A packaging paper humidity detection method based on infrared imaging temperature distribution analysis

Abstract

This invention discloses a method for detecting packaging paper humidity based on infrared imaging temperature distribution analysis. This method relates to the field of humidity detection technology. The method includes triggering a programmable metasurface filter array to dynamically select wavelengths based on the type of packaging paper, collecting multi-band infrared radiation data from the packaging paper surface; employing a dual generative adversarial network to generate a temperature-humidity mapping dataset; processing the central flat region of the packaging paper using the Poincare disk model and the edge abrupt region using the Lorentz model, extracting a mixed feature tensor within a hyperbolic Euclidean mixing space; and inferring humidity distribution using a hyperbolic neural network containing a dynamic curvature convolutional layer and a hyperbolic attention mechanism to output a surface humidity distribution matrix. By simultaneously utilizing the dual generative adversarial network to simulate humidity diffusion paths and generate a noise-resistant temperature field, the method enhances adaptability to different packaging paper materials and microstructures, thereby improving the accuracy of packaging paper humidity detection.

Description

Technical Field

[0001] The present invention relates to the technical field of humidity detection, in particular to a packaging paper humidity detection method based on infrared imaging temperature distribution analysis. Background Art

[0002] With the growing demand for logistics, storage, and industrial packaging, wrapping paper plays a key role in protecting various products. The moisture content of wrapping paper directly affects its mechanical properties and moisture-proofing effectiveness. Due to differences in its material and structure, the moisture distribution of wrapping paper itself is uneven and time-varying. Traditional humidity detection methods, which often rely on contact measurement, suffer from low accuracy, response delays, and environmental interference, making them difficult to meet the requirements of modern production and quality control.

[0003] Furthermore, non-contact detection technology has gained increasing attention in recent years. By monitoring the temperature distribution on the surface of wrapping paper, this technology indirectly reflects humidity information. However, practical applications still face numerous challenges. Significant differences in manufacturing processes, surface textures, and hygroscopicity among different wrapping papers complicate signal acquisition and analysis during the detection process. Improving the accuracy and stability of data processing while maintaining detection efficiency has become a critical issue for both research and industry. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a packaging paper humidity detection method based on infrared imaging temperature distribution analysis to solve the problems of low detection accuracy and poor adaptability to packaging paper materials in traditional packaging paper humidity detection.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a method for detecting moisture content in packaging paper based on infrared imaging temperature distribution analysis, which includes triggering a programmable metasurface filter array to perform dynamic band selection based on the type of packaging paper, and collecting multi-band infrared radiation data from the packaging paper surface;

[0008] A dual generative adversarial network is used to generate a temperature and humidity mapping dataset, where the first generator simulates the humidity diffusion path and the second generator generates a noise-resistant temperature field through magnetorheological thermal excitation.

[0009] The Poincare disk model is used to treat the flat area in the center of the wrapping paper, and the Lorentz model is used to treat the edge mutation area. The mixing characteristic tensor is extracted in the hyperbolic Euclidean mixing space.

[0010] Moisture distribution inference is performed through a hyperbolic neural network, and a moisture distribution matrix of the packaging paper surface is output. The hyperbolic neural network includes a dynamic curvature convolution layer and a hyperbolic attention mechanism.

[0011] As a preferred solution of the packaging paper humidity detection method based on infrared imaging temperature distribution analysis described in the present invention, the first generator combines the packaging paper material properties and the humidity diffusion partial differential equation constraints to generate a humidity diffusion path that matches the microstructure of the packaging paper.

[0012] As a preferred solution of the packaging paper humidity detection method based on infrared imaging temperature distribution analysis described in the present invention, the second generator generates a non-uniform thermal excitation field by regulating the magnetic field distribution, and suppresses the temperature field noise in combination with the heat conduction equation.

[0013] As a preferred solution of the packaging paper moisture detection method based on infrared imaging temperature distribution analysis described in the present invention, the dynamic band selection adjusts the transmission band combination of the filter array in real time by analyzing the spectral absorption characteristics of the packaging paper, and compensates the band weight based on ambient temperature fluctuations.

