Carbon footprint detection method and system based on polarized imaging

Abstract

The application provides a kind of carbon footprint detection method and system based on polarization imaging, it is related to carbon footprint detection technical field.The method comprises the following steps: collecting polarization infrared and spectral image, extracting intensity, polarization degree, polarization angle and spectral feature;Intensity, polarization degree, polarization angle feature is fused and input into target detection network, and the region to be identified material is segmented;Extract the polarization and spectral feature vector of the region, and get the posterior probability after classification;Calculate the polarization similarity, spectral similarity, region purity and density size consistency score, and obtain the carbon footprint factor confidence level;Fuse the foregoing index to obtain the fusion confidence;Construct carbon footprint deviation cost sensitive loss function, calculate the expected loss based on the fusion confidence, and select the minimum loss category as the final recognition result;Statistical material pixel ratio, combined with density size conversion mass ratio, combined with total mass and carbon footprint factor weighting to obtain total carbon footprint.The problems of manual reporting, single-dimensional recognition being easily disturbed and being difficult to automate are solved.

Description

Technical Field

[0001] This invention relates to the field of carbon footprint detection technology, and more specifically, to a carbon footprint detection method and system based on polarization imaging. Background Technology

[0002] Carbon footprint is a key quantitative indicator used throughout the entire product lifecycle for carbon emission accounting and verification. Accurate and efficient carbon footprint accounting helps various market players clearly understand their emissions, identify high-emission stages, and implement targeted carbon reduction technology upgrades. Currently, the mainstream carbon footprint accounting method relies primarily on manual reporting, depending on companies providing their own production process data and emission coefficients. This method is difficult to implement, time-consuming, and prone to errors and omissions, making it difficult to guarantee the objectivity and accuracy of the accounting results. Furthermore, it cannot achieve dynamic real-time monitoring of carbon emissions during the production process. As carbon footprint management develops towards large-scale and dynamic approaches, there is an urgent market demand for rapid, accurate, and automated carbon footprint accounting and monitoring solutions. Traditional manual accounting methods are no longer sufficient to meet practical application requirements.

[0003] To improve accounting efficiency, it is necessary to quickly identify the types and proportions of materials used in products and automatically match carbon footprint factors for calculation. Existing material identification methods typically rely on single spectral information or visual features for analysis. This one-dimensional detection method is easily affected by changes in external lighting, differences in material surface coatings, and complex environmental backgrounds, resulting in unstable identification accuracy and insufficient system robustness. Therefore, there is an urgent need for a detection method that can overcome the above-mentioned shortcomings and achieve high-precision, high-robustness material identification and rapid carbon footprint calculation. Summary of the Invention

[0004] The purpose of this invention is to provide a carbon footprint detection method and system based on polarization imaging, which aims to solve the problems in existing carbon footprint accounting, such as the susceptibility of material identification to light and environmental interference, insufficient accuracy and robustness, and the inability to achieve automated dynamic real-time detection.

[0005] This invention is achieved through the following technical solution: A carbon footprint detection method based on polarization imaging includes the following steps: The polarized infrared image and spectral image of the object under test are acquired and preprocessed to obtain multidimensional feature data; wherein, the multidimensional feature data includes intensity features, polarization degree features, polarization angle features and spectral features; The intensity feature, polarization degree feature and polarization angle feature in the multidimensional feature data are fused in multiple channels and input into the target detection network to locate and segment at least one material region to be identified. The polarization feature vector and spectral feature vector of the material region to be identified are extracted and input into the material classification model to obtain the posterior probability of the material region to be identified corresponding to multiple candidate material categories; Calculate the polarization similarity between the polarization features of each material region to be identified and the standard polarization features of the candidate materials, the spectral similarity between the spectral features of each material region to be identified and the standard spectral features of the candidate materials, the region pixel purity, and the consistency score between the preset density of each candidate material category and the physical size of the material region to be identified; obtain the data credibility level corresponding to each candidate material category in the material carbon footprint factor library; Based on the posterior probability, polarization similarity, spectral similarity, regional purity, consistency score, and data confidence level, the fusion confidence of each candidate material category is calculated. A cost-sensitive loss function with carbon footprint factor bias as the cost of misclassification is constructed. Based on the fusion confidence of each candidate material category, the expected loss of selecting each candidate material category is calculated. The candidate material category with the smallest expected loss is selected as the final identification result of the material region to be identified. Based on the final recognition results of all material regions to be identified, the pixel percentage of each material category is calculated, and combined with the preset density of each candidate material category and the physical size of the material region to be identified, it is converted into a mass percentage. The total mass of the object under test is obtained. The total carbon footprint is obtained by weighted summing of the mass percentage, the total mass, and the unit mass carbon footprint factor matched from the material carbon footprint factor library.

[0006] Optionally, the specific process of fusing the intensity features, polarization degree features, and polarization angle features from the multi-dimensional feature data into multiple channels, inputting them into the target detection network, and locating and segmenting at least one material region to be identified is as follows: The intensity feature, polarization degree feature, and polarization angle feature in the multidimensional feature data are normalized to the same spatial resolution and then stitched together in the channel dimension to form a multi-channel fused feature map; wherein, the multi-channel fused feature map includes: intensity feature map, polarization degree feature map, and polarization angle feature map; The multi-channel fused feature map is input into the target detection network, which uses YOLOv8 as the baseline architecture. A polarization feature extraction branch is added to the backbone network of the target detection network. The polarization feature extraction branch performs convolution and downsampling on the polarization degree feature map and polarization angle feature map in the multi-channel fused feature map to extract polarization texture features that reflect the anisotropy and microstructure of the material surface. These features are then fused layer by layer with the intensity feature map of the corresponding scale in the backbone network to obtain enhanced multimodal features. The target detection network is based on multimodal features. It regresses the bounding box of the object under test through feature pyramid and detection head, performs pixel-level semantic segmentation on the region within the bounding box, generates a foreground mask, and removes background regions, shadow regions and coating coverage regions to obtain at least one material region to be identified.

[0007] Optionally, the specific process for calculating the regional pixel purity of each region of material to be identified and the standard spectral characteristics of candidate materials is as follows: Extract the spectral feature vectors of all pixels within the region of the material to be identified, and construct a regional spectral feature set; For any candidate material category, its standard spectral feature vector is obtained, and the spectral angle between the spectral feature vector of each pixel in the region spectral feature set and the standard spectral feature vector is calculated to generate a pixel-level spectral angle sequence. The arithmetic mean of the pixel-level spectral angle sequence is used as the spectral similarity between the candidate material category and the region of the material to be identified; The number of pixels with spectral angles less than a preset consistency threshold in the pixel-level spectral angle sequence is counted, and the proportion of this number of pixels to the total number of pixels in the material region to be identified is calculated. This proportion is used as the region pixel purity corresponding to the candidate material category.

