cigarette packs inevitably follows a strict linear or planar topology. By adjusting the image scale parameter σ, this bias function can adaptively simulate the width and spacing of the cigarette packs, forcing the model to prioritize message passing along the long axis of the cigarette packs during inference. This spatial guidance is equivalent to injecting a rule of physical common sense into the model, enabling it to reconstruct the complete stacking logic based on spatial proximity even when faced with severely compressed or partially damaged cigarette packs. This represents a leap from viewing pixel features to understanding physical structures. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a rapid, non-invasive screening method for smoke-contaminated packages based on millimeter-wave radar imaging, provided by an embodiment of the present invention. Figure 2 This is a schematic diagram of the X-view projection YZ plane, Y-view projection XZ plane, and Z-view projection XY plane of the radar three-dimensional imaging data provided in the embodiments of the present invention; Figure 3 This is a flowchart provided by an embodiment of the present invention, which describes how to input the feature map into an adaptive feature enhancement recognition model that integrates prior knowledge of stacking morphology to achieve cigarette feature recognition; Figure 4 This is a comparison chart of the performance of the DSATR-Net model provided in this embodiment of the invention compared to traditional algorithms. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0021] Example 1, Figure 1 A flowchart of a rapid, non-invasive screening method for smoke-contaminated packages based on millimeter-wave radar imaging is shown, as follows: Figure 1 As shown, a rapid, non-invasive screening method for smoke-contaminated packages based on millimeter-wave radar imaging includes the following steps: S1: Acquire millimeter-wave radar 3D imaging data of the package to be inspected; In this step, a portable millimeter-wave radar scanning device is used to acquire millimeter-wave radar three-dimensional imaging data of the package to be inspected. The portable millimeter-wave radar scanning device includes a radio frequency front-end, an antenna array, and a motion mechanism. The radio frequency front-end uses frequency-modulated continuous wave, operates in the 77-81GHz frequency band, has a sweep bandwidth of 4GHz, and achieves a distance resolution of 3.75cm.

[0022] The antenna array adopts a MIMO array design, configured with 8 transmit and 16 receive channels, totaling 128 virtual channels, with a horizontal azimuth resolution of 1.5° and a vertical elevation resolution of 3°. The array has a compact size, which is suitable for the miniaturization requirements of handheld devices.

[0023] The motion mechanism includes a built-in miniature inertial measurement unit and encoder, which are used to record the spatial pose changes of the device in real time during the scanning process, providing motion compensation parameters for subsequent synthetic aperture imaging.

[0024] Furthermore, the scanning mode is selected based on the package size; for small packages, i.e., packages with a size of <30cm×30cm×30cm, the portable millimeter-wave radar scanning device activates near-field static scanning, with the radar array closely attached to the surface of the package, and the entire package area is covered by electronic beam scanning, with a scanning time of ≤5 seconds.

[0025] For large packages, i.e., packages ≥30cm×30cm×30cm, the portable millimeter-wave radar scanning device activates dynamic synthetic aperture scanning. Law enforcement officers hold the device and move it in a uniform zigzag pattern along the surface of the package. The scanning path covers the front and sides of the package. The scanning speed is controlled at 5-10cm / s to ensure the continuity of the echo signal, and the scanning time is ≤10 seconds.

[0026] Meanwhile, to reduce the subsequent computational burden, the portable millimeter-wave radar scanning device also integrates a real-time data preprocessing module to preprocess the raw echo data in real time and generate three-dimensional imaging data. Specifically, it performs a fast Fourier transform on the IF signal of each chirp to extract the target's range information and generate a range-amplitude spectrum; it performs a Fourier transform on multiple frames of data in the same range cell to extract the target's radial velocity information; it estimates the target's azimuth and elevation angles using a beamforming algorithm based on the phase difference of the MIMO array to determine the target's two-dimensional angle information; and it converts the range-amplitude spectrum and two-dimensional angle information into three-dimensional coordinates (x, y, z) and combines them with the reflection intensity to form three-dimensional imaging data.