[0014] As a preferred solution of the packaging paper moisture detection method based on infrared imaging temperature distribution analysis of the present invention, wherein: the extraction of the mixed feature tensor in the hyperbolic Euclidean mixed space is carried out in the following specific steps:

[0015] The infrared radiation data of the flat area in the center of the wrapping paper is mapped to the geodesic coordinate system of the Poincare disk model to extract the curvature characteristics in the hyperbolic space.

[0016] The infrared radiation data of the edge mutation area is input into the Lorentz model for gradient field decomposition to extract the mutation characteristics in Euclidean space;

[0017] The hyperbolic and Euclidean features are fused through tensor concatenation and normalization operations to generate a hybrid feature tensor.

[0018] As a preferred solution of the packaging paper humidity detection method based on infrared imaging temperature distribution analysis described in the present invention, the dynamic curvature convolution layer adaptively adjusts the geometric shape of the convolution kernel through the curvature change of the local feature manifold, and dynamically constrains the curvature parameters using the hyperbolic tangent function.

[0019] As a preferred solution of the packaging paper moisture detection method based on infrared imaging temperature distribution analysis described in the present invention, the hyperbolic attention mechanism generates attention weights by calculating the geodesic distance of the mixed feature tensor in the Poincare disk, and performs nonlinear weighted fusion of feature channels based on hyperbolic space projection.

[0020] As a preferred solution of the packaging paper humidity detection method based on infrared imaging temperature distribution analysis described in the present invention, the packaging paper surface humidity distribution matrix is used to generate a visual humidity distribution map and perform humidity detection.

[0021] In a second aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the packaging paper moisture detection method based on infrared imaging temperature distribution analysis as described in the first aspect of the present invention is implemented.

[0022] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, any step of the packaging paper moisture detection method based on infrared imaging temperature distribution analysis as described in the first aspect of the present invention is implemented.

[0023] The beneficial effects of the present invention are as follows: by adopting a programmable metasurface filter array to realize dynamic band selection, the multi-band infrared radiation data of the packaging paper surface is accurately captured, thereby improving the accuracy of humidity detection; at the same time, a dual generative adversarial network is used to simulate the humidity diffusion path and generate a noise-resistant temperature field, thereby enhancing the adaptability to different packaging paper materials and microstructures; by applying the Poincare disk model and the Lorentz model to the central flat area and the edge mutation area of the packaging paper respectively, and extracting the mixed feature tensor in the hyperbolic Euclidean mixed space, the fine modeling of complex temperature and humidity distribution is realized; combining the efficient nonlinear feature processing of the dynamic curvature convolution layer and the hyperbolic attention mechanism in the hyperbolic neural network, not only the computational efficiency is optimized, but also the anti-interference ability is improved; the humidity distribution matrix finally output can be directly used to generate intuitive visualization graphics, providing reliable technical support for quality control and detection decision-making, and has certain industrial integration and application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0025] Figure 1 This is the overall framework diagram of packaging paper humidity detection using infrared imaging temperature distribution analysis in Example 1.

[0026] Figure 2 This is a schematic diagram of the packaging paper humidity detection process using infrared imaging temperature distribution analysis in Example 1.

[0027] Figure 3Schematic diagram of the relationship between infrared imaging and humidity mapping in Example 1.

[0028] Figure 4 Schematic diagram of the hyperbolic neural network structure in Example 1. DETAILED DESCRIPTION

[0029] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0030] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0031] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0032] Example 1, with reference to Figures 1 to 4 This embodiment provides a packaging paper moisture detection method based on infrared imaging temperature distribution analysis, comprising the following steps:

[0033] S1: Trigger the programmable metasurface filter array to perform dynamic band selection based on the type of wrapping paper and collect multi-band infrared radiation data from the wrapping paper surface.

[0034] Specifically, the following steps are included:

[0035] S1.1: Perform packaging paper category identification and spectral characteristic matching.

[0036] Specifically, set the wrapping paper category code ,in The total number of preset wrapping paper types. Stores the absorption spectrum of each type of wrapping paper in the infrared band (3-14μm) . According to the input wrapping paper category code Retrieve the corresponding absorption spectrum , extract its characteristic absorption peak band set ,in The number of effective bands , is the absorption spectrum, Code the category of wrapping paper, is the wavelength, is a collection of characteristic absorption peak bands.

