Edge ai-based printed packaging commodity anti-counterfeiting label verification system and method

Abstract

The application relates to the technical field of anti-counterfeiting verification, and discloses a printed packaging commodity anti-counterfeiting label verification system and method based on edge AI. The system comprises a multi-mode acquisition module, which synchronously acquires a digital image matrix, an original bottom layer radio frequency sequence and a printing process context vector of a label; a space compensation module, which is used for solving a spatial relative pose and performing reverse space compensation on the radio frequency sequence; a feature extraction module, which respectively extracts a visual physical form feature vector and a radio frequency physical degradation feature vector; a collaborative calculation module, which is used for calculating cross-modal collaborative correlation between the feature vectors, splicing and generating a high-dimensional joint feature tensor; and an edge inference module, which is used for taking the printing process context vector as a boundary adjustment parameter of a classification model, performing anti-counterfeiting judgment result output on the high-dimensional joint feature tensor. The application overcomes spatial pose interference in handheld verification, combines cross-modal physical collaboration and dynamic boundary adjustment, and improves the accuracy and robustness of anti-counterfeiting judgment.

Description

Technical Field

[0001] This invention relates to the field of anti-counterfeiting verification technology, specifically to an edge AI-based anti-counterfeiting label verification system and method for printed packaging products. Background Technology

[0002] Currently, in the anti-counterfeiting and traceability process of printed and packaged goods, anti-counterfeiting labels integrating optical patterns and radio frequency identification (RFID) chips are commonly used. In actual business scenarios, operators mostly verify the anti-counterfeiting labels using handheld edge computing devices. Due to the uncontrollability of handheld operation, the distance, pitch angle, and yaw angle between the device and the label will randomly change. This uncertainty in spatial pose will directly interfere with the phase and intensity of the radio frequency signal and cause perspective distortion in visual imaging, resulting in a large spatial propagation error in the underlying signal acquisition.

[0003] After signal acquisition, existing verification systems generally compare optical anti-counterfeiting features and radio frequency (RF) anti-counterfeiting features as two independent judgment branches. This processing mechanism breaks down the physical consistency of the label's own material and does not consider the cross-modal physical coupling relationship between the surface printing ink and the underlying RF antenna when they undergo the same printing process. This makes the anti-counterfeiting system easy to be deceived by criminals who independently imitate visual patterns and clone RF chips.

[0004] In the final classification and determination stage, conventional anti-counterfeiting verification models often employ fixed feature decision boundaries. However, in actual industrial production, fluctuations in the physical environment of the workshop during manufacturing inevitably occur between different batches of labels. This causes reasonable range drift in the microscopic printing features of genuine labels. Classification models using fixed boundaries cannot dynamically accommodate this feature drift caused by changes in the manufacturing environment, easily leading to misjudgments of genuine products and resulting in poor accuracy and adaptability of the anti-counterfeiting verification system. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an edge AI-based anti-counterfeiting label verification system and method for printed packaging products. It solves the problems of existing anti-counterfeiting label verification systems being susceptible to signal acquisition errors caused by handheld operation spatial pose interference, failing to consider cross-modal physical coupling between visual and radio frequency features under the same manufacturing process, and having low judgment accuracy due to the incompatibility of fixed decision boundaries with reasonable feature drift caused by manufacturing environment fluctuations.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides a verification system for anti-counterfeiting labels on printed packaging products based on edge AI, comprising: The multi-mode acquisition module is used to control the image sensor and the UHF RFID read / write hardware to perform synchronous sampling operations, acquire the digital image matrix of the target tag and the original underlying RF sequence, and acquire the printing process context vector through the network communication interface; The spatial compensation module is used to calculate the spatial relative pose parameters between the edge computing device and the physical plane of the target tag, and accordingly perform spatial compensation on the original underlying radio frequency sequence to generate a normalized radio frequency sequence. The feature extraction module is used to extract and generate visual physical morphology feature vectors from the digital image matrix, and simultaneously extract and generate radio frequency physical degradation feature vectors from the normalized radio frequency sequence. The collaborative computing module receives the visual physical morphology feature vector and the radio frequency physical degradation feature vector, and uses them to calculate the cross-modal collaborative correlation between the two feature vectors and concatenate them to generate a high-dimensional joint feature tensor. The edge reasoning module has a pre-built classification model, which is used to configure the printing process context vector as the spatial decision boundary adjustment parameter of the classification model, and input the high-dimensional joint feature tensor into the classification model after boundary adjustment, and output discretized decision result variables.

[0007] The technical solution provided by the first aspect of the present invention, by linking visual and radio frequency dual-mode signals, using spatial calculation to perform reverse compensation on radio frequency signals, and extracting the collaborative features of optical and radio frequency media under the same printing process, dynamically adjusting the decision boundary of the classification model in combination with the printing process context, overcomes the spatial pose interference in the handheld verification process and improves the accuracy of anti-counterfeiting judgment.

[0008] Furthermore, the multi-mode acquisition module is specifically used for: The system's global clock generator generates edge-aligned hardware trigger pulse signals, which are then sent to the global shutter control pin of the image sensor and the baseband control pin of the UHF RFID read / write hardware, respectively. The identification information of the target tag is parsed from the digital baseband payload of the RF communication signal, and the original backscattering timing sequence of the physical layer is extracted to generate the original bottom-layer RF sequence. The identification information is sent to the cloud data platform, and the system receives the printing process context vector returned by the cloud data platform based on the identification information. The printing process context vector represents the set of parameters of the physical manufacturing environment in which the target tag is printed at the factory.

[0009] In the above-mentioned further scheme, the optical imaging exposure start time and the radio frequency carrier transmission time are physically aligned on the time axis through the underlying hardware clock pulse, eliminating data misalignment caused by asynchronous acquisition; at the same time, the physical parameters of the manufacturing workshop are dynamically obtained according to the identification information, providing an environmental constraint benchmark for the dynamic verification of the subsequent network model.

[0010] Furthermore, the specific process by which the spatial compensation module calculates the spatial relative pose parameters includes: The region of interest containing the target label's identification information is extracted from the digital image matrix. The set of observation anchor points within the region of interest is extracted, and a pre-stored set of physical world reference points is retrieved. Based on the perspective mapping relationship between the set of observation anchor points and the set of physical world reference points, the homography matrix is ​​calculated. After the homography matrix is ​​verified to be valid, matrix decomposition is performed on the homography matrix in conjunction with the camera's intrinsic parameter matrix to calculate the spatial relative pose vector of the edge computing device relative to the physical plane of the target label.

[0011] In the above-mentioned further scheme, a perspective mapping model is established by utilizing the correspondence of geometric anchor points in the optical image, and the random distance, pitch angle and yaw angle parameters generated in the handheld verification operation are transformed into definite mathematical vectors, providing a data basis for the reverse spatial compensation of radio frequency signals.

[0012] Furthermore, the specific process by which the spatial compensation module generates the normalized radio frequency sequence includes: Using the distance and angle parameters in the spatial relative pose vector, the equivalent propagation distance from the phase center of the RF antenna to the tag response center is calculated; based on the equivalent propagation distance, inverse spatial phase compensation calculation is performed on each frame of the original phase observation value in the original underlying RF sequence to generate a normalized phase; based on the spatial relative pose vector and the real-time operating frequency, inverse free space loss and antenna pattern gain compensation calculation is performed on each frame of the original received signal strength indication value in the original underlying RF sequence to generate a normalized received signal strength; the normalized phase and the normalized received signal strength are combined to generate the normalized RF sequence.

[0013] In the above further scheme, based on the three-dimensional spatial pose obtained by calculation, the electromagnetic wave phase offset and intensity loss caused by the change of measurement distance and reading angle are eliminated, and the radio frequency parameter sequence characterizing the physical degradation characteristics of the internal antenna itself is restored from the radio frequency communication signal.

[0014] Furthermore, the specific process by which the feature extraction module generates visual physical morphology feature vectors includes: A standardized label image is generated by performing an inverse perspective transformation on the digital image matrix using the homography matrix, and the region of interest for ink verification is extracted. The gradient change amplitude of each pixel in the region of interest for ink verification is calculated and spatially compared with a pre-stored lossless reference template image to calculate and generate an ink bleeding diffusion index. The statistical variance of the halftone dot area coverage in the local spatial neighborhood of the region of interest for ink verification is calculated to generate a dot gain variance. The ink bleeding diffusion index and the dot gain variance are combined to generate the visual physical morphology feature vector.