[0027] S2: Perform data preprocessing and multi-view projection on the three-dimensional imaging data to form a feature map; The preprocessing includes adaptive noise filtering based on density clustering and region of interest clipping based on wrapping contours.

[0028] Millimeter-wave radar is susceptible to multipath reflection and environmental electromagnetic interference, resulting in a large number of discrete noise points in the original 3D image data. In this step, an adaptive DBSCAN algorithm is used for noise filtering. The specific process is as follows: First, the 3D imaging data is calculated in x... The average density ρ in the y-plane xy The search radius for clustering is determined based on the average density. ,Right now =k / ρ xy k=0.8 is an empirical coefficient; then, taking each point as the center, The system searches for neighboring points within the search radius. If the number of neighboring points is greater than or equal to a preset threshold, the point is marked as a core point, and the core point and its neighboring points are merged to form a cluster. This process is repeated until all core points have been visited. Finally, only the top 3 clusters with the most points are retained, and the remaining smaller clusters are removed.

[0029] To reduce subsequent computational load and focus on the internal region of the package, a ROI cropping operation needs to be performed on the 3D imaging data. Specifically, the Alpha Image Shapes algorithm is used to extract the 3D imaging data in x... The convex hull profile in the y-plane is used to determine the bounding rectangular area of ​​the package; based on common package heights, the effective range of the z-axis is set, and ground and overhead interference are removed to generate the cropped ROI region data.

[0030] To comprehensively capture the three-dimensional geometric features of the cigarette pack, the preprocessed three-dimensional imaging data was projected from three orthogonal perspectives (x, y, and z) to generate two-dimensional grayscale images. Specifically, the x-view projection was a projection along the x-axis, mapping points in the preprocessed three-dimensional imaging data to the y-axis. Generate a grayscale image F with size H×W in the z-plane. d,x The y-view projection is a projection along the y-axis, mapping points in the preprocessed 3D imaging data to the x-axis. The z-plane generates a grayscale image F. d,y The z-view projection is a projection along the z-axis, mapping points in the preprocessed 3D imaging data to the x-axis. Generate grayscale image F in the y-plane. d,z .

[0031] like Figure 2 As shown, Figure 2 The image shows the X-view projection (YZ plane), Y-view projection (XZ plane), and Z-view projection (XY plane) of the three-dimensional image data, intuitively demonstrating the original distribution of the millimeter-wave image in three-dimensional space. In the left X-view image, the red dot clusters are densely distributed in three segments along the Y-axis, corresponding to the stacked outline of the cigarette in the depth direction. In the middle Y-view image, the blue dot cloud appears as a wide diffuse pattern in the XZ plane, reflecting the depth scattering characteristics of millimeter waves after penetrating clothing. The right Z-view image uses red, green, and orange to distinguish the projections of the three cigarette clusters on the horizontal plane, clearly showing the linear arrangement of the cigarette in the XY plane.

[0032] Finally, all grayscale images are subjected to resolution unification and channel stitching operations to form a feature map.

[0033] S3: Input the feature map into an adaptive feature enhancement recognition model that integrates prior knowledge of stacking morphology to achieve cigarette feature recognition; Existing AI recognition models for millimeter-wave screening systems, such as standard YOLO and SSD general-purpose target detectors, suffer from high false positive rates and still require unpacking and verification. These general-purpose models typically project millimeter-wave 3D images directly into a single 2D grayscale image, discarding crucial depth dimension information. This makes it impossible to distinguish between a single layer of metal foil tightly adhering to the inner wall of the package and the strong reflections of multiple stacked packs of cigarettes, resulting in a persistently high false positive rate. While standard YOLO and SSD general-purpose target detectors excel at extracting texture and color features, millimeter-wave images are colorless, odorless, and lack rich texture. The core characteristics of cigarette packs lie in their parallelepiped geometry and the topological regularity of multiple packs stacked periodically. Existing models lack the ability to reason about this high-order geometry, making them highly susceptible to missed detection when cigarette packs are entangled in clothing or placed at an angle.