[0037] It should be noted that the absorption spectrum This refers to the infrared absorption spectrum of packaging paper materials in the infrared band (3-14μm), specifically expressed as the absorption rate of the packaging paper material to infrared radiation at different wavelengths. This is generated by measuring the absorption characteristics of the packaging paper material in the infrared band using an infrared spectrometer (such as a Fourier transform infrared spectrometer) and matching characteristic peaks with standard databases (such as the Sadtler spectrum library).

[0038] S1.2: Implement dynamic band selection.

[0039] Specifically, from the absorption peak band collection The optimal band combination is dynamically selected to maximize humidity sensitivity and environmental robustness.

[0040] For each candidate band , calculate its humidity sensitivity factor and environmental interference factor, the expression is:

[0041]

[0042] in, is the humidity sensitivity factor, Environmental interference factors, is the band number index, For wrapping paper category In the characteristic absorption peak band The infrared radiation absorption rate at is the humidity value of the sample packaging paper, For band The half-peak width, is the width suppression coefficient , The ambient temperature is in the band The radiation variance caused by is the ambient temperature, Characteristic absorption peak band The inverse of the signal-to-noise ratio, is the radiation variance weight (can be 0.6), is the signal-to-noise ratio weight (can be 0.4, adjusted through experiments).

[0043] It should be noted that the humidity value of the sample packaging paper refers to the moisture content of the sample packaging paper calculated by the oven drying method. The corresponding sample is selected according to the type of packaging paper.

[0044] Combining humidity sensitivity factor and environmental interference factor to perform dynamic band selection, the band optimization function expression is:

[0045]

[0046] in, is the optimal value of the band, is the natural exponential function, is the reference wavelength (8-10μm), is the wavelength deviation penalty coefficient (can be 2.5).

[0047] Select The top 3 bands are used as active bands.

[0048] S1.3: Control a programmable metasurface filter array.

[0049] It should be understood that the metasurface filter array is composed of multiple metasurface units, and each unit can dynamically adjust the transmission wavelength by applying voltage.

[0050] Specifically, the control strategy is as follows: the metasurface filter array is divided into multiple sub-areas according to fixed rows and columns, and each sub-area is assigned an activation band; the transmission wavelength of the metasurface unit at different voltages is measured experimentally, and a lookup table is established. For the target band corresponding to each sub-area, the lookup table is reversely queried to obtain the corresponding voltage, and the activation band is voltage mapped to obtain the driving voltage of each sub-area; when the number of activated bands is greater than 1, the bands are cyclically switched according to the time slice for sequential polling.

[0051] S1.4: Perform ambient temperature compensation.

[0052] Specifically, after collecting the raw infrared radiation intensity from the wrapping paper surface, the system dynamically compensates for the effects of ambient temperature using the principle of blackbody radiation. A sensor measures the current ambient temperature. In an unloaded state without wrapping paper, baseline background radiation values for each band at different ambient temperatures are pre-recorded to create a temperature-radiation mapping database. During actual testing, the baseline ambient radiation value for the corresponding band is retrieved from the database based on the current ambient temperature. This baseline value is treated as interference from ambient heat sources (such as heat generated by the device itself and air convection) and removed. Finally, the system outputs the compensated infrared radiation data.

[0053] S1.5: Collect multi-band infrared radiation data from the packaging paper surface.

[0054] It should be understood that the infrared radiation data after ambient temperature compensation and the coordinates of the infrared image are stored in a three-dimensional tensor to obtain multi-band infrared radiation data of the packaging paper surface.

[0055] Preferably, the corresponding infrared absorption spectrum is matched according to the packaging paper category code, the characteristic absorption peak band is accurately extracted, and the optimal band combination is dynamically screened in combination with humidity sensitivity and environmental robustness to improve the humidity detection sensitivity and reduce the impact of environmental interference. A programmable metasurface filter array is used to dynamically adjust the transmission wavelength through voltage control, and the target band is efficiently activated according to the sub-region allocation and time slice polling strategy, thereby improving the adaptability and real-time performance of multi-band information acquisition. During the data acquisition process, ambient temperature compensation is performed in combination with the blackbody radiation formula to correct the impact of temperature fluctuations on infrared radiation data and ensure the stability and accuracy of the measurement results. The compensated infrared radiation data is integrated with the infrared image coordinates, and the multi-band radiation data is stored in matrix form to provide high-quality input for subsequent humidity distribution reasoning based on deep learning.