[0008] Optionally, the specific process for calculating the consistency score between the preset density of each candidate material category and the physical size of the material region to be identified is as follows: The calibration pixel size, focal length, and synchronously acquired object distance of the polarized infrared imaging device are obtained. Based on the object distance and imaging geometry, the actual physical side length corresponding to a single pixel is determined, and the pixel area of ​​the material region to be identified is converted into the actual area. Obtain the thickness of the material region to be identified; wherein, the thickness is the measured thickness synchronously obtained by the measuring device, or the standard thickness corresponding to the object category to which the material region to be identified belongs, retrieved from a preset material thickness library; Multiply the actual area by the thickness to obtain the measured volume of the material region to be identified; Extract the preset density of candidate material categories from the material carbon footprint factor library, calculate the estimated mass when assuming the material region to be identified is composed of candidate material categories, and obtain the measured total mass of the object being measured. When the object under test contains multiple regions of material to be identified, the estimated quality of each region other than the current region under the current best candidate material is fixed, and the total estimated quality is obtained by summing them up when the current region is assumed to be the candidate material category, and the quality deviation is calculated; when the object under test contains only a single region of material to be identified, the estimated quality of that region is directly used to calculate the quality deviation with the measured total quality. The quality deviation is mapped to a consistency score using a preset monotonically decreasing scoring function. A smaller deviation results in a higher score, characterizing the physical consistency between the preset density of the candidate material category and the physical dimensions of the material region to be identified. The consistency score ranges from [value missing]. .

[0009] Optionally, the specific process for calculating the fusion confidence of each candidate material category is as follows: First to sixth weight coefficients are assigned to the posterior probability, the polarization similarity, the spectral similarity, the regional purity, the consistency score, and the data credibility level, respectively, and the sum of the first to sixth weight coefficients is 1. For any candidate material category, the posterior probability, polarization similarity, spectral similarity, regional purity, consistency score, and data credibility level of the candidate material category are multiplied by their respective weight coefficients, and the results of each product are added together to obtain the comprehensive score corresponding to the candidate material category. Based on the comprehensive scores of all candidate material categories, a normalized exponential function is applied to all comprehensive scores to obtain the normalized value of each candidate material category. The normalized value is then used as the fusion confidence of the candidate material category.

[0010] Optionally, the specific process of constructing the cost-sensitive loss function with carbon footprint factor bias as the cost of misclassification is as follows: Extract the unit mass carbon footprint factor corresponding to each candidate material category and the data reliability level of the factor library from the material carbon footprint factor library; For decision categories Compared to the real category Construct a cost-sensitive loss function of the following form. :

[0011] in, For indicator functions, in The value is 1 if the condition is met, otherwise it is 0. and Decision categories and real category Carbon footprint factor per unit mass; To prevent extremely small constants with a denominator of zero; For decision categories The data reliability level in the materials carbon footprint factor library, and the value range is [missing information]. ; , and These are the non-negative weight coefficients for the corresponding terms.

[0012] Optionally, after selecting the candidate material category with the minimum expected loss as the final identification result for the material region to be identified, the method further includes a step of hierarchical review of the identification result: Obtain the detection confidence of the material region to be identified in the target detection network, as well as the fusion confidence and expected loss value corresponding to the final identification result; If the detection confidence level is less than the preset detection threshold, the current image quality is determined to be unqualified, and a prompt message is generated to re-acquire or adjust the imaging angle. If the detection confidence is greater than or equal to the detection threshold, but the fusion confidence is less than the preset first confidence threshold, then the current recognition result is determined to be in a fuzzy state, and the corresponding material area to be recognized is marked as an area to be manually reviewed. If the fusion confidence level is greater than or equal to the first confidence level threshold, but the expected loss value is greater than the preset risk threshold, then it is determined that although the current identification result meets the confidence level, the carbon footprint accounting deviation risk is too high, and the corresponding material area to be identified is marked as an area to be reviewed. If the fusion confidence level is greater than or equal to the first confidence threshold, and the expected loss value is less than or equal to the risk threshold, then the current identification result is determined to be valid, and the final identification result is output to subsequent steps.

[0013] Based on the same inventive concept, the present invention also provides a carbon footprint detection system based on polarization imaging, for implementing the aforementioned carbon footprint detection method based on polarization imaging, comprising: The image acquisition module is used to simultaneously acquire polarized infrared and spectral images of the object under test; A data preprocessing module, connected to the image acquisition module, is used to preprocess the polarized infrared image and the spectral image to generate multidimensional feature data containing intensity features, polarization degree features, polarization angle features and spectral features. The material region segmentation module is connected to the data preprocessing module. It is used to receive the intensity features, polarization degree features and polarization angle features in the multi-dimensional feature data and perform multi-channel fusion. The fused feature map is input into the target detection network to locate and segment at least one material region to be identified. The feature extraction and classification module is connected to the material region segmentation module and the data preprocessing module, respectively. It is used to extract polarization feature vectors and spectral feature vectors from the material region to be identified, and input the polarization feature vectors and spectral feature vectors into the material classification model to output the posterior probability of each candidate material category. The storage module is used to store the standard polarization characteristics, standard spectral characteristics, preset density, carbon footprint factor, and data reliability level corresponding to each candidate material category of the candidate material. A multi-dimensional scoring calculation module is connected to the material region segmentation module and the storage module respectively. It is used to calculate the polarization similarity, spectral similarity, region pixel purity, and consistency score between the preset density and the physical size of the material region to be identified for each candidate material category, and to read the data credibility level. The fusion confidence calculation module is connected to the feature extraction and classification module, the multi-dimensional scoring calculation module, and the storage module, respectively, and is used to calculate the fusion confidence of each candidate material category based on the posterior probability, the polarization similarity, the spectral similarity, the regional purity, the consistency score, and the data credibility level. The cost-sensitive decision module is connected to the fusion confidence calculation module and the storage module respectively. It is used to construct a cost-sensitive loss function with carbon footprint factor bias as the cost of misclassification, and to select the expected loss of each candidate material category based on the fusion confidence of each candidate material category. The candidate material category with the smallest expected loss is selected as the final identification result of the material region to be identified. The carbon footprint calculation module is connected to the cost-sensitive decision module and the storage module respectively. It is used to calculate the pixel ratio of each material category based on the final identification results of all material regions to be identified, convert the preset density and the physical size of the material region to be identified into the mass ratio, and then combine the total mass of the measured object with the unit mass carbon footprint factor matched from the storage module to output the total carbon footprint.

[0014] Based on the same inventive concept, the present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the above-described carbon footprint detection method based on polarization imaging.