[0015] In the above-mentioned further scheme, the optical image is reverse-corrected to a standard spatial coordinate system, and the microscopic capillary diffusion phenomenon of ink on the printing substrate and the non-uniform characteristics of dot expansion are quantitatively extracted to characterize the printing features of the label surface in vector form.

[0016] Furthermore, the specific process by which the feature extraction module generates the radio frequency physical degradation feature vector includes: Phase data within all valid sampled frames of the normalized radio frequency sequence is extracted, and the residual impedance phase drift relative to the standard reference phase is calculated. The time-series variance is calculated for the normalized received signal strength sequence within the normalized radio frequency sequence to generate a signal timing fluctuation parameter reflecting the dispersion of radio frequency backscattered energy. The residual impedance phase drift and the signal timing fluctuation parameter are combined to generate the radio frequency physical degradation feature vector.

[0017] In the above further scheme, the phase drift and energy timing fluctuation parameters caused by deformation or impedance mismatch of the antenna conductive substrate are extracted from the normalized radio frequency parameters, and the physical degradation properties of the internal radio frequency medium are parameterized and extracted.

[0018] Furthermore, the specific process by which the collaborative computing module generates the high-dimensional joint feature tensor includes: Calculate the cross product of the visual physical morphology feature vector and the radio frequency physical degradation feature vector to generate an initial cross feature matrix; introduce a preset local covariance weight matrix, perform Hadamard product operation on the initial cross feature matrix to generate a cross-modal physical synergy matrix used to eliminate dimensional differences and amplify physical consistency cross feature terms; perform a dimension flattening operation on the cross-modal physical synergy matrix to generate a one-dimensional synergy feature vector; concatenate the visual physical morphology feature vector, the radio frequency physical degradation feature vector, and the one-dimensional synergy feature vector in the column direction to generate the high-dimensional joint feature tensor.

[0019] In the above further scheme, the cross-modal physical coupling relationship between visual media and radio frequency media generated in the same manufacturing process environment is expressed by isomorphic matrix calculation. The local covariance weight matrix is ​​used to suppress irrelevant noise and amplify the cross terms with physical consistency, and the independent single-domain features are integrated into a joint feature distribution.

[0020] Furthermore, the edge inference module internally deploys a classification model based on a conditionally gated multilayer perceptron architecture. This classification model includes a feature backbone network, a context branch network, and a dynamic feature modulation layer. The specific decision-making process includes: The high-dimensional joint feature tensor is input into the feature backbone network to extract the basic hidden layer feature vector; the printing process context vector is input into the context branch network to calculate and generate the scaling coefficient vector and the bias coefficient vector; in the dynamic feature modulation layer, the scaling coefficient vector and the bias coefficient vector are used to perform an element-based affine transformation on the basic hidden layer feature vector to generate the modulation feature vector.

[0021] In the above further scheme, the acquired environmental constraint variables are converted into boundary adjustment coefficients of the classification model, and the feature space of the backbone layer of the network is stretched or twisted using the affine transformation mechanism, so as to adapt to the reasonable feature drift caused by environmental fluctuations of different manufacturing batches and enhance the generalization performance of the model.

[0022] Furthermore, the edge reasoning module also includes the following process when performing classification decisions: The modulated feature vector is transmitted to the output layer of the classification model to perform dimensionality reduction mapping, and the authenticity probability score of the target label is output. The authenticity probability score is compared with a preset confidence threshold. If it is greater than or equal to the confidence threshold, a discretized business status indication of authenticity release is output. If it is less than the confidence threshold, a discretized business status indication of counterfeiting warning is output. When the printing process context vector is null due to cloud communication anomaly, the edge inference module triggers a bypass fault tolerance mechanism, sets the scaling coefficient vector to an all-one vector, and sets the bias coefficient vector to an all-zero vector, so that the classification model degenerates into a fixed-boundary classifier.

[0023] In the above-mentioned further solutions, a clear business result output is achieved through probability threshold comparison. At the same time, a network anomaly bypass mechanism is set up so that when the external network is interrupted, it automatically degrades to the standard inference mode to ensure the availability of the system under weak network or network outage conditions.

[0024] A second aspect of this invention provides a method for verifying anti-counterfeiting labels on printed packaging products based on edge AI, applied to the aforementioned system, comprising the following steps: S1. Perform the steps of synchronous acquisition of multimodal anti-counterfeiting signals and acquisition of manufacturing parameters. Control the equipment to perform synchronous sampling operations to acquire the digital image matrix and original underlying radio frequency sequence of the target label, and obtain the printing process context vector from the cloud. S2. Perform a dynamic spatial compensation step for radio frequency based on visual anchor points, calculate the spatial relative pose parameters between the edge computing device and the physical plane of the target tag, and perform spatial compensation calculation on the original bottom-layer radio frequency sequence accordingly to generate a normalized radio frequency sequence. S3. Perform a deterministic extraction step of dual physical domain micro-degradation features, and extract and generate visual physical morphology feature vector and radio frequency physical degradation feature vector from the digital image matrix and the normalized radio frequency sequence, respectively. S4. Perform the isomorphic cross-modal physical synergy feature tensor construction step, calculate the cross-modal physical synergy association between the two feature vectors and concatenate them to generate a high-dimensional joint feature tensor; S5. Execute the dynamic classification decision step based on manufacturing environment constraints, configure the printing process context vector as the boundary adjustment constraint condition of the internal preset classification model, perform nonlinear classification mapping calculation on the high-dimensional joint feature tensor through the classification model after boundary adjustment, and output the label anti-counterfeiting judgment result.

[0025] This invention provides a system and method for verifying anti-counterfeiting labels on printed packaging products based on edge AI. It has the following beneficial effects: 1. This invention extracts observation anchor points from the digital image matrix to calculate the spatial relative pose parameters between the edge computing device and the tag, and uses these parameters to perform inverse spatial phase and signal strength compensation on the original underlying radio frequency sequence. This eliminates spatial propagation errors caused by random variations in measurement distance and reading angle during manual handheld verification, obtaining a normalized radio frequency sequence that directly reflects the structural properties of the tag antenna, thereby improving the accuracy of underlying radio frequency signal extraction.

[0026] 2. After simultaneously extracting visual physical morphological feature vectors and radio frequency physical degradation feature vectors, this invention uses a local covariance weight matrix to perform a Hadamard product operation on the cross-feature matrix of the two, generating a cross-modal physical synergy matrix. This establishes a physical coupling constraint between the optical ink and the radio frequency antenna under the same manufacturing process, integrating single-modal features into a high-dimensional joint feature tensor. This overcomes the deficiency of single-modal features being easily counterfeited independently in conventional technologies, improving the reliability of anti-counterfeiting verification.

[0027] 3. This invention obtains the printing process context vector and inputs it into the context branch network of the classification model to generate scaling and bias coefficients. Then, it performs affine transformation on the features of the backbone network. This process dynamically adjusts the decision boundary of the classification model according to the objective physical manufacturing environment when the label leaves the factory, so that the system can adaptively accommodate feature drift between different manufacturing batches and reduce the misjudgment rate caused by batch fluctuations. At the same time, the bypass fault tolerance mechanism configured in the system ensures the continuous availability of the equipment in the case of communication abnormality. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention; Figure 3 This is a schematic diagram of the process for synchronous acquisition of multimodal anti-counterfeiting signals and acquisition of manufacturing parameters in this invention; Figure 4 This is a schematic diagram of the radio frequency spatial dynamic compensation based on visual anchor points according to the present invention. Figure 5 This is a schematic diagram of the deterministic extraction process for dual physical domain degradation features in this invention. Figure 6 This is a schematic diagram illustrating the process of constructing the isomorphic cross-modal physical synergy feature tensor of the present invention; Figure 7 This is a schematic diagram of the dynamic classification decision based on manufacturing environment constraints of the present invention; Figure 8 This is a graph comparing the subject operating characteristic performance of different verification algorithms of this invention.

[0029] Among them, 100 is the multi-mode acquisition module; 200 is the spatial compensation module; 300 is the feature extraction module; 400 is the collaborative computing module; and 500 is the edge reasoning module. Detailed Implementation

[0030] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Reference Figure 1 The present invention provides a verification system for anti-counterfeiting labels of printed packaging products based on edge AI. The system may include: a multi-mode acquisition module 100, a spatial compensation module 200, a feature extraction module 300, a collaborative computing module 400, and an edge reasoning module 500.