[0056] S2: Generate temperature and humidity mapping dataset using dual generative adversarial networks.

[0057] Specifically, the following steps are included:

[0058] S2.1: Construct a dual-generator adversarial network.

[0059] Specifically, the dual-generator adversarial network includes a first generator, a second generator, a first adversary, a second adversary and a feature interaction unit.

[0060] The first generator generates a humidity distribution field based on the microstructure of the packaging paper and the humidity diffusion physical model.

[0061] The second generator generates a noise-resistant temperature field based on a magnetorheological thermal excitation and heat conduction model.

[0062] The first adversary determines the authenticity of the humidity distribution field.

[0063] The second antagonist determines the authenticity of the anti-noise temperature field.

[0064] The feature interaction unit realizes feature coupling between the first generator and the second generator through cross-modal attention.

[0065] Furthermore, the input of the dual generator adversarial network is the multi-band infrared radiation data of the wrapping paper surface, the magnetic field distribution on the wrapping paper surface and the wrapping paper material parameters.

[0066] Finite element simulation is used to obtain the magnetic field distribution on the wrapping paper surface. First, the computational domain is established and parameters are set based on the wrapping paper material properties. A non-uniform adaptive mesh is used to optimize computational accuracy. During the simulation, an external alternating magnetic field is applied as an excitation source, and boundary conditions such as zero scalar potential and absorbing boundaries are set. The finite element method is used to calculate the steady-state or alternating magnetic field distribution at different locations on the wrapping paper surface. The magnetic field data is then interpolated and smoothed. The resulting magnetic field distribution matrix is then generated.

[0067] Wrapping paper material parameters include porosity, fiber orientation and thickness.

[0068] The output is the normalized humidity distribution matrix and temperature field matrix.

[0069] S2.2: Modeling of the first generator humidity diffusion path.

[0070] Specifically, the encoder uses 5 layers of 3D convolution to extract the geometric features of the wrapping paper material parameters.

[0071] The implicit finite difference method is used to discretize the PDE, construct the residual constraint term, and add the physical residual to the loss function. The expression is:

[0072]

[0073] in, is the physical residual loss, is the total number of samples for calculating physical residuals, is the time variable, is the humidity diffusion term based on Friedel-Crafts law, is the gradient, is the diffusion coefficient, is the material parameter of the wrapping paper, is the spatial horizontal coordinate, is the spatial ordinate, The humidity source term generated by the first generator.

[0074] The decoder uses transposed convolution to generate the humidity distribution matrix, and the output layer is Sigmoid activation.

[0075] S2.3: Second generator anti-noise temperature field generation.

[0076] Specifically, the multi-band infrared radiation data of the wrapping paper surface is input and the multi-band features are extracted by the Inception unit, and the magnetic field distribution on the wrapping paper surface is encoded into a spatial modulation vector through the fully connected layer.

[0077] The thermal excitation power density is regulated based on the magnetic field distribution on the packaging paper surface, and the heat conduction equation constraint is embedded in the temperature field generation. The expression is:

[0078]

[0079] in, is the surface temperature of the wrapping paper, is the thermal diffusivity, is the thermal excitation power density, is the material density, is the specific heat capacity, is the noise suppression factor (optimized by adversarial training), is a Gaussian noise field, Thermal conversion efficiency coefficient , is the vacuum permeability, is the magnetic susceptibility.

[0080] Using the U-Net structure, the jump connection injects the magnetic field modulation information. The output layer uses Tanh activation and maps it to the temperature range. .

[0081] S2.4: Set up adversarial training strategy.

[0082] It should be understood that the discrimination loss is generated based on the surface temperature of the wrapping paper generated by the second generator and the surface humidity of the wrapping paper generated by the first generator, and a weighted sum is performed to establish a total loss function for joint optimization.

[0083] The total loss function includes the first adversary discriminant loss, the second adversary discriminant loss, the physical residual loss and the temperature field reconstruction loss, and the weight coefficients are set to 1, 1, 0.5 and 0.2 respectively.

[0084] S2.5: Generate temperature and humidity mapping dataset.

[0085] Specifically, the multi-band infrared radiation data of the wrapping paper surface are standardized, and the magnetic field distribution on the wrapping paper surface is obtained through finite element simulation.