[0015] Based on the same inventive concept, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described carbon footprint detection method based on polarization imaging.

[0016] The technical solution of the present invention has at least the following advantages and beneficial effects: By acquiring polarized infrared and spectral images, multi-dimensional features such as intensity, degree of polarization, polarization angle, and spectrum are simultaneously obtained and fused for analysis. This overcomes the shortcomings of traditional single-dimensional recognition methods, which are susceptible to changes in illumination, differences in surface coatings, and interference from complex backgrounds, thereby improving the accuracy of material recognition and the robustness of the system.

[0017] By using a target detection network to automatically locate and segment the material region to be identified, and combining the material classification model with multi-source information to calculate the fusion confidence score, the entire process from image acquisition to material identification is automated. This effectively avoids the data misreporting and omission problems that exist in manual reporting, improves the efficiency and objectivity of carbon footprint accounting, and can meet the needs of dynamic real-time detection.

[0018] By comprehensively considering posterior probability, polarization similarity, spectral similarity, regional pixel purity, physical size consistency score, and data credibility level, a multi-dimensional fusion confidence score is constructed. A cost-sensitive loss function with carbon footprint factor bias as the cost of misclassification is introduced, and the material category with the smallest expected loss is selected as the final result. This makes the material identification process directly aimed at minimizing the carbon footprint accounting error, further ensuring the accuracy of the carbon footprint accounting results.

[0019] By combining the material identification results with pixel ratio, preset density and physical size to convert into mass ratio, and then weighting and summing the total mass and carbon footprint factor per unit mass to obtain the total carbon footprint, a complete, fast and accurate automated detection solution for product carbon footprint is provided. Attached Figure Description

[0020] Figure 1 This is a schematic flowchart of the carbon footprint detection method based on polarization imaging according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a carbon footprint detection system based on polarization imaging according to an embodiment of the present invention. Detailed Implementation

[0021] The following is a detailed description of the embodiments, in conjunction with the accompanying drawings.

[0022] Reference Figure 1 A carbon footprint detection method based on polarization imaging includes the following steps: Step 1: Acquire polarized infrared and spectral images of the object under test, and obtain multidimensional feature data through preprocessing; wherein, the multidimensional feature data includes intensity features, polarization degree features, polarization angle features and spectral features.

[0023] In some embodiments, a polarized infrared imaging module and a short-wave infrared multispectral polarization camera can be used to simultaneously acquire multi-dimensional optical information of the target object. An intensity image containing polarization components in four directions (0°, 45°, 90°, and 135°) is obtained through a polarization focal plane detector, and the degree of polarization (DOLP) and angle of polarization (AOP) features are calculated. At the same time, a spectral imaging module captures spectral data covering absorption bands of key functional groups such as C–H, O–H, and C=O. After preprocessing operations such as nonlocal mean filtering and median filtering for noise reduction, polarization correction, radiometric calibration, and geometric correction, infrared noise, stripe interference, and environmental errors are eliminated. Finally, a four-dimensional feature data cube with spatial dimension alignment and deep fusion of intensity, polarization, and spectral information is constructed, thereby acquiring and outputting multi-dimensional feature data covering intensity, degree of polarization, angle of polarization, and spectral absorption characteristics in one go.

[0024] Furthermore, the polarization infrared imaging module can employ an uncooled long-wavelength infrared polarization focal plane detector as its core sensor, operating in the 8μm–12μm band to capture molecular vibrational characteristics carried by the material's own thermal radiation, unaffected by ambient visible light interference. A subwavelength linear grating micro-polarizer array is monolithically integrated on the detector's focal plane. This array consists of micro-polarizer units with four polarization directions (0°, 45°, 90°, and 135°) arranged in a 2×2 superpixel period, with each micro-polarizer having a period of 200nm and a height of 300nm. After the incident infrared radiation is filtered by the micro-polarizer array, the polarization components in the four directions within each superpixel are synchronously received by the corresponding pixel, thus enabling the acquisition of a four-directional polarization intensity image in a single exposure. The optical system consists of a zinc selenide infrared window, a liquid crystal polarization grating, and an achromatic infrared lens, arranged sequentially. The zinc selenide infrared window has a transmittance of over 95% in the 8μm–12μm wavelength range, protecting internal components. The liquid crystal polarization grating has a diffraction efficiency of over 90% at a center wavelength of 10μm, further suppressing stray polarization crosstalk. The achromatic infrared lens images the target object onto the detector's focal plane, achieving large field-of-view, high-resolution polarization imaging. In addition, the module integrates an active infrared illumination unit (operating wavelength 8μm–12μm), a high-precision gimbal (angle control accuracy ±0.1°), and environmental temperature and humidity sensors and object distance sensors to compensate for the impact of environmental factors on polarization measurement accuracy.

[0025] After the device is powered on, polarization response calibration and dark current correction are performed first. By acquiring images of a standard diffuse reflection reference plate and an opaque baffle with known polarization characteristics, the responsivity curves and dark current reference values ​​of each pixel under different polarization directions are established and stored in the system's storage module for subsequent calibration.

[0026] After calibration, the high-precision gimbal is controlled to drive the polarization infrared imaging module to perform polarization imaging of the target object at four angles: 0°, 45°, 90°, and 135°. Alternatively, an electrically controlled polarization state switching can be achieved using a liquid crystal polarization grating without mechanical rotation. At each imaging angle, the detector simultaneously outputs intensity images I0, I2, and I3 in four polarization directions. 45 I 90 I 135 Based on the four intensity images above, the Stokes parameters are calculated pixel-by-pixel according to the Stokes vector calculation formula. (Total light intensity / Total radiation intensity) (Difference between horizontal and vertical linear polarization components) (The difference between the 45° and 135° linear polarization components), and further derive the polarization degree DOLP and polarization angle AOP images.

[0027] The following preprocessing steps are performed on each of the solved images in sequence: Non-local mean filtering combined with median filtering is used to suppress fixed-mode noise and random stripe interference from the infrared detector; radiometric calibration is performed based on the radiometric calibration coefficients in the system parameter library to convert the original digital signal into a radiometric intensity value; atmospheric transmission attenuation compensation and geometric correction are performed on the degree of polarization based on the measurements from the ambient temperature and humidity sensor and the object distance sensor to eliminate environmental and equipment errors.

[0028] Finally, the intensity image I, polarization degree image DOLP, and polarization angle image AOP are aligned in space to form a data cube containing multi-dimensional features of intensity and polarization, which is then output to the subsequent target detection and material recognition stages.

[0029] Step 2: Perform multi-channel fusion of intensity features, polarization degree features, and polarization angle features from the multi-dimensional feature data, input the fusion into the target detection network, and locate and segment at least one material region to be identified.