[0032] This anti-counterfeiting verification system is deployed in an edge computing device. The edge computing device is equipped with an image sensor, UHF RFID reader / writer hardware, and a network communication interface. The image sensor is used to acquire the optical pixel array on the surface of the target label; the UHF RFID reader / writer hardware is used to read the underlying timing characteristics in the RF communication link; and the network communication interface is used to establish a data interaction channel with the cloud data platform.

[0033] The multi-mode acquisition module 100 is used to control the image sensor and the UHF RFID read / write hardware to perform synchronous sampling operations when the verification process starts. The multi-mode acquisition module 100 acquires the current tag's identification information, digital image matrix, and raw underlying RF sequence.

[0034] The multi-mode acquisition module 100 is also used to send the identification information to the cloud data platform via a network communication interface, and to receive the printing process context vector returned by the cloud data platform based on the identification information. The printing process context vector is used to characterize the objective manufacturing environment when the target label is printed at the factory.

[0035] In one implementation, the cloud-based data platform determines the production batch to which the target label belongs based on the identification information and returns a set of batch baseline parameters bound to that production batch. The set of batch baseline parameters includes one or more of the following: standard reference phase, local covariance weight matrix, confidence threshold, reference template image index, classification model version identifier, and process normalized extreme value parameters. The edge computing device caches this set of batch baseline parameters in local storage for use by the spatial compensation module 200, feature extraction module 300, collaborative computing module 400, and edge inference module 500.

[0036] The spatial compensation module 200 is communicatively connected to the multi-mode acquisition module 100. The spatial compensation module 200 extracts boundary geometric anchor points within the digital image matrix and calculates the spatial relative pose parameters between the edge computing device and the physical plane of the target tag. Based on the spatial relative pose parameters, the spatial compensation module 200 performs inverse spatial path loss calculation and aligns the RF phase offset on the original underlying RF sequence to generate a normalized RF sequence.

[0037] The feature extraction module 300 is connected to the output of the spatial compensation module 200. The feature extraction module 300 calculates the ink edge gradient distribution data of a preset region in the digital image matrix using morphological algorithms, generating a visual physical morphological feature vector. Simultaneously, the feature extraction module 300 extracts the residual impedance phase drift and signal temporal fluctuation variance within the normalized radio frequency sequence, generating a radio frequency physical degradation feature vector.

[0038] The collaborative computing module 400 receives visual physical morphology feature vectors and radio frequency physical degradation feature vectors. The collaborative computing module 400 constructs a cross-modal physical coordination degree matrix by calculating the outer product mapping and local covariance between the two feature vectors. The collaborative computing module 400 then performs a dimension flattening operation, concatenating the visual physical morphology feature vectors, radio frequency physical degradation feature vectors, and the cross-modal physical coordination degree matrix into a high-dimensional joint feature tensor.

[0039] The edge inference module 500 acquires the printing process context vector output by the multi-mode acquisition module 100 and the high-dimensional joint feature tensor output by the collaborative computing module 400. The edge inference module 500 has a pre-built classification model. It configures the printing process context vector as a spatial decision boundary adjustment parameter for the classification model and inputs the high-dimensional joint feature tensor into the adjusted boundary classification model to perform nonlinear mapping calculations, outputting discretized decision result variables.

[0040] Reference Figure 2 This invention provides a method for verifying anti-counterfeiting labels on printed packaging products based on edge AI. This method is executed based on the aforementioned anti-counterfeiting verification system and includes the following steps: In step S1, the multi-mode acquisition module 100 acquires the optical image and underlying radio frequency communication sequence of the target tag, and parses the identification information from the underlying radio frequency communication sequence. The multi-mode acquisition module 100 sends the identification information to the cloud and receives the printing process context vector retrieved and returned by the cloud based on the identification information.

[0041] In step S2, the spatial compensation module 200 receives the optical image and the underlying radio frequency communication sequence. The spatial compensation module 200 calculates the spatial pose parameters of the verification device relative to the tag plane using the geometric feature anchor points within the optical image. Based on the spatial pose parameters, the spatial compensation module 200 performs spatial path loss calculation and radio frequency phase offset compensation on the underlying radio frequency communication sequence, outputting a normalized radio frequency sequence.

[0042] In step S3, the feature extraction module 300 processes the optical image and extracts the ink edge gradient distribution data to generate a visual physical morphology feature vector. The feature extraction module 300 also extracts the impedance phase shift and signal timing fluctuation parameters within the normalized radio frequency sequence to generate a radio frequency physical degradation feature vector.

[0043] In step S4, the collaborative computing module 400 receives the visual physical morphology feature vector and the radio frequency physical degradation feature vector. The collaborative computing module 400 performs an outer product mapping operation and a local covariance operation on the two feature vectors to generate a cross-modal physical coherence matrix. The collaborative computing module 400 then flattens and concatenates the visual physical morphology feature vector, the radio frequency physical degradation feature vector, and the cross-modal physical coherence matrix to generate a high-dimensional joint feature tensor.

[0044] In step S5, the edge inference module 500 receives the printing process context vector and the high-dimensional joint feature tensor. The edge inference module 500 configures the printing process context vector as the boundary adjustment constraint of the classification model, performs nonlinear classification mapping calculation on the high-dimensional joint feature tensor through the adjusted classification model, and outputs the label anti-counterfeiting judgment result.

[0045] Reference Figure 3 Step S1 specifically includes the following sub-steps: In step S11, the multi-mode acquisition module 100 sends a synchronization trigger command to the image sensor and UHF RFID read / write hardware inside the edge computing device. The synchronization trigger command generates an edge-aligned hardware trigger pulse signal based on the system's global clock generator.

[0046] In the specific hardware implementation, the multi-mode acquisition module 100 sends synchronous level transition signals to the image sensor and the UHF RFID read / write hardware respectively through the GPIO interface of the edge computing device. The global shutter control pin of the image sensor and the baseband control pin of the UHF RFID read / write hardware receive the hardware trigger pulse signal, aligning the optical imaging exposure start time with the RF carrier transmission time on the same physical time axis.

[0047] In step S12, the image sensor responds to the hardware trigger pulse signal, captures the optical diffuse reflection signal from the surface of the target label, and generates a two-dimensional digital image matrix through analog-to-digital conversion. ,in and These represent the horizontal and vertical spatial coordinates of a pixel on a two-dimensional image plane, respectively.

[0048] Within the synchronous time window of optical image acquisition, the UHF RFID read / write hardware responds to the same hardware trigger pulse signal, transmits continuous wave RF energy to the target tag, and receives the RF communication signal formed by backscattering from the antenna inside the target tag.

[0049] The multi-mode acquisition module 100 calls the low-level reader protocol to access the baseband registers of the UHF RFID read / write hardware to obtain low-level physical layer data that has not been filtered by the application layer interface. The UHF RFID read / write hardware is a read / write module that supports the output of physical layer phase values, received signal strength indication values, and sampling timestamps. Its baseband registers include at least a phase observation register, a received signal strength register, and a frame timestamp register. The multi-mode acquisition module 100 can read the data in the aforementioned registers through a serial peripheral interface, a universal asynchronous transceiver interface, or a high-speed peripheral bus.

[0050] The multi-mode acquisition module 100 parses the unique identifier of the target tag from the digital baseband payload of the radio frequency communication signal and generates identification information. Simultaneously, the multi-mode acquisition module 100 extracts the original backscattering time sequence of the radio frequency communication signal at the physical layer, generating the original bottom-layer radio frequency sequence. The original underlying radio frequency sequence here. It contains several consecutive radio frequency sampling frames, each sampling frame consisting of a timestamp. Original phase observations and the original received signal strength indication value These are all components. These underlying radio frequency parameters preserve the objective physical state of electromagnetic wave signals during free space propagation and backscattering, without software smoothing or filtering.

[0051] Step S13, the multi-mode acquisition module 100 will transfer the identification information The request is encapsulated into a data request message and sent to the cloud data platform via the wireless communication interface of the edge computing device. The server in the cloud data platform receives the data request message, performs a primary key index query in the relational database, and retrieves the identification information. The bound production batch file data.

[0052] The multi-mode acquisition module 100 receives the response message sent by the cloud data platform based on the production batch file data, and parses out the printing process context vector from it. This printing process context vector represents the set of parameters of the objective physical manufacturing environment in which the target label is located during the printing process.