[0086] The first generator, the second generator, the first adversary, and the second adversary are updated alternately, and the Adam optimizer is used. After each round of training, the sample pairs generated by the first generator and the second generator are frozen.

[0087] A random affine transformation is applied to the generated wrapping paper surface temperature, and random impulse noise is injected into the generated wrapping paper surface temperature.

[0088] The temperature and humidity mapping dataset contains multi-band infrared radiation data from the wrapping paper surface, the magnetic field distribution on the wrapping paper surface, and the generated surface temperature and humidity of the wrapping paper surface. The dataset is divided into training, validation, and test sets in a 70:15:15 ratio.

[0089] The dual-generator architecture combines a humidity diffusion physics model with a magnetorheological heat conduction model, ensuring that the generated humidity distribution and noise-resilient temperature field are more consistent with the physical properties of the wrapping paper material, enhancing the architecture's physical consistency and generalization capabilities. The first generator models the wrapping paper's microstructure and humidity diffusion pathways, introducing physical residual constraints to ensure that the generated humidity distribution conforms to Friedel's law. The second generator combines a magnetic field-controlled thermal excitation mechanism with the heat conduction equation to make the generated temperature field more noise-resilient. The feature interaction unit uses a cross-modal attention mechanism to enable interaction between temperature and humidity data during the generation process, enhancing the realism and robustness of the generated data. An adversarial training strategy comprehensively optimizes the adversarial loss, physical residual loss, and temperature reconstruction loss to achieve higher generation quality. Finite element simulation, random affine transformation, and impulse noise injection are combined to ensure the generated temperature and humidity mapping dataset has greater diversity and robustness. The training, validation, and test sets are rationally divided to ensure the rationality of the training.

[0090] S3: Use the Poincare disk model to process the flat area in the center of the wrapping paper, use the Lorentz model to process the edge mutation area, and extract the mixing feature tensor in the hyperbolic Euclidean mixing space.

[0091] Specifically, the following steps are included:

[0092] S3.1: Perform regional division and preprocessing.

[0093] Specifically, based on the temperature and humidity mapping dataset, for each band , the Sobel operator is used to calculate the radiation intensity gradient amplitude , the expression is:

[0094]

[0095] in, For band Infrared radiation data, is the radiation intensity gradient amplitude, is the band index.

[0096] Setting the gradient threshold ,in is the mean value of the gradient amplitude, is the standard deviation of the gradient amplitude.

[0097] The central flat area and the edge mutation area are divided according to the gradient threshold. The expression is:

[0098]

[0099] in, It is a central flat area. The edge mutation zone, coordinate The radiation intensity gradient amplitude at is the gradient threshold.

[0100] S3.2: Treating the central flat region via the Poincare disk model.

[0101] It should be understood that the goal of this step is to map the flat area data into a hyperbolic space to capture the hierarchical moisture diffusion pattern.

[0102] Specifically, for the central flat area Infrared radiation data Normalize and map to the Poincare disk , the expression is:

[0103]

[0104] in, is a hyperbolic space mapping point, for The mean of for The standard deviation of is the natural base, is an imaginary unit, satisfying , is the local gradient direction angle.

[0105] Guaranteed by the scaling factor .

[0106] Furthermore, we construct a geodesic line coordinate system, define the hyperbolic tensor and calculate the local curvature characteristics, which are expressed as:

[0107]

[0108] in, is the hyperbolic metric tensor, is the local curvature feature.

[0109] Extracting the curvature tensor ,in, is the curvature dimension, is the set of real numbers.

[0110] S3.3: Treating edge mutation regions using the Lorentz model.

[0111] It should be understood that the goal of this step is to decompose the gradient field in the Euclidean space and capture the edge humidity mutation characteristics.

[0112] Specifically, for the edge mutation zone Infrared radiation data , constructing the Lorentz manifold The tangent space projection on , defines the pseudo-Riemannian metric and decomposes the gradient into the normal vector and the tangent vector, the expression is:

[0113]

[0114] in, is the normal vector, is the tangent vector, is a random perturbation vector.

[0115] Furthermore, based on the normal vector and the tangent vector, the covariant derivative is calculated as follows:

[0116]

[0117] in, The tangent vector is The direction of the component, The tangent vector is Component in the direction. Extract the mutation feature tensor .