[0030] In some embodiments, the specific process of fusing intensity features, polarization degree features, and polarization angle features from multi-dimensional feature data into multiple channels, inputting the data into a target detection network, and locating and segmenting at least one material region to be identified is as follows: The intensity feature, polarization degree feature, and polarization angle feature in the multidimensional feature data are normalized to the same spatial resolution and then stitched together in the channel dimension to form a multi-channel fused feature map; wherein, the multi-channel fused feature map includes: intensity feature map, polarization degree feature map, and polarization angle feature map; The multi-channel fused feature map is input into the target detection network, which uses YOLOv8 as the baseline architecture. A polarization feature extraction branch is added to the backbone network of the target detection network. The polarization feature extraction branch performs convolution and downsampling on the polarization degree feature map and polarization angle feature map in the multi-channel fused feature map to extract polarization texture features that reflect the anisotropy and microstructure of the material surface. These features are then fused layer by layer with the intensity feature map of the corresponding scale in the backbone network to obtain enhanced multimodal features. The target detection network is based on multimodal features. It regresses the bounding box of the object under test through feature pyramid and detection head, performs pixel-level semantic segmentation on the region within the bounding box, generates a foreground mask, and removes background regions, shadow regions and coating coverage regions to obtain at least one material region to be identified.

[0031] It should be noted that the input to the target detection network only uses a three-channel fused feature map composed of intensity features, polarization degree features, and polarization angle features, and does not directly incorporate spectral features. This is because the number of spectral channels output by the spectral imaging module is typically tens to hundreds of dimensions. Directly stacking these as image channels and inputting them into the target detection network would significantly increase the number of network parameters and computational overhead, which is detrimental to real-time detection. Furthermore, the spatial resolution of spectral imaging is usually lower than that of polarized infrared imaging, and direct fusion may lead to a decrease in spatial alignment accuracy. Therefore, this invention places the utilization of spectral features in the material classification stage of step three, performing deep fusion at the feature vector level through a parallel architecture of one-dimensional convolution and Transformer, balancing detection efficiency and recognition accuracy.

[0032] Step 3: Extract the polarization feature vector and spectral feature vector of the material region to be identified, and input them into the material classification model to obtain the posterior probability of the material region to be identified corresponding to multiple candidate material categories.

[0033] In some embodiments, for each material region to be identified located and segmented in step two, the polarization feature vector and spectral feature vector are extracted from its multidimensional feature data. The polarization feature vector contains the polarization degree, polarization angle and polarization state spatial distribution statistical features obtained by Stokes parameters. The spectral feature vector covers the absorption peak position and intensity information of material characteristic functional groups such as C-H, O-H and C=O in the short-wave infrared band. The two are strictly aligned in the spatial dimension and together constitute the joint feature vector of the region. The joint feature vector is input into a material classification model constructed in parallel by a one-dimensional convolutional neural network and a Transformer encoder. The one-dimensional convolutional network branch extracts distinguishable features such as local absorption peak morphology and polarization response details by sliding multi-scale convolutional kernels along the spectral and polarization dimensions. The Transformer branch uses a multi-head self-attention mechanism to model the long-range dependencies between the dimensions of the feature vector and the global context, so as to effectively distinguish different material categories with similar spectral features. The deep features output from the two branches are weighted and fused, and then mapped to the posterior probability distribution of each candidate material category through a fully connected layer and a softmax normalization layer. This posterior probability distribution comprehensively reflects the probability that a material belongs to each category in the multidimensional polarization-spectral feature space, providing a basic probability input for the subsequent introduction of physical consistency constraints and cost-sensitive decision-making.

[0034] Step 4: Calculate the polarization similarity between the polarization characteristics of each material region to be identified and the standard polarization characteristics of the candidate materials, the spectral similarity between the spectral characteristics of each material region to be identified and the standard spectral characteristics of the candidate materials, the region pixel purity, and the consistency score between the preset density of each candidate material category and the physical size of the material region to be identified; obtain the data credibility level corresponding to each candidate material category in the material carbon footprint factor library.

[0035] In some embodiments, the area of ​​material to be identified may include For each region of material to be identified, the polarization degree image, polarization angle image, and normalized linear polarization component calculated from the Stokes parameters are extracted based on the corresponding spatial location for each pixel. Simultaneously, to capture the spatial texture information of the polarization state, the polarization degree image and polarization angle image are calculated separately, centered on the current pixel. The local standard deviation and orientation entropy within the window. These features are concatenated along the channel dimension to form the polarization feature vector for each pixel. , .

[0036] For candidate material categories in the carbon footprint material library Using the same polarization infrared imaging and feature construction method as in step one, the regional polarization feature vector set of the standard sample of this type of material is obtained under standard imaging conditions. After clustering and outlier removal, the cluster center is taken as the standard polarization feature vector. This vector is stored in the storage module and used for similarity calculations. For pixels within the region to be identified... Its polarization eigenvector With candidate materials Standard polarization eigenvector Polarization cosine similarity between The definition is as follows:

[0037] in, Represents the dot product of vectors; It is the Euclidean norm; To prevent constants with a denominator of zero (such as...) The closer the cosine value is to 1, the more consistent the polarization characteristics of the pixel are with the standard characteristics of the candidate material.

[0038] The arithmetic mean of the polarization cosine similarity of all pixels within the region to be identified is used as the original polarization similarity between the candidate material category and the entire region to be identified. As shown in the following formula:

[0039] in, The range of values ​​is In practical applications, it can be linearly mapped to As shown in the following formula:

[0040] in, This is the normalized regional polarization similarity, ensuring that it is aligned with indicators such as posterior probability and spectral similarity during fusion. The larger this value, the better the overall polarization characteristics of the region match the candidate material.

[0041] In some embodiments, the specific process for calculating the region pixel purity of each region of material to be identified and the standard spectral characteristics of candidate materials is as follows: Extract the spectral feature vectors of all pixels within the region of the material to be identified to construct a regional spectral feature set; for any candidate material category, obtain its standard spectral feature vector, and calculate the spectral angle between the spectral feature vector of each pixel in the regional spectral feature set and the standard spectral feature vector to generate a pixel-level spectral angle sequence; use the arithmetic mean of the pixel-level spectral angle sequence as the spectral similarity between the candidate material category and the region of the material to be identified; count the number of pixels in the pixel-level spectral angle sequence whose spectral angle is less than a preset consistency threshold, and calculate the proportion of this number of pixels to the total number of pixels in the region of the material to be identified, using this proportion as the region pixel purity corresponding to the candidate material category.