[0053] The response message may also include a set of batch reference parameters corresponding to the production batch to which the target tag belongs. This set of batch reference parameters is used to provide benchmark data for subsequent RF benchmark comparison, visual template comparison, cross-modal collaborative computation, and classification decision. The set of batch reference parameters includes a standard reference phase. Local covariance weight matrix Non-destructive benchmark template image or its index, confidence threshold One or more of the following: classification model version identifier and process normalization extreme value parameters.

[0054] For printed anti-counterfeiting labels, the ink on the label surface and the conductive silver paste of the internal antenna are affected by environmental humidity, printing pressure, and the pore structure of the substrate during printing, lamination, or curing. These manufacturing parameters can cause batch-consistent physical deviations in the visual morphology and radio frequency impedance response of genuine labels. Therefore, obtaining the printing process context vector is beneficial for setting reasonable tolerance boundaries for subsequent cross-modal feature verification.

[0055] The specific matrix representation of the printing process context vector acquired by the multi-mode acquisition module 100 is as follows: ; In the formula, Represents the printing process context vector; superscript This represents the transpose operation of a vector; This represents the average value of continuous environmental humidity in the manufacturing workshop where the target label was printed, reflecting the microscopic influence of air moisture on the printing medium. This represents the measured value of the printing pressure set by the printing press and applied directly to the surface of the target label. The fiber porosity index represents the fiber porosity index of the substrate used for printing the target label. This index characterizes the substrate's microscopic capillary permeation physical properties to liquid printing media.

[0056] As an feasible approach, to facilitate dimensionless processing of matrix operations in subsequent modules, the measured values ​​of environmental humidity, printing pressure, and fiber porosity index issued by the cloud data platform have all been mapped to normalized scalars between 0 and 1 based on preset corresponding process extreme values. The multi-mode acquisition module 100 acquires the digital image matrix. Original underlying radio frequency sequence and printing process context vector It is cached in the memory of the edge computing device for use by other logical modules within the system.

[0057] Reference Figure 4 Step S2 specifically includes the following sub-steps: In handheld or mobile verification scenarios, the distance, pitch angle, and yaw angle between the edge computing device and the target tag typically vary randomly. These spatial variations simultaneously affect the free-space propagation loss and propagation phase of the radio frequency (RF) signal. The spatial compensation module 200 calculates the relative pose between the device and the tag using visual anchor points and compensates for the underlying RF signal accordingly, thereby reducing the impact of the on-site scanning posture on RF feature extraction.

[0058] Step S21, the spatial compensation module 200 receives the digital image matrix buffered by the multi-mode acquisition module 100. The spatial compensation module 200 captures a digital image matrix. The region of interest contains the identification information. As one possible approach, this identification information is embedded within the QR code graphic, and the region of interest is the pixel area containing the QR code and its positioning reference graphic.

[0059] The target label can be set with both a QR code visual identifier and an RFID electronic identifier. The identifier code in the QR code visual identifier and the UID in the UHF RFID electronic identifier are pre-bound in the cloud data platform. The multi-mode acquisition module 100 preferentially uses the UID parsed from the RF communication signal as the primary key for cloud query; when the RF reading is abnormal, the identifier code in the QR code visual identifier can also be used as a backup query index.

[0060] The spatial compensation module 200 uses a corner detection algorithm to extract four standard boundary geometric anchor points within the region of interest, generating a set of observation anchor points in the image pixel coordinate system. For the corner feature recognition and extraction process of digital images, those skilled in the art can configure computer vision algorithms such as Harris corner detection to implement it. The underlying gradient optimization logic is a well-known technology in this field and will not be elaborated upon here. The spatial compensation module 200 synchronously retrieves the standard physical size parameters of the target labels pre-stored in the edge computing device, and generates a corresponding set of physical world reference points based on these standard physical size parameters. .

[0061] Step S22, the spatial compensation module 200, based on the set of observation anchor points... With the set of reference points of the physical world The geometric correspondence between them is used to calculate the homography matrix of the current digital image. Homography matrix It characterizes the perspective mapping relationship between the two-dimensional physical plane where the target label is located and the imaging plane of the image sensor.

[0062] Spatial compensation module 200 homography matrix The validity of the matrix calculation is verified. If the matrix calculation condition number is higher than the preset convergence threshold, it indicates that the current scanning angle is too large or local occlusion has occurred. At this time, a resampling command is generated and fed back to the operation interface to prompt the user to re-align. If it is valid, the homography matrix is ​​then combined with the camera intrinsic parameter matrix calibrated by the edge computing device at the factory to verify the homography matrix. Perform matrix factorization to calculate the spatial relative pose vector of the edge computing device relative to the physical plane of the target tag. .

[0063] Edge computing devices undergo pre-calibration of extrinsic parameters between the image sensor coordinate system and the UHF RFID antenna phase center coordinate system during the manufacturing process, and the corresponding rigid transformation matrix is ​​stored in local memory. After obtaining the pose of the image sensor relative to the physical plane of the target tag, the spatial compensation module 200 uses this rigid transformation matrix to calculate the equivalent spatial pose parameters of the RF antenna phase center relative to the physical plane of the target tag. These equivalent spatial pose parameters are used for subsequent RF phase compensation and received signal strength compensation. Spatial relative pose vector The specific form of expression is: ; In the formula, Represents the spatial relative pose vector; This represents the normal distance from the phase center of the built-in radio frequency antenna of the edge computing device to the physical plane of the target tag; Indicates the pitch angle during device scanning and reading; This represents the yaw angle. The above spatial calculation process transforms the operational randomness generated when manually holding the verification equipment into deterministic spatial geometric parameters.

[0064] Step S23: Spatial compensation module 200 retrieves the original low-level radio frequency sequence cached by multi-mode acquisition module 100. .

[0065] The spatial compensation module 200 utilizes spatial relative pose vectors normal distance in Or use normal distance Pitch angle and yaw angle The calculated equivalent propagation distance from the phase center of the RF antenna to the center of the tag response For the original underlying radio frequency sequence Each frame of raw phase observations Perform inverse spatial phase compensation calculations to generate a normalized phase. Where the tag response center and the normal projection point of the RF antenna phase center approximately coincide, the equivalent propagation distance is... It can be approximated as the normal distance. When the device is in a significantly tilted reading posture, the spatial compensation module 200 adjusts the tilt angle accordingly. With yaw angle With normal distance Perform geometric corrections to obtain the equivalent propagation distance. The specific compensation formula is as follows: ; In the formula, This indicates the operating wavelength of the radio frequency carrier transmitted by the edge computing device; Indicates taking Modulo remainder operation; Indicates the normalized phase; Represents the original phase observation value; Indicates the normal distance; This represents the constant pi. The remainder operation is used to fold the compensated absolute phase value to the standard range of zero to two p-threshold principal values. This calculation eliminates spatial electromagnetic wave phase offset caused by variations in the physical distance between the device and the tag.

[0066] Spatial compensation module 200 continues to utilize normal distance and spatial relative pose vector pitch angle With yaw angle For the original underlying radio frequency sequence Raw received signal strength indication value for each frame within Perform inverse free-space loss and antenna pattern gain compensation calculations to generate normalized received signal strength. The calculation formula is as follows: ; In the formula, Indicates the normalized received signal strength; This indicates the original received signal strength indication value; This is a mathematical operator representing a common logarithmic function with base 10; Indicates the normal distance; This indicates the real-time operating frequency of the UHF RFID read / write hardware. This indicates the target tag's internal antenna at a specific elevation angle. With yaw angle The relative gain function value is generated by fitting the microwave anechoic chamber measurement parameters at the time of tag manufacturing and stored in the edge computing device. Specifically, the cloud data platform or calibration device collects the relative gain data of the target tag antenna under multiple discrete pitch and yaw angle combinations during the sample manufacturing stage, and generates an angle-gain lookup table or parameterized fitting function. When performing compensation, the edge computing device uses the currently calculated pitch angle... With yaw angle The corresponding relative gain function value can be obtained by looking up a table, matching the nearest angle, or bilinear interpolation. ; This represents the system's static calibration constant. The specific values ​​are determined by the edge computing device at the factory stage based on the measured data of the internal RF feeder insertion loss and antenna transmit / receive gain, and are stored in the device's read-only memory.

[0067] The spatial compensation module 200 will be composed of normalized phase With normalized received signal strength The newly constructed time series are combined and packaged to generate a normalized radio frequency sequence. Normalized radio frequency sequence This represents the pure set of radio frequency parameters after removing interference from macroscopic spatial distance and angle attitude at the verification site. The spatial compensation module 200 normalizes the radio frequency sequence. It is cached in the system memory pool for subsequent feature extraction logic to call.