[0118] S3.4: Generate mixed feature tensor.

[0119] It should be understood that the goal of this step is to fuse hyperbolic and Euclidean features while maintaining geometric consistency.

[0120] Specifically, yes Perform a logarithmic projection and apply a hyperbolic to Euclidean projection. Perform normalization.

[0121] Splice along the spatial dimension and perform channel normalization to extract the mixed feature tensor .

[0122] Preferably, based on gradient amplitude analysis, the central flat area and the edge mutation area are divided to adapt to the geometric characteristics of different areas. For the central flat area, the Poincare disk model is used to map the infrared radiation data to the hyperbolic space, define the hyperbolic metric tensor and calculate the local curvature characteristics to capture the hierarchical humidity diffusion pattern. For the edge mutation area, the gradient field is decomposed using the Lorentz manifold, the normal vector and tangent vector are extracted, and the covariant derivative is calculated to analyze the mutation characteristics of humidity changes. By applying the hyperbolic to Euclidean projection, a unified hybrid feature tensor is formed. It effectively combines the hierarchical information expression ability of the hyperbolic space with the local change analysis ability of the Euclidean space, enhances the modeling ability of complex humidity distribution patterns, and provides accurate and efficient support for packaging paper humidity monitoring and quality control.

[0123] S4: Perform humidity distribution inference through a hyperbolic neural network and output the humidity distribution matrix of the packaging paper surface.

[0124] Specifically, the following steps are included:

[0125] S4.1: Construct a hyperbolic neural network.

[0126] Specifically, the input of the hyperbolic neural network is the mixed feature tensor, and the output is the moisture distribution matrix of the packaging paper surface.

[0127] The structure of the hyperbolic neural network includes: dynamic curvature convolution layer (3 layers), which is used to adaptively adjust the geometry of the convolution kernel; hyperbolic attention unit (2 layers), which is used for weighted fusion of cross-channel features; hyperbolic pooling layer, which is used to downsample and maintain the hyperbolic flow structure; hyperbolic regression head, which is used to map to the Euclidean space output humidity matrix.

[0128] S4.2: Adjust the dynamic curvature convolution layer.

[0129] Specifically, the input mixed feature tensor is subjected to hyperbolic tangent constraint to calculate the local curvature, and the expression is:

[0130]

[0131] in, is the local curvature, is the mixed feature tensor, is the Poincare disk logarithmic map.

[0132] The basic kernel is deformed according to the local curvature to obtain the curvature adaptive convolution kernel, which is expressed as:

[0133]

[0134] in, is the curvature adaptive convolution kernel, As the basic core, is the learnable scaling factor, , is the element-wise product.

[0135] Furthermore, based on the curvature adaptive convolution kernel, the convolution operation is used in the Poincare disk to perform hyperbolic convolution operation.

[0136] S4.3: Compute cross-channel attention weights via the hyperbolic attention mechanism.

[0137] Specifically, the geodesic distance is calculated as follows:

[0138]

[0139] in, is the geodesic distance, is the feature channel, is another feature channel, is the inverse hyperbolic cosine function.

[0140] The similarity weights between feature channels are calculated based on the geodesic distance and normalized. The expression is:

[0141]

[0142] in, is the similarity weight between feature channels, is the number of channels, is the channel number index, is the feature dimension.

[0143] Hyperbolic feature fusion is performed based on the weighted aggregation features of the similarity between channels.

[0144] S4.4: Map the hyperbolic features to Euclidean space and output the moisture distribution matrix of the wrapping paper surface.

[0145] Specifically, a geometric projection is performed on the output feature tensor. Each hyperbolic feature point is projected radially onto the tangent space at the origin using the Poincare logarithm map based on the origin. This process uses the inverse hyperbolic tangent function to calculate the scaling factor and normalize the original features to ensure that the projected features are completely in Euclidean space.

[0146] The spatial dimensions are adjusted, and the projected feature map is expanded using bilinear interpolation. After aligning the feature dimensions, a nonlinear transformation is performed on the feature vector of each pixel. Multi-channel information is compressed into a single-channel moisture estimate. A sigmoid activation function is applied to the output of each pixel, constraining the value between 0 and 1, corresponding to the dry and wet states of the wrapping paper surface.