[0034] Based on this, this embodiment constructs an adaptive feature enhancement recognition model that integrates prior knowledge of stacking morphology. Through a deep hierarchical attention mechanism, the model can see through the package and accurately lock the cigarette signal in continuous depth layers, reducing the false alarm rate of metal camouflage to below 1%. Through topological reasoning of stacked structures, the model can identify partially obscured or randomly stacked cigarettes, and the recognition accuracy of tilted and staggered stacking morphology is improved to over 99%.

[0035] Specifically, the overall architecture of the adaptive feature enhancement recognition model that integrates prior knowledge of stacked morphology consists of four parts: a backbone feature extraction network, a deep hierarchical attention module, a stacked structure topology inference module, and a detection head. The backbone feature extraction network uses CSPDarknet to extract multi-scale feature maps from the feature maps. The deep hierarchical attention module is connected to the backbone feature extraction network and is used to perform weighted filtering on the feature maps to form a fused feature map. The stacked structure topology inference module is connected to the end of the deep hierarchical attention module and is used to transform the fused feature map into a topology map for inference. The detection head is connected to the end of the stacked structure topology inference module and uses an anchor-free decoupling head to output the bounding box, confidence score, and category of the cigarette.

[0036] Specifically, such as Figure 3 As shown, S3 specifically includes: S3.1: Multi-scale feature map extraction is performed on the feature map using a backbone feature extraction network; The backbone feature extraction network uses the CSPDarknet structure as its backbone network. Compared to the standard Darknet, the CSPDarknet structure duplicates the feature map twice. One copy is processed through dense convolutional blocks, while the other copy skips connections. Finally, the two copies are concatenated at the output. This design reduces the number of parameters and computational cost, while ensuring the diversity of gradient propagation.

[0037] Given that the feature map has 1536 channels, a channel compression convolutional layer is added at the beginning of the backbone feature extraction network. This layer uses a 1×1 image convolution kernel to compress the 1536 channels to 64 dimensions, significantly reducing the computational load of subsequent convolutional layers while preserving key spatial and depth information. After the channel compression convolutional layer outputs the compressed feature map, the Stem layer of the CSPDarknet structure receives the compressed feature map and performs preliminary downsampling and feature abstraction through two 3×3 convolutional layers. Then, it passes through four CSP modules sequentially. Each CSP module contains a residual convolutional block and a cross-stage local fusion structure. As the network deepens, the spatial resolution of the compressed feature map gradually decreases from 256×256 to 32×32, while the number of channels gradually increases from 64 to 512. Finally, three feature maps of different scales are output, namely the large-size feature map F. P3 Mesoscale feature map F P4 and small-scale feature map F P5 The large-size feature map F P3 The size is 80×80, rich in detailed information, and used to detect small targets or single cigarettes. The mesoscale feature map F P4The size is 40×40, used for detecting medium-sized targets, the small-scale feature map F P5 The 20×20 pixel map is rich in semantic information and is used to detect large targets or entire boxes of cigarettes. After processing by the backbone feature extraction network, the multi-scale feature map has been transformed from the original pixel-intensity representation into a semantic-vector representation, which has noise resistance and abstraction properties.

[0038] S3.2: Input the multi-scale feature map into the deep hierarchical attention module to perform deep dimension feature weighting to form a fused feature map; The deep hierarchical attention module receives the feature map output by the backbone feature extraction network. Since the original depth information is encoded in the channel dimension, it first needs to restore the channel dimension to a depth layer, and then perform inter-layer correlation modeling. That is, a deep sequence modeling sub-network with a Transformer image encoder structure is introduced. The deep sequence modeling sub-network treats the depth layer as a time series, and then performs self-attention calculation on the time series to capture the self-attention score between any two depth layers. Finally, a physical bias term based on Gaussian distribution is added to the self-attention score. This physical bias term encourages the deep hierarchical attention module to focus on depth layers within continuous intervals and penalizes discrete depth layers. Then, attention mask generation and application are performed to output attention weights for 512 depth layers. The attention weights are extended to match the shape of the multi-scale feature map and multiplied element-wise with the multi-scale feature map. This enhances feature values ​​belonging to continuous depth intervals and suppresses feature values ​​belonging to discrete depth points. This step realizes the transformation from data-driven to physically guided. Finally, upsampling and concatenation are performed to form a fused feature map.

