Electronic warehouse receipt goods authenticity verification method based on AI image recognition

Abstract

The invention relates to the technical field of block chains and information security, and particularly discloses an electronic warehouse receipt cargo authenticity verification method based on AI image recognition, which comprises the following steps: acquiring a cargo polarized light image sequence, a multi-angle visible light image and environment information, and generating binary visual physical feature information through hierarchical feature purification and fusion; the method comprises the following steps of: collecting and processing data according to a same protocol to generate a to-be-verified feature, dynamically distributing a feature comparison weight and an adjustment judgment threshold value based on environment information, and outputting a verification result through comparison with registration information stored in a block chain when the to-be-verified feature and the electronic warehouse receipt information are bound to generate a registration hash value and then are jointly written into the block chain for evidence storage, and when verification is triggered, collecting and processing data according to the same protocol to generate the to-be-verified feature, and dynamically distributing the feature comparison weight and the adjustment judgment threshold value based on environment information. According to the method, the inherent physical characteristics of the goods are extracted by using the multi-modal data, and the anti-counterfeiting capability and the environmental adaptability of verification are improved by combining the non-tampering property of the block chain and the dynamic adaptation strategy, so that the end-to-end credible closed loop of the physical goods and the digital warehouse receipt is constructed, and the safety of supply chain finance and movable property financing is ensured.

Description

Technical Field

[0001] This invention relates to the fields of blockchain and information security technology, and in particular to a method for verifying the authenticity of goods in electronic warehouse receipts based on AI image recognition. Background Technology

[0002] With the rapid development of supply chain finance and movable asset financing, electronic warehouse receipts, as important digital certificates of physical assets, rely on the authenticity and uniqueness of the physical goods they correspond to for their core value. Therefore, how to efficiently and reliably verify the authenticity of goods under electronic warehouse receipts has become a key technical challenge in ensuring financial security and transaction trust.

[0003] Currently, common cargo verification methods suffer from the following limitations: traditional methods rely on manual on-site inspections and paper documents, resulting in low efficiency, strong subjectivity, easy tampering of records, and difficulty in traceability. Furthermore, existing technical solutions typically employ image recognition methods based on ordinary visible light, comparing images of the cargo's appearance or labels. These methods rely solely on easily counterfeited features such as color and shape, failing to penetrate the surface to obtain the cargo's intrinsic physical properties. They are easily deceived by high-quality imitations, copied labels, or photographic reproductions, resulting in weak anti-counterfeiting capabilities. The feature extraction process of existing methods is usually fixed and simplistic, lacking adaptability to complex data collection environments, leading to unstable verification accuracy. Existing solutions also have shortcomings in data storage and process reliability. They only store easily tamper-proof local data, or, although using blockchain, only store hash values. During verification, there is a lack of directly comparable and tamper-proof original feature benchmarks, making it difficult to construct an end-to-end trusted closed loop from physical goods to digital warehouse receipts.

[0004] Therefore, there is an urgent need for an electronic warehouse receipt verification method based on AI image recognition to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide a method for verifying the authenticity of goods in electronic warehouse receipts based on AI image recognition, comprising the following steps: Acquire multimodal image data of the goods to be verified and information on the acquisition environment. The multimodal image data includes polarized light image sequences and multi-angle visible light images. The multimodal image data is subjected to hierarchical feature purification and fusion processing to generate binary visual physical feature information representing the unique physical attributes of the goods; The visual physical feature information is bound to the corresponding electronic warehouse receipt information to generate a registration hash value. The visual physical feature information and the registration hash value are then written into the blockchain for storage, thus completing the registration and evidence storage of the cargo features. When verification is triggered, multimodal image data of the goods are collected according to the same protocol, and the hierarchical feature purification and fusion processing is performed to generate visual physical feature information to be verified. Based on the real-time collected environmental information, the feature weights and decision thresholds are dynamically calculated. The visual physical feature information to be verified is compared with the registration information stored in the blockchain, and the authenticity verification result is output.

[0006] Furthermore, the steps of acquiring multimodal image data of the goods to be verified and collecting environmental information include: By using an image acquisition device integrated with a polarizing filter, surface images of the goods are acquired at at least two different polarization angles to form a polarized light image sequence. Under the illumination of multiple controllable light sources with known spatial relationships, visible light images of the cargo are acquired from at least two different perspectives to form a multi-angle visible light image set, and the on / off state and position coordinates of each light source are recorded simultaneously. The ambient light intensity, ambient temperature and humidity data are acquired synchronously at the time of data collection using environmental sensors. Based on the multi-angle visible light image set, the three-dimensional point cloud of the cargo surface is calculated using a multi-view stereo vision algorithm, and the overall positioning accuracy of this acquisition is evaluated based on the reprojection error of the three-dimensional point cloud, generating a positioning accuracy evaluation value.