[0147] The final output is a moisture distribution matrix, where each matrix element corresponds to the normalized moisture value at a specific location on the wrapping paper.

[0148] S4.5: Set the loss function and training strategy.

[0149] Specifically, the expression of humidity distribution loss (main loss) is:

[0150]

[0151] in, is the humidity distribution loss function, is the width of the humidity distribution matrix, is the depth of the humidity distribution matrix, For the coordinates The predicted humidity value at For the coordinates The actual humidity value.

[0152] The expression of hyperbolic manifold constraint loss (regular term) is:

[0153]

[0154] in, is the hyperbolic manifold constraint loss function, is the number of neural network layers, is the neural network layer index, is the first layer neural network hidden representation, is the curvature constant of the hyperbolic space, For the The characteristic radius of the layer.

[0155] The Riemannian Adam optimizer is used, and the learning rate is set to , the momentum is set to 0.9 and 0.999.

[0156] Preferably, the state curvature convolution layer adaptively adjusts the convolution kernel shape according to the local curvature to more accurately capture the humidity characteristics in non-Euclidean space. The hyperbolic attention mechanism calculates the similarity between channels through geodesic distance and weightedly fuses the information of different feature channels to improve feature expression capabilities. The hyperbolic features are projected into Euclidean space through logarithmic mapping, and the humidity distribution matrix is generated using a fully connected layer to ensure that the output meets the actual application requirements. During training, the humidity distribution loss function constrains the error between the predicted value and the true value, while the hyperbolic manifold constraint loss further maintains the manifold structure of the network in the hyperbolic space and improves generalization ability. Combined with the efficient training strategy of the Riemannian Adam optimizer, it is possible to improve the convergence speed and stability of the hyperbolic neural network while maintaining hyperbolic geometric consistency.

[0157] This embodiment also provides a computer device suitable for the packaging paper moisture detection method based on infrared imaging temperature distribution analysis, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the packaging paper moisture detection method based on infrared imaging temperature distribution analysis proposed in the above embodiment.

[0158] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0159] This embodiment further provides a storage medium having a computer program stored thereon. When the program is executed by a processor, the program implements the packaging paper moisture detection method based on infrared imaging temperature distribution analysis as proposed in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0160] In summary, the present invention achieves dynamic band selection by adopting a programmable metasurface filter array, accurately capturing multi-band infrared radiation data on the surface of packaging paper, and improving the accuracy of humidity detection; at the same time, it utilizes a dual generative adversarial network to simulate the humidity diffusion path and generate a noise-resistant temperature field, thereby enhancing the adaptability to different packaging paper materials and microstructures; by applying the Poincare disk model and the Lorentz model to the central flat area and the edge mutation area of the packaging paper respectively, and extracting the mixed feature tensor in the hyperbolic Euclidean mixed space, it realizes the fine modeling of complex temperature and humidity distribution; combining the efficient nonlinear feature processing of the dynamic curvature convolution layer and the hyperbolic attention mechanism in the hyperbolic neural network, it not only optimizes the computational efficiency, but also improves the anti-interference ability; the humidity distribution matrix finally output can be directly used to generate intuitive visualization graphics, providing reliable technical support for quality control and detection decision-making, and has certain industrial integration and application prospects.