[0039] The expression for the Gaussian-based physical bias term is as follows: ; In the formula, The attention weights are adjusted after adding spatial bias, where λ is the bias intensity coefficient, a learnable hyperparameter, i and j represent the spatial indices of the nodes, and σ represents the spatial scale parameter, which controls the width or range of influence of the Gaussian function.

[0040] Due to the sparsity and noisy environment characteristics of millimeter-wave images, millimeter-wave imaging typically exhibits a sparse, point-like distribution and is highly susceptible to interference from strongly reflective objects such as metal zippers and buttons. The introduction of the Gaussian kernel function allows spatially adjacent nodes to be identified even if their semantic feature similarity is low due to occlusion, through images with minimal spatial distance (i...). j) The image triggers a high-intensity bias bonus, thereby prioritizing the enhancement of the continuous contour lines of the cigarette packs during the feature aggregation stage. This effectively overcomes the shortcomings of traditional attention mechanisms that rely solely on semantic similarity and are prone to incorrectly severing broken edges. Simultaneously, it exhibits extremely high adaptability in suppressing background noise. For isolated high-reflectivity artifacts, due to the lack of spatially adjacent effective nodes, the Gaussian term imposes an exponentially decaying penalty, ensuring they are almost entirely excluded from the effective topological structure in subsequent graph convolutions, significantly reducing the false alarm rate. Secondly, this function greatly enhances the model's implicit perception of the physical arrangement of cigarette packs. As regular rigid bodies, the stacking of cigarette packs inevitably follows a strict linear or planar topology. By adjusting the image scale parameter σ, this bias function can adaptively simulate the width and spacing of the cigarette packs, forcing the model to prioritize message passing along the long axis of the cigarette packs during inference. This spatial guidance is equivalent to injecting a rule of physical common sense into the model, enabling it to reconstruct the complete stacking logic based on spatial proximity even when faced with severely compressed or partially damaged cigarette packs. This represents a leap from viewing pixel features to understanding physical structures.

[0041] S3.3: Input the fused feature map into the stacked structure topology reasoning module to model spatial topology relationships and generate a structured feature map; Traditional stacked topology inference modules are usually based on static graph convolution, which assumes that the connections and edges between nodes are fixed during the inference process, relying only on the initial spatial distance and feature similarity. However, when the cigarette pack is partially obscured by clothing, the initially constructed adjacency matrix may not be able to connect the segmented cigarette pack parts, resulting in the inability to propagate features and causing missed detections.

[0042] To address the aforementioned issues, this step reconstructs the stacked topology inference module and proposes a dynamic hierarchical graph convolutional network. The core improvement lies in introducing a physically guided dynamic edge update mechanism, enabling the model to dynamically correct connection relationships and inference hierarchy.

[0043] Specifically, S3.3 is as follows: Dynamic graph construction based on physical priors; The fused feature map is divided into an S×S grid to generate an initial node set V0. Each node in the initial node set V0 contains a position coordinate image and a feature vector f. i .

[0044] Physically constrained edge candidate set generation: Construct a physically guided candidate edge set. For each node, only if candidate node v j A potential connection is established only if the following conditions are met: Non-penetrating constraint: Based on the depth information provided by the depth layer attention module, if the difference between the depth values ​​of two nodes is greater than the difference threshold, they are determined to be objects in the preceding and following layers, and no edge is established. Coplanarity constraint: Calculate the angle between the normal vectors of the planes containing the two nodes. If the angle is less than the angle threshold, they are determined to be coplanar, and edges are established first. Minimum Spacing Constraint: There are usually tiny gaps (2-5mm) between cigarette packs. If the distance between two nodes is less than the preset threshold for the thickness error of the cigarette pack, they are determined to be different sampling points of the same object, and no edge is established to prevent a single cigarette pack from being identified as multiple.