[0007] Furthermore, the step of performing hierarchical feature extraction and fusion processing on the multimodal image data to generate binary visual physical feature information representing the unique physical attributes of the goods includes: The polarized light image sequence is fused to calculate the degree of polarization image and the angle of polarization image, and the local binary mode feature vector is extracted from the degree of polarization image as the polarization feature. For the multi-angle visible light image set, the normal map of the cargo surface is recovered using a photometric stereo vision algorithm, and the statistical histogram feature vector of the normal direction is calculated as a visible light feature. Based on the cargo specification database, the standard texture period and surface roughness range are obtained, the degree of conformity between the polarization characteristics and the standard specifications is calculated, and the specification conformity coefficient is obtained. Based on the multi-angle visible light image set and the known light source position, the physical consistency of the collected data is evaluated by verifying whether the continuity of shadows and highlights conforms to the preset optical reflection model, and the physical consistency coefficient is obtained. The polarization characteristics and visible light characteristics are then weighted and corrected using the specification compliance coefficient and the physical consistency coefficient, respectively. The corrected polarization features are concatenated with the visible light features to obtain a high-dimensional fused feature vector. A pre-trained autoencoder is then used to reduce the dimensionality of the high-dimensional fused feature vector to obtain a low-dimensional feature vector. For each dimension of the low-dimensional feature vector, it is compared with the historical statistical mean of that dimension. Based on the comparison result, it is converted into a binary value to generate binary visual physical feature information.

[0008] Furthermore, the step of jointly writing the visual physical feature information and the registration hash value into the blockchain for storage to complete the registration and notarization of the cargo features includes: Based on the collected environment information, obtain the electronic warehouse receipt number, the collection timestamp, and the image acquisition device identifier; The visual physical feature information, electronic warehouse receipt number, collection timestamp, and image acquisition device identifier are concatenated together; Apply a cryptographic hash algorithm to the concatenated data stream to generate a unique registration hash value; The visual physical feature information and the registration hash value are encapsulated together into a transaction data, uploaded to a pre-built blockchain network for storage and consensus verification, and the immutable binding of the physical features of the goods and the electronic warehouse receipt is completed.

[0009] Furthermore, the step of collecting multimodal image data of the goods according to the same protocol and performing the hierarchical feature purification and fusion processing to generate visual physical feature information to be verified when verification is triggered includes: Receive verification instructions for a specific electronic warehouse receipt, and according to the cargo identifier and acquisition protocol identifier carried in the instructions, control the image acquisition device to re-acquire the polarized light image sequence and multi-angle visible light image of the cargo according to the same parameter settings as in the registration stage, and simultaneously acquire the current acquisition environment information. For the re-acquired multimodal image data, perform the same feature purification and fusion processing flow as in the registration phase to generate visual physical feature information to be verified; Based on the re-acquired environmental information and multimodal image data, the physical consistency coefficient and positioning accuracy assessment value of the current acquisition are recalculated as the basis for evaluating the reliability of the acquired data in this verification.

[0010] Furthermore, the step of comparing the visual physical feature information to be verified with the registration information stored in the blockchain and outputting the authenticity verification result includes: Based on the physical consistency coefficient and the positioning accuracy evaluation value, the comparison weights of polarization features and visible light features are dynamically allocated, and the weights of polarization features and visible light features are obtained. The higher the physical consistency coefficient and the higher the positioning accuracy evaluation value, the greater the weight allocated to the polarization features. At the same time, it is ensured that the sum of the weights of polarization features and visible light features is a fixed value. Calculate the similarity between the visual physical feature information to be verified and the registered visual physical feature information retrieved from the blockchain in the polarization feature segment and the visible light feature segment, respectively. Based on the polarization feature weights and visible light feature weights, the similarities of the polarization feature sub-segments and visible light feature sub-segments are weighted and summed to obtain a weighted comprehensive similarity. The judgment threshold will be dynamically adjusted based on the preset risk level of this verification business and the current overall collection quality. The higher the risk level or the lower the overall collection quality, the higher the judgment threshold will be. The weighted comprehensive similarity is compared with the dynamically adjusted judgment threshold. If the weighted comprehensive similarity is greater than or equal to the judgment threshold, the goods are judged to be genuine and the verification is passed. If the weighted comprehensive similarity is less than the judgment threshold, the goods are judged to be fake and the verification is failed. The key data and decision results of this verification process are then used to generate a new hash value and stored on the blockchain.

[0011] Furthermore, this application also discloses an electronic warehouse receipt goods authenticity verification system based on AI image recognition, including: The acquisition module is used to acquire multimodal image data of the goods to be verified and information on the acquisition environment. The multimodal image data includes polarized light image sequences and multi-angle visible light images. The generation module is used to perform hierarchical feature purification and fusion processing on the multimodal image data to generate binary visual physical feature information representing the unique physical attributes of the goods. The registration module is used to bind the visual physical feature information with the corresponding electronic warehouse receipt information, generate a registration hash value, and write the visual physical feature information and the registration hash value together into the blockchain for storage, thereby completing the registration and evidence storage of the cargo features. The fusion module is used to collect multimodal image data of the goods according to the same protocol when verification is triggered, and to perform the hierarchical feature purification and fusion processing to generate visual physical feature information to be verified. The output module is used to dynamically calculate feature weights and decision thresholds based on real-time collected environmental information, compare the visual physical feature information to be verified with the registration information stored in the blockchain, and output the authenticity verification result.