[0042] Furthermore, each pixel can be set The spectral eigenvector is ( (where is the number of spectral channels), then the regional spectral feature set is: For candidate material categories Its standard spectral eigenvector is denoted as Pixels The spectral angle (SA) between the spectral eigenvector and the standard spectral vector is defined as follows:

[0043] in, For the first region of the material to be identified Individual pixels and candidate material categories The spectral angle between the two reflects the degree of deviation of the spectral shape of a single pixel from the standard spectral shape of the candidate material. The smaller the angle, the more similar the spectral features of the two. Represents the dot product operation of vectors; and Pixel spectral vectors and standard spectral vector The Euclidean norm; Generate pixel-level spectral angle sequences Candidate material categories The spectral similarity with the region of material to be identified is taken as the arithmetic mean of the pixel-level spectral angle sequence, as shown in the following formula:

[0044] in, Candidate material categories The spectral similarity between the material and the region to be identified; the smaller the value, the higher the spectral similarity.

[0045] Furthermore, a preset consistency threshold is set. (For example The number of pixels with a spectral angle less than the consistency threshold (rad) is calculated as follows:

[0046] in, For candidate material categories All materials within the area to be identified that meet the following conditions The sum of the number of pixels in the condition, that is, the number of pixels in the area that are determined to be related to the material. The number of "pure pixels" with a consistent spectrum; This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise. Then, the candidate material categories... The corresponding region pixel purity is defined as follows:

[0047] in, Candidate material categories The corresponding region pixel purity, with a value range of . The closer the pixel purity value of a region is to 1, the more homogeneous and uniform the material of that region is; conversely, if the value is low, it indicates that the region contains impurities, coatings, or is located at the boundary between two materials.

[0048] Traditional material identification methods often rely solely on regional average spectra or overall matching degrees (such as the average spectral angle of the entire region) as the sole criterion for spectral similarity, neglecting the discreteness of the spectral responses of different pixels within a region. When the segmented region contains a small amount of heterogeneous material, coatings, or boundary effects, relying solely on average similarity may lead to misjudgment. This invention introduces regional pixel purity, statistically quantifying the uniformity of material regions at the pixel level based on the proportion of spectral consistency. This metric can effectively suppress interference from "pseudo-pure" regions in subsequent multi-feature fusion decisions (such as comprehensive material judgment scoring), improving robustness to material identification in complex scenarios.

[0049] In some embodiments, the specific process for calculating the consistency score between the preset density of each candidate material category and the physical size of the material region to be identified is as follows: The calibration pixel size, focal length, and synchronously acquired object distance of the polarized infrared imaging device are obtained. Based on the object distance and imaging geometry, the actual physical side length corresponding to a single pixel is determined, and the pixel area of ​​the material region to be identified is converted into an actual area. The thickness of the material region to be identified is obtained; wherein the thickness is the measured thickness synchronously acquired by the measuring device, or a standard thickness corresponding to the object category to which the material region to be identified belongs, retrieved from a preset material thickness library. The actual area is multiplied by the thickness to obtain the measured volume of the material region to be identified. A preset density of candidate material categories is extracted from a material carbon footprint factor library, and an estimated density is calculated assuming the material region to be identified is composed of candidate material categories. The quality is measured, and the total measured mass of the object under test is obtained. When the object under test contains multiple regions of material to be identified, the estimated mass of each region other than the current region under the current optimal candidate material is fixed, and the total estimated mass is obtained by summing them up when the current region is assumed to be the candidate material category, and the quality deviation is calculated. When the object under test contains only a single region of material to be identified, the estimated mass of that region is directly used to calculate the quality deviation with the total measured mass. The quality deviation is mapped to a consistency score through a preset monotonically decreasing scoring function. The smaller the deviation, the higher the score, which is used to characterize the physical consistency between the preset density of the candidate material category and the physical size of the material region to be identified. The consistency score ranges from [value missing]. .

[0050] Furthermore, the measured volume of the material region to be identified, obtained through polarized infrared imaging and geometric calibration, can be set as follows: Calculated using the following formula:

[0051] in, The actual area of ​​the region is calculated from pixel area, pixel size, object distance, and focal length; The thickness of the area is obtained through thickness measurement or by retrieving from a preset library.

[0052] For candidate material categories Extracting its preset density from the material carbon footprint factor library Then the area is made of material Estimated quality of the composition As shown in the following formula:

[0053] If the object being measured contains A region to be identified, excluding the current region. The predicted quality of the remaining regions under their respective current best candidate materials. ( ), then assume region For materials Total estimated mass of objects at time As shown in the following formula:

[0054] Obtain the measured total mass of the object being measured. Define relative quality deviation As shown in the following formula:

[0055] When an object contains only a single region The above deviation formula also applies.

[0056] A smooth, monotonically decreasing function based on adaptive measurement uncertainty tolerance can be used to account for the relative quality deviation. Mapping to a range of values Physical consistency score The smaller the deviation, the higher the score, as shown in the following formula:

[0057] in, The tolerance coefficient is determined jointly by the system mass measurement uncertainty, volume measurement error, density variation coefficient, etc., through calibration experiments. This indicates a high consistency score when the relative deviation is within 5%. This function... When the maximum value is obtained , indicating complete agreement; with As the score increases, it decreases smoothly and approaches a certain level. The function is differentiable everywhere, which facilitates gradient optimization for subsequent multi-feature fusion.

[0058] Furthermore, when the object under test contains multiple regions of material to be identified, an iterative convergence strategy is used to calculate the consistency score of each region: In the initialization phase, for each region other than the current region, the candidate material category corresponding to the maximum posterior probability output by its material classification model is used as the initial candidate material for that region, and the initial estimated quality for each region is calculated. In the iteration phase, in the t-th iteration, for the current region, the estimated quality of the other regions under the optimal candidate material determined in the (t-1)-th iteration is fixed, and the total estimated quality is obtained by summing them up when the current region is assumed to be the candidate material category, and the quality deviation is calculated. After performing the above process in parallel for all regions, it is checked whether the temporary identification result of each region is consistent with the result of the previous iteration. If all regions converge, the final consistency score and identification result are output. If they do not converge, the candidate materials of each region are updated with the result of this iteration and the next iteration is entered. When the number of iterations reaches the preset maximum number of iterations and still does not converge, the result with the minimum expected loss of the last iteration is used as the final output.

[0059] Step 5: Based on the posterior probability, polarization similarity, spectral similarity, regional purity, consistency score, and data confidence level, calculate the fusion confidence of each candidate material category by performing a mapping on the spectral similarity (adjusting the index direction) and then weighted summing.