[0045] Dynamic edge weight update introduces an edge weight prediction subnetwork, which uses the feature vector f of each node. i As input, the reliability score of the predicted edge is predicted in real time. During the training process, the edge weight prediction sub-network learns that parallel edges with uniform spacing have high weights and non-parallel edges with non-uniform spacing have low weights. During the inference phase, the adjacency matrix is ​​dynamically adjusted. As the number of network layers increases, incorrect connections are gradually weakened and correct connections are strengthened. Finally, multi-granularity convolution is performed on each node according to the dynamic edge weights.

[0046] Topological confidence field generation and feature reprojection; After multi-granularity convolution, each node is endowed with rich hierarchical semantic information. Through an MLP layer, a topological confidence score is output for each node. Then, graph feature reprojection is performed to map the learning results from the graph space back to the image space. Using bilinear interpolation, the feature vector and topological confidence score of each node are mapped back to the feature map generated in step S2, generating a structured feature map.

[0047] The stacked structure topology inference module provided in this embodiment solves the feature breakage problem caused by occlusion and stacking disorder when traditional static graph convolution is used to process millimeter-wave smoke-related cases by introducing a physical-guided dynamic edge update mechanism. The dynamic edge weight prediction subnetwork can eliminate erroneous connections that do not conform to the physical properties of cigarette packs in real time based on physical priors. The embedding of physical constraints greatly reduces the false alarm rate of the model. By hard-constraining non-penetration and coplanarity, the model eliminates the low-level error of misjudging two different layers of items as a single-layer stack of cigarette packs. At the same time, due to the introduction of inductive bias based on physical laws, the stacked structure topology inference module does not need to rely on massive amounts of data for pure statistical learning. It can grasp the essential rules of cigarette pack stacking with only a small number of labeled samples, which is particularly important in the actual law enforcement environment where samples of smoke-related cases are scarce.

[0048] S4: Input the structured feature map into the detection head and output the screening result.

[0049] The detection head is a Path Aggregation Network, which is used to further interpret the structured feature map and finally output the screening results.

[0050] like Figure 4 As shown, Figure 4 This paper demonstrates the performance advantages of the DSATR-Net model of this invention compared to traditional algorithms. The horizontal axis lists four schemes in order: general YOLOv8, general Faster R-CNN, standard GCN, and the present invention. The vertical axis represents the percentage of recognition accuracy. The industry benchmark marked in the figure is at 95%. General YOLOv8, with a red bar at 76.5%, exposes its inherent weakness in semantic feature matching in millimeter-wave sparse point cloud scenarios. General Faster... While R-CNN improved the accuracy to 79.2%, it was still limited by the confusion between disguised objects and cigarette packs. Standard GCN achieved 82.1% with basic graph structure inference but could not solve the problem of topological disconnection caused by occlusion and breakage. In contrast, this invention, with 99.2% of green bars, almost reached the 100% mark. This 22.7% performance leap directly verifies the strengthening effect of the deep hierarchical attention module on continuous deep features and the reasoning ability of dynamic graph convolution on stacked topological structures. In particular, the Gaussian bias function gives the model a spatial proximity prior, which enables it to accurately restore the physical arrangement logic of cigarette packs even in extremely complex scenarios such as cotton clothing wrapping and cardboard box interlayers. It completely breaks through the false alarm bottleneck caused by the lack of physical constraints in traditional algorithms, fully meets the almost demanding practical requirements of tobacco monopoly law enforcement for recognition accuracy, and provides solid technical confidence support for the subsequent zero-error law enforcement closed loop.

[0051] Example 2: This invention also provides a rapid non-invasive screening system for smoke-contaminated packages based on millimeter-wave radar imaging. The system employs a rapid non-invasive screening method for smoke-contaminated packages based on millimeter-wave radar imaging as described in Example 1. The system includes: The data acquisition module is used to acquire millimeter-wave radar three-dimensional imaging data of the package to be inspected; The feature extraction module is used to perform data preprocessing and multi-view projection on the three-dimensional imaging data to form a feature map; The recognition module is used to input the feature map into an adaptive feature enhancement recognition model that integrates prior knowledge of stacking morphology to achieve cigarette feature recognition; The result output module is used to input the structured feature map into the detection head and output the screening result.