[0068] Reference Figure 5 Step S3 specifically includes the following sub-steps: During the actual manufacturing and printing process of the label, the ink will physically penetrate the substrate, and at the same time, the conductive silver paste of the embedded radio frequency antenna will also change its impedance characteristics due to the process pressure. The feature extraction module 300 quantitatively extracts these two degradation features that are on the same physical medium.

[0069] Step S31: Feature extraction module 300 retrieves the digital image matrix cached by multi-mode acquisition module 100. The feature extraction module 300 first uses the homography matrix H obtained by the spatial compensation module 200 to compare the digital image matrix. An inverse perspective transformation is performed to correct the target label image to the standard label plane coordinate system, generating a standardized label image. The feature extraction module 300 filters the macroscopic geometric coding data within this digital image matrix, or filters the macroscopic geometric coding data within the standardized label image, and extracts the ink verification region of interest (ROI) for anti-counterfeiting verification based on the spatial coordinate offset generated with the geometric anchor points extracted by the spatial compensation module 200 as the origin. The position, size, and orientation of the ink verification ROI are determined by a preset coordinate offset relative to the geometric anchor points, ensuring that the ink verification ROI in the currently acquired image and the corresponding region in the non-destructive reference template image are under the same spatial coordinate reference.

[0070] The feature extraction module 300 calculates the gradient change amplitude of each pixel in the region and spatially compares it with the non-destructive reference template image pre-stored in the edge computing device to calculate the ink staining diffusion index. The non-destructive benchmark template image can be generated from standard sampled images of genuine labels from the same batch after noise reduction, perspective correction, and mean fusion processing. Alternatively, it can be generated from the label design master image after mapping to printing resolution. The non-destructive benchmark template image is bound to the production batch, label format, or template index of the target label and is then distributed from the cloud data platform to edge computing devices or pre-stored in edge computing devices. The specific calculation formula is as follows: ; In the formula, Indicates the ink bleeding and diffusion index; This indicates the total number of edge pixels contained within the region of interest for ink verification. The pixel index variable represents the pixel number being calculated. One edge pixel; This represents the actual pixel gradient vector of the digital image matrix at that coordinate point. This represents the diffusionless reference gradient vector of the reference template image at the corresponding coordinate point; This refers to performing an L2 norm operation on the difference between two gradient vectors, i.e., calculating the Euclidean distance. In the algorithm execution logic, the feature extraction module 300 first verifies the total number of pixels. The value of . If If the result is zero or less than the set effective threshold, it indicates that the current image is severely out of focus or overexposed, resulting in the loss of edge features. The module will then interrupt the calculation and generate an image acquisition abnormality status word to avoid triggering a program abnormal interruption with a denominator of zero. For the extraction of edge gradient vectors in digital images, those skilled in the art can use the Sobel operator or the Cannibal operator to perform spatial convolution calculations. The specific matrix convolution operation logic is a well-known technique in the field and will not be elaborated here.

[0071] In the principles of printing physics, halftone dots undergo nonlinear physical expansion under printing pressure during the printing process, and this expansion exhibits statistical non-uniformity on the surface of a rough substrate. Based on this principle, the feature extraction module 300 further calculates the halftone dot area coverage of the ink within the region of interest, calculates the statistical variance of this coverage within the local spatial neighborhood, and generates the dot gain variance. The feature extraction module 300 extracts the ink diffusion index. Increased variance with network points Merge into column vector form to generate visual physical morphology feature vectors Visual physical morphology feature vector The specific structural characterization is as follows .

[0072] Step S32, the feature extraction module 300 retrieves the normalized radio frequency sequence generated by the spatial compensation module 200. After eliminating spatial interference parameters such as external measurement distance and reading angle, the normalized RF sequence The residual signal distortion is characterized in the system theoretical model as the impedance change of the antenna inside the target tag. The feature extraction module 300 extracts the phase data from all valid sampled frames in the sequence and calculates the residual impedance phase drift. The calculation formula is as follows: ; In the formula, This indicates the residual impedance phase shift. Represents the normalized radio frequency sequence The total number of valid sampled frames contained within, and the system settings. To ensure the validity of the mean calculation; The index variable is the sampling frame, representing the frame currently being calculated. Frame data; Indicates the first position in the sequence The normalized phase value of the frame; This indicates the standard reference phase inherent to the batch to which the target tag belongs at the time of manufacture. This standard reference phase data is obtained along with the message sent by the cloud data platform, or pre-stored in the memory of the edge computing device, representing the antenna phase feedback value under ideal conditions without cracks or oxidation.

[0073] Feature extraction module: 300 pairs of normalized radio frequency sequences Normalized received signal strength within The sequence performs time variance calculation to generate signal time fluctuation parameters. The timing fluctuation parameters of this signal This reflects the dispersion of RF backscattered energy caused by impedance mismatch in the antenna microstrip line. The feature extraction module 300 extracts the residual impedance phase shift. With signal timing fluctuation parameters Merge and splice to generate radio frequency physical degradation feature vectors Radio frequency physical degradation eigenvectors The specific structural characterization is as follows The feature extraction module 300 will generate a visual-physical feature vector. With radio frequency physical degradation feature vector It is output to the system's data bus.

[0074] Reference Figure 6 Step S4 specifically includes the following sub-steps: In the manufacturing process of printed packaging, the printing ink on the substrate and the conductive substrate of the embedded radio frequency antenna are in similar physical environments. Temperature and humidity fluctuations and mechanical imprinting stress during the operation of the manufacturing equipment simultaneously affect these two physical media. Therefore, within the system's theoretical framework, there is a certain physical coupling relationship between the physical morphological degradation at the visual level and the impedance parameter bias at the radio frequency level. While the characteristic parameters of a single mode are at risk of being independently counterfeited using existing counterfeiting methods, the cross-modal physical collaborative relationships formed based on the manufacturing process are difficult to replicate independently. Based on this physical law, the collaborative computing module 400 uses matrix operations to mathematically represent this physical collaborative relationship.

[0075] Step S41, the collaborative computing module 400 receives the visual physical morphology feature vector output by the feature extraction module 300. With radio frequency physical degradation feature vector The collaborative computing module 400 performs data validity checks on the two input feature vectors, detecting whether the vectors contain non-numerical data or infinite outliers. If an outlier is detected, the system marks the current verification process as invalid acquisition, outputs a rescan prompt to the operation interface, and suspends the current computing thread to prevent abnormal data overflow from causing system crashes; if the data is valid, the collaborative computing module 400 calculates the visual physical morphology feature vector. With radio frequency physical degradation feature vector The outer product of the vectors generates the initial cross feature matrix. The outer product operation maps the independent one-dimensional eigenvectors into a two-dimensional matrix space. The calculation formula is as follows: ; In the formula, the initial cross feature matrix Each element within represents a direct cross-product of visual feature variables and radio frequency feature variables; Represents a visual physical morphological feature vector; Represents the characteristic vector of radio frequency physical degradation; Indicates the ink bleeding and diffusion index; This indicates that the variance of the network points has increased; This indicates the residual impedance phase shift. This represents the signal timing fluctuation parameters. As a specific example, and All are 2×1 column vectors, and the outer product generates the initial cross feature matrix. It is a 2×2 square matrix. Since the dimensions and numerical magnitudes of the various physical feature parameters differ, directly using this cross-feature matrix may cause gradient skew in the network weights of subsequent classification models. The collaborative computing module 400 introduces a local covariance weight matrix. For the initial cross feature matrix Perform the Hadamard product operation to generate the cross-modal physical cooperability matrix. The specific calculation formula is as follows: ; In the formula, Represents the initial cross feature matrix; Represents the local covariance weight matrix; Represents the cross-modal physical cooperability matrix; Represents the Hadamard product. Local covariance weight matrix. This is used to eliminate dimensional differences between various physical parameters. The specific numerical matrix is ​​pre-obtained by the cloud data platform through statistical analysis of the feature dataset of genuine anti-counterfeiting label samples from the same batch. As one feasible approach, the cloud server extracts feature vectors from a large number of genuine samples to construct a prior sample set. Following the same feature extraction process as the edge computing platform, it generates visual physical morphology feature vectors and radio frequency physical degradation feature vectors, and further generates corresponding cross-feature matrix sample sets. Based on this cross-feature matrix sample set, the cloud server calculates the covariance distribution of the cross terms between visual features and radio frequency features, and then performs normalization, inversion, regularization, or scaling processing on this covariance distribution to generate a matrix... Local covariance weight matrix of the same dimension Local covariance weight matrix By establishing a binding relationship with the production batch, label format, or process parameter range of the target label, and sending it to the edge computing device as part of the batch baseline parameter set, or pre-caching it in the edge computing device, the weight matrix can amplify the cross-features with physical consistency in the genuine product sample and suppress irrelevant environmental noise variables.