[0060] In some embodiments, the specific process for calculating the fusion confidence of each candidate material category is as follows: First to sixth weight coefficients are assigned to the posterior probability, the polarization similarity, the spectral similarity, the regional purity, the consistency score, and the data credibility level, respectively, and the sum of the first to sixth weight coefficients is 1. For any candidate material category, the posterior probability, polarization similarity, spectral similarity, regional purity, consistency score, and data credibility level of the candidate material category are multiplied by their respective weight coefficients, and the results of each product are added together to obtain the comprehensive score corresponding to the candidate material category. Based on the comprehensive scores of all candidate material categories, a normalized exponential function is applied to all comprehensive scores to obtain a normalized value for each candidate material category. This normalized value is then used as the fusion confidence score for that candidate material category. The normalized exponential function can be the softmax normalization function.

[0061] Step 6: Construct a cost-sensitive loss function with carbon footprint factor bias as the cost of misclassification, and calculate the expected loss for each candidate material category based on the fusion confidence of each candidate material category. Select the candidate material category with the smallest expected loss as the final identification result of the material region to be identified.

[0062] In some embodiments, the specific process of constructing a cost-sensitive loss function with carbon footprint factor bias as the cost of misclassification is as follows: Extract the unit mass carbon footprint factor corresponding to each candidate material category and the data reliability level of the factor library from the material carbon footprint factor library; For decision categories Compared to the real category Construct a cost-sensitive loss function of the following form. :

[0063] in, For indicator functions, in The value is 1 if the condition is met, otherwise it is 0. and Decision categories and real category Carbon footprint factor per unit mass; To prevent extremely small constants with a denominator of zero; For decision categories The data reliability level in the materials carbon footprint factor library, and the value range is [missing information]. ; , and These are the non-negative weight coefficients for the corresponding terms.

[0064] It should be noted that the above cost-sensitive loss function The calculations for each item do not depend on the true category label, but only on the known parameters and posterior probability distribution in the carbon footprint factor library. Therefore, they can be used directly during the inference phase. During the inference phase, the true category... Unknown, using posterior probability For all possible real categories Calculate the weighted expectation to obtain the decision category. The expected loss is calculated, and the candidate material category with the smallest expected loss is selected as the final identification result. Represents the given polarization-spectral joint eigenvector. (i.e., the joint feature vector in step three), the true category of this material region is The probability is output by the material classification model in step three.

[0065] In some embodiments, after selecting the candidate material category with the minimum expected loss as the final identification result for the material region to be identified, the method further includes a step of hierarchical review of the identification result: Obtain the detection confidence of the material region to be identified in the target detection network, as well as the fusion confidence and expected loss value corresponding to the final identification result; If the detection confidence level is less than the preset detection threshold, the current image quality is determined to be unqualified, and a prompt message is generated to re-acquire or adjust the imaging angle. If the detection confidence is greater than or equal to the detection threshold, but the fusion confidence is less than the preset first confidence threshold, then the current recognition result is determined to be in a fuzzy state, and the corresponding material area to be identified is marked as an area to be manually reviewed. Subsequently, the material will be identified through manual screening and recorded in the database. This process is repeated to improve the system's recognition capability. If the fusion confidence level is greater than or equal to the first confidence level threshold, but the expected loss value is greater than the preset risk threshold, then it is determined that although the current identification result meets the confidence level, the carbon footprint accounting deviation risk is too high, and the corresponding material area to be identified is marked as an area to be reviewed. If the fusion confidence level is greater than or equal to the first confidence threshold, and the expected loss value is less than or equal to the risk threshold, then the current identification result is determined to be valid, and the final identification result is output to subsequent steps.

[0066] Step 7: Based on the final recognition results of all material regions to be identified, calculate the pixel percentage of each material category, and convert it into a mass percentage by combining the preset density of each candidate material category with the physical size of the material region to be identified.

[0067] In some embodiments, for each material region to be identified, based on its final identified material category, the standard density and porosity correction coefficients of the corresponding material are retrieved from the material carbon footprint factor library. Combined with the actual physical area of ​​the region obtained by imaging geometric calibration, the measured thickness synchronously acquired by the thickness measurement module, or the standard thickness of the corresponding component retrieved from the preset standard thickness library, the net volume of the region after deducting the pore volume is calculated, and the theoretical mass of the region is further obtained. The theoretical mass of the region is multiplied by the fusion confidence of the final identification result and the region pixel purity to obtain the weighted effective mass of the region. At the same time, abnormal regions and boundary transition pixels with a fusion confidence lower than the preset invalid threshold are removed. The weighted effective masses of all effective regions are summarized, and the total effective mass of each type of material is statistically obtained. Based on the measured total mass of the object being measured, the total effective mass of each type of material is proportionally normalized to obtain the final mass ratio of each type of material that is strictly conserved with the total mass of the object. The mass deviation value before and after normalization is output synchronously for the error range calibration of the subsequent carbon footprint calculation results.

[0068] Step 8: Obtain the total mass of the object being measured. Weight the mass percentage, total mass, and unit mass carbon footprint factor obtained from the material carbon footprint factor library to obtain the total carbon footprint.

[0069] In some embodiments, the measured total mass of the object under test, collected by the integrated weighing module, or the calibrated total mass manually input by the user, is obtained. The unit mass carbon footprint factor corresponding to the final identification result of each material region to be identified, along with the data confidence level and uncertainty parameters of this factor in the material carbon footprint factor library, are retrieved. First, the unit mass carbon footprint factor of each material is corrected; the corrected effective unit mass carbon footprint factor is equal to the original factor multiplied by the fusion confidence level of the material identification result. Subsequently, the final mass proportion of each material after mass conservation normalization in step seven, the corrected effective unit mass carbon footprint factor, and the measured total mass of the object under test are weighted and summed to obtain... The nominal value of the total carbon footprint is obtained. Based on this, regional pixel purity deviation, mass normalization deviation, imaging geometric measurement error, and inherent uncertainty of factor library data in the material identification process are introduced. Monte Carlo method is used to conduct full-link error propagation analysis, and data confidence level is introduced to calculate the upper and lower limits of the total carbon footprint at a 95% confidence level. Finally, the nominal value of the total carbon footprint, the confidence interval corresponding to the total carbon footprint, the correction factor details of each material, the error source decomposition, and the corresponding compliance statement are integrated into a standardized accounting traceability unit, which is encrypted and stored together with the original four-dimensional imaging data, intermediate results of material identification, and mass proportion calculation process.