[0052] Example 3: The present invention also provides an electronic device, including one or more processors and a memory.

[0053] A processor can be a central processing unit (CPU) or other form of processing unit with data processing and / or instruction execution capabilities, and can control other components in an electronic device to perform desired functions.

[0054] The memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and a processor may execute the program instructions to implement a rapid non-invasive screening method for smoke-contaminated packages based on millimeter-wave radar imaging, as described above in any embodiment of this application, and / or other desired functions. Various contents such as initial extrinsic parameters and thresholds may also be stored in the computer-readable storage medium.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A rapid, non-invasive screening method for smoke-contaminated packages based on millimeter-wave radar imaging, characterized in that, The method includes the following steps: S1: Acquire millimeter-wave radar 3D imaging data of the package to be inspected; S2: Perform data preprocessing and multi-view projection on the three-dimensional imaging data to form a feature map; S3: Input the feature map into the adaptive feature enhancement recognition model that integrates prior knowledge of stacked morphology to realize cigarette feature recognition; the overall architecture of the adaptive feature enhancement recognition model that integrates prior knowledge of stacked morphology consists of four parts: backbone feature extraction network, deep hierarchical attention module, stacked structure topology inference module and detection head; Specifically, S3 consists of: S3.1: using a backbone feature extraction network to extract multi-scale feature maps from the feature maps; S3.2: inputting the multi-scale feature maps into a deep hierarchical attention module for depth dimension feature weighting to form a fused feature map; S3.3: inputting the fused feature maps into a stacked structure topology reasoning module for spatial topology modeling to generate a structured feature map; S4: Input the structured feature map into the detection head and output the screening result.

2. The rapid, non-invasive screening method for smoke-contaminated packages based on millimeter-wave radar imaging according to claim 1, characterized in that, In S3.1, the backbone feature extraction network uses the CSPDarknet structure as the backbone network, and the feature map is copied twice. One copy is processed by dense convolutional blocks, and the other copy skips the connection directly. Finally, the two copies are spliced ​​at the output end. A channel compression convolutional layer is added at the beginning of the backbone feature extraction network. This layer uses a 1×1 image convolution kernel to compress the 1536-dimensional channels to 64 dimensions. After the channel compression convolutional layer outputs the compressed feature map, the Stem layer of the CSPDarknet structure receives the compressed feature map and performs preliminary downsampling and feature abstraction through two 3×3 convolutional layers. Then, it passes through four CSP modules in sequence. Each CSP module contains a residual convolutional block and a cross-stage local fusion structure. Finally, it outputs three feature maps of different scales, namely the large-size feature map F. P3 Mesoscale feature map F P4 and small-scale feature map F P5 .

3. The rapid, non-invasive screening method for smoke-contaminated packages based on millimeter-wave radar imaging according to claim 2, characterized in that, In step S3.2, the deep hierarchical attention module receives the feature map output by the backbone feature extraction network. First, it restores the channel dimension to a depth layer, and then performs inter-layer correlation modeling. That is, it introduces a deep sequence modeling sub-network with a Transformer image encoder structure. The deep sequence modeling sub-network treats the depth layer as a time series, and then performs self-attention calculation on the time series to capture the self-attention score between any two depth layers. Finally, it adds a physical bias term based on Gaussian distribution to the self-attention score. Then, it performs attention mask generation and application, outputting attention weights for 512 depth layers. The attention weights are extended to match the shape of the multi-scale feature map and multiplied element-wise with the multi-scale feature map. Then, it performs upsampling and concatenation to form a fused feature map.

4. The rapid non-invasive screening method for smoke-contaminated packages based on millimeter-wave radar imaging according to claim 3, characterized in that, The expression for the physical bias term based on the Gaussian distribution is: ； In the formula, The modified attention weights are defined by adding spatial bias, where λ is the bias intensity coefficient, i and j represent the spatial indices of the nodes, and σ represents the spatial scale parameter.