[0076] Step S42, the collaborative computing module 400 performs a computation on the generated two-dimensional cross-modal physical coordination degree matrix. Perform a dimension flattening operation. This operation extracts each element of the matrix in row-major order, transforming the 2×2 cross-modal physical cooperability matrix. Dimensionality reduction to a 4×1 dimension one-dimensional collaborative feature vector .

[0077] The collaborative computing module 400 processes the original visual-physical feature vectors according to the predefined data structure order. Radio frequency physical degradation feature vector and the flattened synergy feature vector Perform column-wise concatenation to generate a high-dimensional joint feature tensor. High-dimensional joint feature tensor The mathematical structure is expressed as follows: ; In the formula, Represents a high-dimensional joint feature tensor; Represents a visual physical morphological feature vector; Represents the characteristic vector of radio frequency physical degradation; Represents a one-dimensional eigenvector of synergy; superscript This represents the transpose operation of a vector; This high-dimensional joint feature tensor It not only includes the physical degradation state parameters of the target tag in a single optical and radio frequency domain, but also integrates cross-modal cooperative parameters characterizing the coupling relationship of manufacturing processes. Combining the derivations of the aforementioned dimensions, It is a one-dimensional tensor containing 8 elements. As an implementable approach, high-dimensional joint feature tensors... The elements in the high-dimensional joint feature tensor are fixed in order as follows: ink diffusion index, dot gain variance, residual impedance phase shift, signal timing fluctuation parameters, and four cross-modal physical coordination degree matrix elements expanded in row priority. Before writing to the cache queue, the collaborative computing module 400 performs dimensionality and numerical range checks on the high-dimensional joint feature tensor to ensure that its input dimension and element order are consistent with the input layer structure of the classification model inside the edge inference module 500. It is written into the dedicated video memory or cache queue of the tensor processing unit of the edge computing device and provided to the downstream classification model as the underlying input data source for network inference.

[0078] Reference Figure 7 Step S5 specifically includes the following sub-steps: In conventional fixed-boundary classification models, the system often assumes that the features of all genuine labels are distributed within the same fixed feature space. However, in actual industrial production, environmental parameters such as temperature, humidity, and imprinting force vary between different batches of labels during manufacturing. These physical fluctuations may cause reasonable spatial drift of the microscopic features of genuine labels. If a rigid decision boundary is used, it is easy to misjudge legitimate drift features located at the boundary edge. The edge inference module 500 introduces parameters representing the physical environment of manufacturing into the network structure to dynamically correct the classification decision boundary of the feature space, which can improve the generalization ability of the verification system across different manufacturing batches to a certain extent.

[0079] Step S51, the edge inference module 500 retrieves the printing process context vector cached by the multi-mode acquisition module 100. and the high-dimensional joint feature tensor generated by the collaborative computing module 400 The edge inference module 500 internally deploys a classification model based on a conditionally gated multilayer perceptron architecture. The internal topology of this classification model consists of three parts: a feature backbone network, a context branch network, and a dynamic feature modulation layer.

[0080] The edge reasoning module 500 will use high-dimensional joint feature tensors The input is fed into the feature backbone network of the classification model. The feature backbone network extracts the basic hidden layer feature vectors through fully connected linear mapping and non-linear activation operations. The specific calculation formula is as follows: ; In the formula, Represents the feature vector of the basic hidden layer; Represents a high-dimensional joint feature tensor; This represents the input layer weight matrix of the feature backbone network; as one embodiment of the present invention, it is combined with the aforementioned steps. Given an 8×1 dimension one-dimensional input tensor, The dimension corresponding to the configuration is ×8, The preset number of hidden layer neurons, for example, can be set to 32 or 64; The dimension is ×1 input layer bias vector; This represents the linear rectified activation function, used to introduce nonlinear expressiveness and alleviate the gradient vanishing phenomenon. Basic hidden layer feature vectors. It characterizes the underlying physical degradation mapping state of the current scan tag under unconstrained environmental conditions.

[0081] Step S52, the edge inference module 500 will generate the printing process context vector The input is fed into the context branch network of the classification model. Based on this environmental constraint, the context branch network calculates and generates a scaling coefficient vector for feature linear modulation. With bias coefficient vector The calculation formula is as follows: ; ; In the formula, Represents the scaling factor vector; Represents the bias coefficient vector; and These represent the mapping weight matrices for generating scaling and bias coefficients, respectively, taking into account... It is a 3×1 dimensional context vector. and All dimensions are set to ×3; and This represents the corresponding bias constant vector, with dimensions of 1. ×1. This is used to calculate the output scaling factor vector. With bias coefficient vector Dimensions and basic hidden layer feature vectors Strict alignment, all ×1.

[0082] Subsequently, the edge inference module 500 utilizes the scaling factor vector in the dynamic feature modulation layer. With bias coefficient vector For the basic hidden layer feature vector Perform an element-based affine transformation to generate a modulation feature vector. The calculation logic is as follows: ; In the formula, Represents the modulation feature vector; Represents the scaling factor vector; Represents the feature vector of the basic hidden layer; Represents the bias coefficient vector; This represents the Hadamard product operation. The mathematical principle of this affine transformation mechanism lies in using environmental parameters to perform element-wise scaling and translation operations on the backbone feature space. In this embodiment, the classification decision boundary adjustment does not directly modify the fixed threshold of the output layer, but rather generates a scaling coefficient vector through the context branch network. With bias coefficient vector And using this scaling factor vector The bias coefficient vector β is compared with the base hidden layer feature vector. Affine modulation is performed, which effectively changes the classification decision boundary of the output layer in the original feature space. When the humidity or pressure of the manufacturing environment exceeds the standard range, the scaling factor and bias factor will change the data distribution range of the backbone features accordingly. In the geometric sense of multidimensional space, this is manifested as the model adaptively stretching or twisting the classification decision boundary according to the state of the manufacturing environment, thereby accommodating the drift of legitimate features caused by the environment.

[0083] Step S53, the edge inference module 500 modulates the feature vector The data is transmitted to the output layer of the classification model. The output layer performs dimensionality reduction mapping using the logistic function and outputs the authenticity probability score of the target label. The formula is as follows: ; In the formula, This represents the probability score of the target label being genuine. Represents the modulation feature vector; and Let these represent the classification weight matrix and classification bias vector of the output layer, respectively. The dimension is 1× , For scalar bias; This represents the logistic function, used to map linear outputs in the real number field to a probability space between 0 and 1; It represents the base of the natural logarithm.

[0084] Edge reasoning module 500 vs. genuine product probability score Compared with the preset confidence threshold The confidence threshold The specific value is determined by technicians during the factory calibration phase, based on the cross-equivalence point between the false positive and false negative rates in the test set, and is typically set between 0.80 and 0.90. If the product's probability score is... Greater than or equal to the confidence threshold The edge inference module 500 determines that the physical coordination characteristics of the target tag conform to the original manufacturing rules, and outputs a discrete business status indication of genuine product release to the interactive interface of the edge computing device; if the genuine product probability score is... Less than this confidence threshold If the cross-modal physical characteristics are decoupled, the device will output a discrete business status indication of the fake warning to the interactive interface and drive the device's general input / output pins to trigger the audible and visual alarm module.

[0085] To avoid algorithm crashes caused by network transmission latency, the edge inference module 500 employs a bypass fault tolerance mechanism before executing step S51. Given that the edge computing device has already cached the standard reference phase, local covariance weight matrix, reference template parameters, and classification model parameters for the production batch to which the target label belongs, if the printing process context vector is affected due to cloud communication anomalies... If the value is null, the edge inference module 500 will scale the coefficient vector. Set as a one vector, bias coefficient vector The vector is set to all zeros. At this point, the model mathematically degenerates into a fixed-boundary classifier, ensuring the basic continuity of the verification process in a weak network environment. If the edge computing device does not cache the aforementioned batch baseline parameters, or if the batch baseline parameters cannot match the identification information of the current target label, the edge inference module 500 suspends the current verification process and outputs a low-confidence verification status or a reconnection request prompt to the interactive interface. It should be understood that the edge inference module 500 also performs parallel verification of the high-dimensional joint feature tensor when receiving data. The module ensures the integrity of feature dimensions. If a missing dimension is detected, it will directly interrupt the current network inference process and throw a data acquisition anomaly flag to the system to prevent system crashes caused by out-of-bounds access to the underlying array.