[0161] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for detecting moisture content in packaging paper based on infrared imaging temperature distribution analysis, characterized in that: include, According to the type of wrapping paper, the programmable metasurface filter array is triggered to perform dynamic band selection and collect multi-band infrared radiation data from the wrapping paper surface; A dual generative adversarial network is used to generate a temperature and humidity mapping dataset, where the first generator simulates the humidity diffusion path and the second generator generates a noise-resistant temperature field through magnetorheological thermal excitation. The first generator simulates the humidity diffusion path as follows: The encoder uses 5 layers of 3D convolution to extract the geometric features of the wrapping paper material parameters; The implicit finite difference method is used to discretize the PDE, construct the residual constraint term, and add the physical residual into the loss function; The decoder uses transposed convolution to generate the humidity distribution matrix, and the output layer is Sigmoid activation; The second generator generates an anti-noise temperature field through magnetorheological thermal excitation, specifically as follows: The multi-band infrared radiation data of the wrapping paper surface is input and the multi-band features are extracted by the Inception unit. The magnetic field distribution on the wrapping paper surface is encoded into a spatial modulation vector through the fully connected layer. The thermal excitation power density is regulated based on the magnetic field distribution on the packaging paper surface, and the heat conduction equation constraint is embedded in the temperature field generation; Using the U-Net structure, the skip connection injects the magnetic field modulation information, and the output layer adopts Tanh activation to map it to the temperature range; The Poincare disk model is used to process the flat area in the center of the wrapping paper, and the Lorentz model is used to process the edge mutation area. The mixed feature tensor is extracted in the hyperbolic Euclidean mixing space as follows: Based on the temperature and humidity mapping dataset, the Sobel operator is used to calculate the radiation intensity gradient amplitude for each band, and the gradient threshold is set to divide the central flat area and the edge mutation area; The infrared radiation data of the central flat area is normalized and mapped to the Poincare disk, and a geodetic line coordinate system is constructed. The hyperbolic tensor is defined to calculate the local curvature characteristics and extract the curvature tensor. For the infrared radiation data of the edge mutation area, a tangent space projection on the Lorentz manifold is constructed, a pseudo-Riemannian metric is defined, and the gradient is decomposed into a normal vector and a tangent vector. The covariant derivative is calculated, and the mutation feature tensor is extracted. Perform logarithmic mapping on the curvature tensor, apply hyperbolic to Euclidean projection, normalize the mutation feature tensor, splice along the spatial dimension, and perform channel normalization to extract the mixed feature tensor; Moisture distribution inference is performed through a hyperbolic neural network, and a moisture distribution matrix of the packaging paper surface is output. The hyperbolic neural network includes a dynamic curvature convolution layer and a hyperbolic attention mechanism.

2. The method for detecting packaging paper humidity based on infrared imaging temperature distribution analysis according to claim 1, characterized in that: The first generator combines the material properties of the wrapping paper and the constraints of the humidity diffusion partial differential equation to generate a humidity diffusion path that matches the microstructure of the wrapping paper.

3. The method for detecting packaging paper humidity based on infrared imaging temperature distribution analysis according to claim 1, wherein: The second generator generates a non-uniform thermal excitation field by regulating the magnetic field distribution, and suppresses the temperature field noise in combination with the heat conduction equation.

4. The method for detecting packaging paper humidity based on infrared imaging temperature distribution analysis according to claim 1, wherein: The dynamic band selection adjusts the transmission band combination of the filter array in real time by analyzing the spectral absorption characteristics of the wrapping paper and compensates the band weights based on ambient temperature fluctuations.

5. The method for detecting packaging paper humidity based on infrared imaging temperature distribution analysis according to claim 1, wherein: The specific steps of extracting the mixed feature tensor in the hyperbolic Euclidean mixed space are as follows: The infrared radiation data of the flat area in the center of the wrapping paper is mapped to the geodesic coordinate system of the Poincare disk model to extract the curvature characteristics in the hyperbolic space. The infrared radiation data of the edge mutation area is input into the Lorentz model for gradient field decomposition to extract the mutation characteristics in Euclidean space; The hyperbolic and Euclidean features are fused through tensor concatenation and normalization operations to generate a hybrid feature tensor.

6. The method for detecting packaging paper humidity based on infrared imaging temperature distribution analysis according to claim 1, wherein: The dynamic curvature convolution layer adaptively adjusts the geometric shape of the convolution kernel according to the curvature change of the local feature manifold, and dynamically constrains the curvature parameters using the hyperbolic tangent function.

7. The method for detecting packaging paper humidity based on infrared imaging temperature distribution analysis according to claim 1, wherein: The hyperbolic attention mechanism generates attention weights by calculating the geodesic distance of the mixed feature tensor in the Poincare disk, and performs nonlinear weighted fusion of feature channels based on hyperbolic space projection.

8. The method for detecting packaging paper humidity based on infrared imaging temperature distribution analysis according to claim 1, wherein: The packaging paper surface humidity distribution matrix is used to generate a visual humidity distribution diagram and perform humidity detection.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the packaging paper moisture detection method based on infrared imaging temperature distribution analysis according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the packaging paper moisture detection method based on infrared imaging temperature distribution analysis according to any one of claims 1 to 8 are implemented.

Images