[0070] Based on the same inventive concept, and corresponding to any of the above embodiments, refer to... Figure 2 This invention provides a carbon footprint detection system based on polarization imaging, used to implement the aforementioned carbon footprint detection method based on polarization imaging, comprising: The image acquisition module is used to simultaneously acquire polarized infrared and spectral images of the object under test; A data preprocessing module, connected to the image acquisition module, is used to preprocess the polarized infrared image and the spectral image to generate multidimensional feature data containing intensity features, polarization degree features, polarization angle features and spectral features. The material region segmentation module is connected to the data preprocessing module. It is used to receive the intensity features, polarization degree features and polarization angle features in the multi-dimensional feature data and perform multi-channel fusion. The fused feature map is input into the target detection network to locate and segment at least one material region to be identified. The feature extraction and classification module is connected to the material region segmentation module and the data preprocessing module, respectively. It is used to extract polarization feature vectors and spectral feature vectors from the material region to be identified, and input the polarization feature vectors and spectral feature vectors into the material classification model to output the posterior probability of each candidate material category. The storage module is used to store the standard polarization characteristics, standard spectral characteristics, preset density, carbon footprint factor, and data reliability level corresponding to each candidate material category of the candidate material. A multi-dimensional scoring calculation module is connected to the material region segmentation module and the storage module respectively. It is used to calculate the polarization similarity, spectral similarity, region pixel purity, and consistency score between the preset density and the physical size of the material region to be identified for each candidate material category, and to read the data credibility level. The fusion confidence calculation module is connected to the feature extraction and classification module, the multi-dimensional scoring calculation module, and the storage module, respectively, and is used to calculate the fusion confidence of each candidate material category based on the posterior probability, the polarization similarity, the spectral similarity, the regional purity, the consistency score, and the data credibility level. The cost-sensitive decision module is connected to the fusion confidence calculation module and the storage module respectively. It is used to construct a cost-sensitive loss function with carbon footprint factor bias as the cost of misclassification, and to select the expected loss of each candidate material category based on the fusion confidence of each candidate material category. The candidate material category with the smallest expected loss is selected as the final identification result of the material region to be identified. The carbon footprint calculation module is connected to the cost-sensitive decision module and the storage module respectively. It is used to calculate the pixel ratio of each material category based on the final identification results of all material regions to be identified, convert the preset density and the physical size of the material region to be identified into the mass ratio, and then combine the total mass of the measured object with the unit mass carbon footprint factor matched from the storage module to output the total carbon footprint.

[0071] Based on the same inventive concept, corresponding to any of the above embodiments, the present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the carbon footprint detection method based on polarization imaging of the embodiments.

[0072] Alternatively, the aforementioned electronic device may be a server.

[0073] In addition, this embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the carbon footprint detection method based on polarization imaging of the embodiment.

[0074] It is understood that the processor in the embodiments of the present invention may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0075] The method steps in the embodiments of the present invention can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0076] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a storage medium or transmitted through a storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive (SSD)).

Claims

1. A carbon footprint detection method based on polarization imaging, characterized in that, Includes the following steps: The polarized infrared image and spectral image of the object under test are acquired and preprocessed to obtain multidimensional feature data; wherein, the multidimensional feature data includes intensity features, polarization degree features, polarization angle features and spectral features; The intensity feature, polarization degree feature and polarization angle feature in the multidimensional feature data are fused in multiple channels and input into the target detection network to locate and segment at least one material region to be identified. The polarization feature vector and spectral feature vector of the material region to be identified are extracted and input into the material classification model to obtain the posterior probability of the material region to be identified corresponding to multiple candidate material categories; Calculate the polarization similarity between the polarization features of each material region to be identified and the standard polarization features of the candidate materials, the spectral similarity between the spectral features of each material region to be identified and the standard spectral features of the candidate materials, the region pixel purity, and the consistency score between the preset density of each candidate material category and the physical size of the material region to be identified; obtain the data credibility level corresponding to each candidate material category in the material carbon footprint factor library; Based on the posterior probability, polarization similarity, spectral similarity, regional purity, consistency score, and data confidence level, the fusion confidence of each candidate material category is calculated. A cost-sensitive loss function with carbon footprint factor bias as the cost of misclassification is constructed. Based on the fusion confidence of each candidate material category, the expected loss of selecting each candidate material category is calculated. The candidate material category with the smallest expected loss is selected as the final identification result of the material region to be identified. Based on the final recognition results of all material regions to be identified, the pixel percentage of each material category is calculated, and combined with the preset density of each candidate material category and the physical size of the material region to be identified, it is converted into a mass percentage. The total mass of the object under test is obtained. The total carbon footprint is obtained by weighted summing of the mass percentage, the total mass, and the unit mass carbon footprint factor matched from the material carbon footprint factor library.

2. The carbon footprint detection method based on polarization imaging as described in claim 1, characterized in that, The specific process of fusing intensity features, polarization degree features, and polarization angle features from multi-dimensional feature data into multiple channels, inputting the data into the target detection network, and locating and segmenting at least one region of material to be identified is as follows: The intensity feature, polarization degree feature, and polarization angle feature in the multidimensional feature data are normalized to the same spatial resolution and then stitched together in the channel dimension to form a multi-channel fused feature map; wherein, the multi-channel fused feature map includes: intensity feature map, polarization degree feature map, and polarization angle feature map; The multi-channel fused feature map is input into the target detection network, which uses YOLOv8 as the baseline architecture. A polarization feature extraction branch is added to the backbone network of the target detection network. The polarization feature extraction branch performs convolution and downsampling on the polarization degree feature map and polarization angle feature map in the multi-channel fused feature map to extract polarization texture features that reflect the anisotropy and microstructure of the material surface. These features are then fused layer by layer with the intensity feature map of the corresponding scale in the backbone network to obtain enhanced multimodal features. The target detection network is based on multimodal features. It regresses the bounding box of the object under test through feature pyramid and detection head, performs pixel-level semantic segmentation on the region within the bounding box, generates a foreground mask, and removes background regions, shadow regions and coating coverage regions to obtain at least one material region to be identified.

3. The carbon footprint detection method based on polarization imaging as described in claim 1, characterized in that, The specific process for calculating the regional pixel purity of each region of the material to be identified and the standard spectral characteristics of the candidate material is as follows: Extract the spectral feature vectors of all pixels within the region of the material to be identified, and construct a regional spectral feature set; For any candidate material category, its standard spectral feature vector is obtained, and the spectral angle between the spectral feature vector of each pixel in the region spectral feature set and the standard spectral feature vector is calculated to generate a pixel-level spectral angle sequence. The arithmetic mean of the pixel-level spectral angle sequence is used as the spectral similarity between the candidate material category and the region of the material to be identified; The number of pixels with spectral angles less than a preset consistency threshold in the pixel-level spectral angle sequence is counted, and the proportion of this number of pixels to the total number of pixels in the material region to be identified is calculated. This proportion is used as the region pixel purity corresponding to the candidate material category.