[0086] In step S54, before the classification model is deployed to the edge computing device, its internal network weight parameters are pre-trained and iterated on the cloud computing server. For the execution logic of model building and training, the cloud computing server collects genuine labels and counterfeit labels made using known imitation processes from the manufacturing line and quality inspection stages, simultaneously recording the corresponding printing process environment parameters to construct a multimodal sample dataset. For each training sample, the cloud computing server generates a high-dimensional joint feature tensor using the same multimodal acquisition, spatial compensation, feature extraction, and collaborative computing process as the edge computing device, and uses the printing process context vector corresponding to that training sample as the context input, ensuring that the dimension, element order, and normalization method of the input data during the training stage are consistent with those during the edge inference stage.

[0087] During the training phase, the cloud computing server sets the data sample labels for genuine labels to a constant 1, and the data sample labels for counterfeit labels to a constant 0. The training process uses a binary cross-entropy loss function with an added L2 regularization term to calculate the residual gradient between the network output and the real sample labels. The calculation formula is: ; In the formula, Represents the loss function; This represents the number of mini-batch samples in a single training iteration. Indicates the first The true label value of each sample; The pixel index variable represents the pixel number being calculated. One edge pixel; This represents the predicted probability calculated from the network's forward propagation. Represents the L2 regularization coefficient, used to constrain the weight matrix. Various parameters The modulus length is adjusted to avoid overfitting of the model to the training set. This represents the log probability that the model predicts the sample to be a fake (negative class). Indicates the variable currently involved in the calculation. It belongs to the model weight set One of the specific weight parameters; Represents weight parameters The square of.

[0088] The cloud computing server uses the Adaptive Moments Estimation (Adam) optimization algorithm to update network parameters in reverse based on the residual gradient, with an initial learning rate typically set to 0.001. After the loss function converges to a stable state, the cloud computing server truncates the network structure and compiles it into a static computation graph, which is then downloaded to the edge inference module 500 for local computation via over-the-air (OTA) download technology. When downloading the static computation graph, the cloud computing server simultaneously generates a model version identifier and binds this identifier to the corresponding feature input dimension, element order, normalization parameters, and applicable production batch range. Before performing classification inference, the edge inference module 500 verifies whether the model version identifier in the current batch's baseline parameter set is consistent with the locally loaded classification model; if they are inconsistent, it triggers a model update request or outputs a model version mismatch warning.

[0089] Specific application examples: As a specific application scenario, suppose the verification system is deployed at the quality inspection terminal of a pharmaceutical packaging production line. The target label is made of PET substrate, with an embedded UHF RFID aluminum etched antenna, and a QR code containing an anti-counterfeiting UID printed on its surface.

[0090] Phase 1: Data Acquisition and Environmental Parameter Acquisition. The multi-mode acquisition module controls the terminal device to scan a specific target label in the current batch. The cloud-based data platform returns the printing process context vector of that batch based on the parsed UID. Assuming the normalized value of the measured ambient humidity at the factory is 0.65, the normalized value of the printing pressure is 0.80, and the substrate porosity index is 0.50, then the constructed printing process context vector... .

[0091] The second stage: feature extraction. After pose correction by the spatial compensation module, the feature extraction module calculates the ink diffusion index in the visual domain. Increased variance of outlets Generate visual physical morphology feature vectors .

[0092] Simultaneously extract residual impedance phase shift in the radio frequency domain Radius, signal timing fluctuation parameter Generate radio frequency physical degradation feature vector .

[0093] Phase 3: Collaborative computation. The collaborative computation module calculates the outer product of the two modality vectors mentioned above to generate the initial cross-feature matrix. : ; The system retrieves the local covariance weight matrix of the same batch distributed from the cloud. Perform the Hadamard product to eliminate dimensionless differences: ; Flatten the matrix row-wise and concatenate it with the basic features to generate an 8-dimensional high-dimensional joint feature tensor. .

[0094] Phase Four: Dynamic Classification and Decision Making, Edge Reasoning Module will Input the feature backbone network of the classification model and extract the basic hidden layer feature vectors. Meanwhile, the context branch network receives... The scaling factor vector is calculated based on the pre-trained weights. With bias coefficient vector .

[0095] Because the humidity (0.65) and pressure (0.80) of the current batch are higher than the baseline values, the scaling factor generated by the model is adaptively adjusted. The corresponding component in the hidden layer will be greater than 1, thus affecting the hidden layer. Perform stretch transformation ( This operation mathematically expands the tolerance for legitimate product decisions under current environmental conditions.

[0096] The final output layer calculates the probability score of genuine products. Preset reliability threshold System judgment The system outputs a release instruction for genuine products. If a rigid model without environmental modulation is used, the score of this slightly contaminated sample may be only 0.81, leading to a false alarm (false negative).

[0097] Reference Figure 8 , Figure 8 This is a performance comparison curve of receiver operating characteristic (ROC) of different verification algorithms according to an embodiment of the present invention. The curve shows the overall performance differences of four different verification algorithm configurations on a large-scale test set containing complex manufacturing batch spans and multiple known forgery methods.

[0098] The horizontal axis of the coordinate system in the figure represents the false positive rate (FPR), which indicates the probability that the system incorrectly identifies counterfeit labels as genuine products and allows them to pass through, reflecting the anti-counterfeiting system's ability to intercept counterfeit products; the vertical axis represents the true positive rate (TPR), which indicates the probability that the system correctly identifies genuine product labels as genuine products and reflects the anti-counterfeiting system's tolerance for the release of legitimate products.

[0099] To verify the effectiveness of the technical solution of this invention across complex manufacturing batch spans, a verification dataset containing 20,000 test samples was constructed. This dataset includes 10,000 genuine samples (distributed across four production batches with environmental differences, such as high humidity in spring and low humidity in winter); and 10,000 counterfeit samples (covering three counterfeiting methods: high-precision optical copying, RF chip cloning, and independent splicing of visual and RF components).

[0100] The experiment included four control groups: Single vision model (conventional): Only extracts ink dot features and uses fixed classification boundaries.

[0101] Single RF model (conventional): Only impedance phase and signal strength features are extracted, and a fixed classification boundary is used.

[0102] Dual-modal unconstrained model: Visual and radio frequency features are extracted and stitched together, but the printing process context vector is not introduced for boundary modulation.

[0103] The present invention proposes to introduce a cross-modal physical coherence feature tensor and utilize the printing process context vector to perform dynamic feature modulation.

[0104] Combination Figure 8 The receiver operating characteristic (ROC) performance comparison curve shown is illustrated. The horizontal axis of the curve represents the false positive rate (FPR, the probability of misidentifying a counterfeit product as a genuine product), and the vertical axis represents the true positive rate (TPR, the probability of correctly identifying a genuine product). The closer the curve is to the upper left corner, the better the overall anti-counterfeiting performance of the system.

[0105] Experimental data show that when encountering cross-batch samples, both the single visual model (dotted line) and the single radio frequency model (dotted line) experience overall curve collapse due to legitimate drift of physical features exceeding fixed boundaries, resulting in area under the curve (AUC) values ​​of only 0.86 and 0.82, respectively. The bimodal unconstrained model (dashed line) shows improved performance (AUC=0.91) by introducing feature fusion, but its true positive rate struggles to exceed 0.90 in the lower false positive rate range (FPR < 0.05), indicating a significant problem of excessively high kill rate.

[0106] As can be seen from the comparison, the present invention (solid line) eliminates on-site reading errors through dynamic spatial pose compensation, characterizes cross-modal physical synergy using the outer product and local covariance, and adaptively adjusts the decision boundary through manufacturing environment constraints. Under strict quality control conditions that ensure a false positive rate of less than 0.02, the present invention can still maintain a true positive rate of over 98.5% (AUC=0.98), encompassing legitimate manufacturing errors and intercepting piecemeal counterfeit samples, thus possessing engineering application value.