5. A rapid, non-invasive screening method for smoke-contaminated packages based on millimeter-wave radar imaging according to claim 3, characterized in that, Specifically, S3.3 is as follows: Dynamic graph construction based on physical priors; the fused feature map is divided into an S×S grid to generate an initial node set V0, where each node in the initial node set V0 contains position coordinates and a feature vector f. i ; Physically constrained edge candidate set generation: Construct a physically guided candidate edge set. For each node, only if candidate node v j A potential connection is established only if the following conditions are met: Non-penetrating constraint: Based on the depth information provided by the depth layer attention module, if the difference between the depth values ​​of two nodes is greater than the difference threshold, they are determined to be objects in the preceding and following layers, and no edge is established. Coplanarity constraint: Calculate the angle between the normal vectors of the planes containing the two nodes. If the angle is less than the angle threshold, they are determined to be coplanar, and edges are established first. Minimum Spacing Constraint: There are usually tiny gaps (2-5mm) between cigarette packs. If the distance between two nodes is less than the preset threshold for the thickness error of the cigarette pack, they are determined to be different sampling points of the same object, and no edge is established to prevent a single cigarette pack from being identified as multiple. Dynamic edge weight update introduces an edge weight prediction subnetwork, which uses the feature vector f of each node. i As input, the reliability score of the predicted edge is predicted in real time. During the training process, the edge weight prediction subnetwork learns that parallel edges with uniform spacing have high weights and non-parallel edges with non-uniform spacing have low weights. During the inference phase, the adjacency matrix is ​​dynamically adjusted at the same time, and finally, multi-granularity convolution is performed on each node according to the dynamic edge weight. Topological confidence field generation and feature reprojection: After multi-granularity convolution, each node is endowed with rich hierarchical semantic information. Through an MLP layer, a topological confidence is output for each node. Then, graph feature reprojection is performed to map the learning results of the graph space back to the image space. Using bilinear interpolation, the feature vector and topological confidence of each node are mapped back to the feature map generated in step S2 to generate a structured feature map.

6. A rapid, non-invasive screening method for smoke-contaminated packages based on millimeter-wave radar imaging according to claim 3, characterized in that, In step S4, the detection head is a Path Aggregation Network.

7. A rapid, non-invasive screening method for smoke-contaminated packages based on millimeter-wave radar imaging according to claim 1, characterized in that, In step S1, a portable millimeter-wave radar scanning device is used to acquire millimeter-wave radar three-dimensional imaging data of the package to be inspected. The portable millimeter-wave radar scanning device includes a radio frequency front-end, an antenna array, and a motion mechanism. The radio frequency front-end uses frequency-modulated continuous wave with a working frequency band of 77-81GHz, a sweep bandwidth of 4GHz, and a distance resolution of 3.75cm.

8. A rapid, non-invasive screening method for smoke-contaminated packages based on millimeter-wave radar imaging according to claim 7, characterized in that, Select the scanning mode according to the package size; for small packages, i.e., packages with a size of <30cm×30cm×30cm, the portable millimeter-wave radar scanning device starts near-field static scanning, with the radar array close to the surface of the package, and the entire package area is covered by electronic beam scanning, with a scanning time of ≤5 seconds.

9. A rapid, non-invasive screening method for smoke-contaminated packages based on millimeter-wave radar imaging according to claim 7, characterized in that, In S2, the preprocessing includes adaptive noise filtering based on density clustering and region of interest clipping based on the wrapping contour.

10. A rapid, non-invasive screening system for smoke-contaminated packages based on millimeter-wave radar imaging, characterized in that, The system employs a rapid, non-invasive screening method for smoke-contaminated packages based on millimeter-wave radar imaging, as described in any one of claims 1-9. The system comprises: The data acquisition module is used to acquire millimeter-wave radar three-dimensional imaging data of the package to be inspected; The feature extraction module is used to perform data preprocessing and multi-view projection on the three-dimensional imaging data to form a feature map; The recognition module is used to input the feature map into an adaptive feature enhancement recognition model that integrates prior knowledge of stacking morphology to achieve cigarette feature recognition; The result output module is used to input the structured feature map into the detection head and output the screening result.

Images