4. The carbon footprint detection method based on polarization imaging as described in claim 1, characterized in that, The specific process for calculating the consistency score between the preset density of each candidate material category and the physical size of the material region to be identified is as follows: The calibration pixel size, focal length, and synchronously acquired object distance of the polarized infrared imaging device are obtained. Based on the object distance and imaging geometry, the actual physical side length corresponding to a single pixel is determined, and the pixel area of ​​the material region to be identified is converted into the actual area. Obtain the thickness of the material region to be identified; wherein, the thickness is the measured thickness synchronously obtained by the measuring device, or the standard thickness corresponding to the object category to which the material region to be identified belongs, retrieved from a preset material thickness library; Multiply the actual area by the thickness to obtain the measured volume of the material region to be identified; Extract the preset density of candidate material categories from the material carbon footprint factor library, calculate the estimated mass when the material region to be identified is assumed to be composed of candidate material categories, and obtain the measured total mass of the object being measured. When the object under test contains multiple regions of material to be identified, the estimated quality of each region other than the current region under the current best candidate material is fixed. The target material is iteratively screened through preliminary optimization screening. The total estimated quality is obtained when the current region is assumed to be a candidate material category, and the quality deviation is calculated. When the object under test contains only a single region of material to be identified, the estimated quality of that region is directly used to calculate the quality deviation with the measured total quality. The quality deviation is mapped to a consistency score using a preset monotonically decreasing scoring function. A smaller deviation results in a higher score, characterizing the physical consistency between the preset density of the candidate material category and the physical dimensions of the material region to be identified. The consistency score ranges from [value missing]. .

5. The carbon footprint detection method based on polarization imaging as described in claim 1, characterized in that, The specific process for calculating the fusion confidence of each candidate material category is as follows: First to sixth weight coefficients are assigned to the posterior probability, the polarization similarity, the spectral similarity, the regional purity, the consistency score, and the data credibility level, respectively, and the sum of the first to sixth weight coefficients is 1. For any candidate material category, the posterior probability, polarization similarity, spectral similarity, regional purity, consistency score, and data credibility level of the candidate material category are multiplied by their respective weight coefficients, and the results of each product are added together to obtain the comprehensive score corresponding to the candidate material category. Based on the comprehensive scores of all candidate material categories, a normalized exponential function is applied to all comprehensive scores to obtain the normalized value of each candidate material category. The normalized value is then used as the fusion confidence of the candidate material category.

6. The carbon footprint detection method based on polarization imaging as described in claim 1, characterized in that, The specific process of constructing a cost-sensitive loss function with carbon footprint factor bias as the cost of misclassification is as follows: Extract the unit mass carbon footprint factor corresponding to each candidate material category and the data reliability level of the factor library from the material carbon footprint factor library; For decision categories Compared to the real category Construct a cost-sensitive loss function of the following form. : in, For indicator functions, in The value is 1 if the condition is met, otherwise it is 0. and Decision categories and real category Carbon footprint factor per unit mass; To prevent extremely small constants with a denominator of zero; For decision categories The data reliability level in the materials carbon footprint factor library, and the value range is [missing information]. ; , and These are the non-negative weight coefficients for the corresponding terms.

7. The carbon footprint detection method based on polarization imaging as described in claim 1, characterized in that, After selecting the candidate material category with the minimum expected loss as the final identification result for the material region to be identified, the method further includes a step of hierarchical review of the identification result: Obtain the detection confidence of the material region to be identified in the target detection network, as well as the fusion confidence and expected loss value corresponding to the final identification result; If the detection confidence level is less than the preset detection threshold, the current image quality is determined to be unqualified, and a prompt message is generated to re-acquire or adjust the imaging angle. If the detection confidence is greater than or equal to the detection threshold, but the fusion confidence is less than the preset first confidence threshold, then the current recognition result is determined to be in a fuzzy state, and the corresponding material area to be recognized is marked as an area to be manually reviewed. If the fusion confidence level is greater than or equal to the first confidence level threshold, but the expected loss value is greater than the preset risk threshold, then it is determined that although the current identification result meets the confidence level, the carbon footprint accounting deviation risk is too high, and the corresponding material area to be identified is marked as the area to be reviewed. If the fusion confidence level is greater than or equal to the first confidence threshold, and the expected loss value is less than or equal to the risk threshold, then the current identification result is determined to be valid, and the final identification result is output to subsequent steps.

8. A carbon footprint detection system based on polarization imaging, used to implement the carbon footprint detection method based on polarization imaging as described in any one of claims 1-7, characterized in that, include: The image acquisition module is used to simultaneously acquire polarized infrared and spectral images of the object under test; A data preprocessing module, connected to the image acquisition module, is used to preprocess the polarized infrared image and the spectral image to generate multidimensional feature data containing intensity features, polarization degree features, polarization angle features and spectral features. The material region segmentation module is connected to the data preprocessing module. It is used to receive the intensity features, polarization degree features and polarization angle features in the multi-dimensional feature data and perform multi-channel fusion. The fused feature map is input into the target detection network to locate and segment at least one material region to be identified. The feature extraction and classification module is connected to the material region segmentation module and the data preprocessing module, respectively. It is used to extract polarization feature vectors and spectral feature vectors from the material region to be identified, and input the polarization feature vectors and spectral feature vectors into the material classification model to output the posterior probability of each candidate material category. The storage module is used to store the standard polarization characteristics, standard spectral characteristics, preset density, carbon footprint factor, and data reliability level corresponding to each candidate material category of the candidate material. A multi-dimensional scoring calculation module is connected to the material region segmentation module and the storage module respectively. It is used to calculate the polarization similarity, spectral similarity, region pixel purity, and consistency score between the preset density and the physical size of the material region to be identified for each candidate material category, and to read the data credibility level. The fusion confidence calculation module is connected to the feature extraction and classification module, the multi-dimensional scoring calculation module, and the storage module, respectively, and is used to calculate the fusion confidence of each candidate material category based on the posterior probability, the polarization similarity, the spectral similarity, the regional purity, the consistency score, and the data credibility level. The cost-sensitive decision module is connected to the fusion confidence calculation module and the storage module respectively. It is used to construct a cost-sensitive loss function with carbon footprint factor bias as the cost of misclassification, and to select the expected loss of each candidate material category based on the fusion confidence of each candidate material category. The candidate material category with the smallest expected loss is selected as the final identification result of the material region to be identified. The carbon footprint calculation module is connected to the cost-sensitive decision module and the storage module respectively. It is used to calculate the pixel ratio of each material category based on the final identification results of all material regions to be identified, convert the preset density and the physical size of the material region to be identified into the mass ratio, and then combine the total mass of the measured object with the unit mass carbon footprint factor matched from the storage module to output the total carbon footprint.

9. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the carbon footprint detection method based on polarization imaging as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the carbon footprint detection method based on polarization imaging as described in any one of claims 1-7.

Images