Claims

1. A verification system for anti-counterfeiting labels on printed packaging products based on edge AI, characterized in that, include: The multi-mode acquisition module is used to control the image sensor and the UHF RFID read / write hardware to perform synchronous sampling operations, acquire the digital image matrix of the target tag and the original low-level RF sequence, and acquire the printing process context vector through the network communication interface. The printing process context vector represents the parameter set of the objective physical manufacturing environment in which the target tag is printed at the factory. The spatial compensation module is used to calculate the spatial relative pose parameters between the edge computing device and the physical plane of the target tag, and accordingly perform spatial compensation on the original underlying radio frequency sequence to generate a normalized radio frequency sequence. The feature extraction module is used to extract and generate visual physical morphology feature vectors from the digital image matrix, and simultaneously extract and generate radio frequency physical degradation feature vectors from the normalized radio frequency sequence. The collaborative computing module receives the visual physical morphology feature vector and the radio frequency physical degradation feature vector, and uses them to calculate the cross-modal collaborative correlation between the two feature vectors and concatenate them to generate a high-dimensional joint feature tensor. The edge reasoning module has a pre-built classification model, which is used to configure the printing process context vector as the spatial decision boundary adjustment parameter of the classification model, and input the high-dimensional joint feature tensor into the classification model after boundary adjustment, and output discretized decision result variables.

2. The edge AI-based anti-counterfeiting label verification system for printed packaging products according to claim 1, characterized in that, The multi-mode acquisition module is specifically used for: The system's global clock generator generates edge-aligned hardware trigger pulse signals, which are then sent to the global shutter control pin of the image sensor and the baseband control pin of the UHF RFID read / write hardware, respectively. The identification information of the target tag is parsed from the digital baseband payload of the radio frequency communication signal, and the original backscattering time sequence of the physical layer is extracted to generate the original bottom layer radio frequency sequence. The identification information is sent to the cloud data platform, and the printing process context vector returned by the cloud data platform based on the identification information is received.

3. The edge AI-based anti-counterfeiting label verification system for printed packaging products according to claim 1, characterized in that, The specific process by which the spatial compensation module calculates the spatial relative pose parameters includes: Extract the region of interest containing the target label identification information from the digital image matrix, extract the set of observation anchor points within the region of interest, and retrieve the pre-stored set of physical world reference points; Based on the perspective mapping relationship between the set of observation anchor points and the set of physical world reference points, the homography matrix is ​​calculated; After the homography matrix verification is valid, matrix decomposition is performed on the homography matrix in conjunction with the camera's intrinsic parameter matrix to calculate the spatial relative pose vector of the edge computing device relative to the physical plane of the target tag.

4. The edge AI-based anti-counterfeiting label verification system for printed packaging products according to claim 3, characterized in that, The specific process by which the spatial compensation module generates the normalized radio frequency sequence includes: Using the distance and angle parameters in the spatial relative pose vector, the equivalent propagation distance from the phase center of the radio frequency antenna to the tag response center is calculated; Based on the equivalent propagation distance, reverse spatial phase compensation calculation is performed on each frame of the original phase observation value in the original underlying radio frequency sequence to generate a normalized phase. Based on the spatial relative pose vector and the real-time operating frequency, reverse free space loss and antenna pattern gain compensation calculations are performed on the original received signal strength indication value of each frame in the original underlying radio frequency sequence to generate a normalized received signal strength. The normalized phase and the normalized received signal strength are combined to generate the normalized radio frequency sequence.

5. The edge AI-based anti-counterfeiting label verification system for printed packaging products according to claim 3, characterized in that, The specific process by which the feature extraction module generates visual physical morphology feature vectors includes: A standardized label image is generated by performing an inverse perspective transformation on the digital image matrix using a homography matrix, and the region of interest for ink verification is extracted. The gradient change amplitude of each pixel in the region of interest of the ink verification is calculated, and it is spatially compared with the pre-stored non-destructive reference template image to calculate and generate the ink staining diffusion index. The statistical variance of the halftone dot area coverage rate within the local spatial neighborhood of the ink verification region of interest is statistically analyzed to generate the dot magnification variance. The ink diffusion index and the dot gain variance are combined to generate the visual physical morphology feature vector.

6. The edge AI-based anti-counterfeiting label verification system for printed packaging products according to claim 1, characterized in that, The specific process by which the feature extraction module generates radio frequency physical degradation feature vectors includes: Extract the phase data within all valid sampled frames of the normalized radio frequency sequence and calculate its residual impedance phase drift relative to the standard reference phase. Perform time-series variance calculation on the normalized received signal strength sequence within the normalized radio frequency sequence to generate signal time-series fluctuation parameters that reflect the dispersion of radio frequency backscatter energy; The residual impedance phase drift is combined with the signal timing fluctuation parameters to generate the radio frequency physical degradation feature vector.

7. The edge AI-based anti-counterfeiting label verification system for printed packaging products according to claim 1, characterized in that, The specific process by which the collaborative computing module generates the high-dimensional joint feature tensor includes: Calculate the cross product of the visual physical morphology feature vector and the radio frequency physical degradation feature vector to generate an initial cross feature matrix; A preset local covariance weight matrix is ​​introduced, and a Hadamard product operation is performed on the initial cross feature matrix to generate a cross-modal physical coherence matrix for eliminating dimensional differences and amplifying physical consistency cross feature terms. Perform a dimension flattening operation on the cross-modal physical synergy matrix to generate a one-dimensional synergy feature vector; The visual physical morphology feature vector, the radio frequency physical degradation feature vector, and the one-dimensional synergy feature vector are concatenated and spliced ​​in the column direction to generate the high-dimensional joint feature tensor.

8. The edge AI-based anti-counterfeiting label verification system for printed packaging products according to claim 1, characterized in that, The edge inference module internally deploys a classification model based on a conditionally gated multilayer perceptron architecture. This classification model includes a feature backbone network, a context branch network, and a dynamic feature modulation layer. The specific decision-making process includes: The high-dimensional joint feature tensor is input into the feature backbone network to extract the basic hidden layer feature vector; The printing process context vector is input into the context branch network to calculate and generate the scaling factor vector and the bias factor vector; In the dynamic feature modulation layer, the scaling coefficient vector and the bias coefficient vector are used to perform an element-based affine transformation on the basic hidden layer feature vector to generate a modulation feature vector.

9. The edge AI-based anti-counterfeiting label verification system for printed packaging products according to claim 8, characterized in that, The edge reasoning module also includes the following feature process when performing classification decisions: The modulation feature vector is transmitted to the output layer of the classification model to perform dimensionality reduction mapping, and the authenticity probability score of the target label is output. Compare the genuine product probability score with a preset confidence threshold. If the score is greater than or equal to the confidence threshold, output a discrete business status indication for genuine product release; if the score is less than the confidence threshold, output a discrete business status indication for counterfeit warning. When the printing process context vector becomes null due to cloud communication anomalies, the edge inference module triggers a bypass fault tolerance mechanism, sets the scaling coefficient vector to an all-one vector and the bias coefficient vector to an all-zero vector, causing the classification model to degenerate into a fixed-boundary classifier.

10. A method for verifying anti-counterfeiting labels on printed packaging products based on edge AI, applied to the system described in any one of claims 1-9, characterized in that, Includes the following steps: S1. Perform the steps of synchronous acquisition of multimodal anti-counterfeiting signals and acquisition of manufacturing parameters. Control the equipment to perform synchronous sampling operations to acquire the digital image matrix and original underlying radio frequency sequence of the target label, and obtain the printing process context vector from the cloud. S2. Perform a dynamic spatial compensation step for radio frequency based on visual anchor points, calculate the spatial relative pose parameters between the edge computing device and the physical plane of the target tag, and perform spatial compensation calculation on the original bottom-layer radio frequency sequence accordingly to generate a normalized radio frequency sequence. S3. Perform a deterministic extraction step of dual physical domain micro-degradation features, and extract and generate visual physical morphology feature vector and radio frequency physical degradation feature vector from the digital image matrix and the normalized radio frequency sequence, respectively. S4. Perform the isomorphic cross-modal physical synergy feature tensor construction step, calculate the cross-modal physical synergy correlation between the two feature vectors and concatenate them to generate a high-dimensional joint feature tensor; S5. Execute the dynamic classification decision step based on manufacturing environment constraints, configure the printing process context vector as the boundary adjustment constraint condition of the internal preset classification model, perform nonlinear classification mapping calculation on the high-dimensional joint feature tensor through the classification model after boundary adjustment, and output the label anti-counterfeiting judgment result.

Images