[0012] Furthermore, the registration module includes: The acquisition unit is used to acquire the electronic warehouse receipt number, acquisition timestamp, and image acquisition device identifier based on the acquisition environment information. The splicing unit is used to splice visual physical feature information, electronic warehouse receipt number, acquisition timestamp, and image acquisition device identifier; The generation unit is used to apply a cryptographic hash algorithm to the concatenated data stream to generate a unique registration hash value. The binding unit is used to encapsulate the visual physical feature information and the registration hash value into a transaction data, upload it to the pre-built blockchain network for storage and consensus verification, and complete the tamper-proof binding of the physical features of the goods with the electronic warehouse receipt.

[0013] This application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method for verifying the authenticity of goods in electronic warehouse receipts based on AI image recognition.

[0014] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method for verifying the authenticity of goods in electronic warehouse receipts based on AI image recognition.

[0015] The beneficial effects of this application are as follows: Firstly, this invention extracts inherent physical features that cannot be cloned, such as microscopic texture and three-dimensional morphology, from the surface of goods by fusing polarized light images and multi-angle visible light images, generating a unique visual fingerprint. This fundamentally resists common fraud methods such as appearance counterfeiting, label duplication, and image re-photography. Through hierarchical feature purification, feature correction guided by prior knowledge, and a dynamic weighted decision-making mechanism, it can adapt to environmental changes and distinguish feature reliability, greatly improving the accuracy and stability of authenticity determination.

[0016] Secondly, by constructing an end-to-end trusted verification closed loop, the present invention stores the generated visual physical feature information and its hash value together on the blockchain, ensuring that the feature benchmark and verification record are tamper-proof and traceable throughout the entire process. This achieves verifiable trust across the entire chain from physical goods to digital warehouse receipts. By evaluating the collection quality in real time and dynamically allocating feature weights and adjusting judgment thresholds accordingly, the invention can maintain optimal performance in complex real-world scenarios, balancing verification efficiency and security risks. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a method flow proposed in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the system structure proposed in an embodiment of the present invention.

[0019] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0021] like Figure 1As shown, this application provides a method for verifying the authenticity of goods in electronic warehouse receipts based on AI image recognition, including the following steps: S1, acquire multimodal image data of the goods to be verified and information on the acquisition environment, wherein the multimodal image data includes polarized light image sequences and multi-angle visible light images; S2, perform hierarchical feature purification and fusion processing on the multimodal image data to generate binary visual physical feature information representing the unique physical attributes of the goods; S3, bind the visual physical feature information with the corresponding electronic warehouse receipt information, generate a registration hash value, and write the visual physical feature information and the registration hash value together into the blockchain for storage, thereby completing the registration and evidence storage of the cargo features; S4. When verification is triggered, multimodal image data of the goods are collected according to the same protocol, and the hierarchical feature purification and fusion processing is performed to generate visual physical feature information to be verified. S5 dynamically calculates feature weights and decision thresholds based on real-time collected environmental information, compares the visual physical feature information to be verified with the registration information stored in the blockchain, and outputs the authenticity verification result.

[0022] As described in steps S1-S5 above, the core financial value of electronic warehouse receipts lies in the authenticity and uniqueness of the physical goods. The physical attributes of goods are irreproducible. By extracting the inherent physical attribute characteristics of the goods and binding them to the warehouse receipt for evidence storage, fraudulent activities such as appearance counterfeiting and label tampering can be prevented at the source. However, complex collection environments can affect the accuracy of feature extraction, requiring dynamic adjustment strategies to ensure verification stability, ultimately solving the problem of credibility and reliability in goods authenticity verification in supply chain finance. Traditional manual verification is inefficient and records are easily tampered with. Existing single visible light image recognition relies solely on appearance features, resulting in weak anti-counterfeiting capabilities. Fixed feature extraction methods lack environmental adaptability, and evidence storage methods struggle to form a reliable closed loop. This invention employs multimodal data acquisition using polarized light and multi-angle visible light to extract the inherent physical characteristics of the goods. Combined with hierarchical purification and fusion technology, it strengthens the uniqueness of the features. Blockchain is used to achieve an immutable binding of features to the warehouse receipt. Furthermore, the comparison strategy is dynamically adjusted based on the collection environment, comprehensively improving the anti-counterfeiting capability and environmental adaptability of verification.

[0023] This invention achieves highly reliable and counterfeit-resistant verification of goods corresponding to electronic warehouse receipts through a closed-loop process of multimodal image data acquisition, hierarchical feature extraction and fusion, blockchain notarization, and dynamic threshold comparison. It establishes an end-to-end trusted association between physical goods and digital warehouse receipts. The core logic of this invention is to build a closed-loop process around the trusted association between electronic warehouse receipts and physical goods. In principle, it leverages the uncopyability of the inherent physical attributes of goods and the immutability of blockchain. Addressing the problems of low efficiency in traditional manual verification, weak anti-counterfeiting capabilities of existing single-image recognition, lack of environmental adaptability in feature extraction, and insufficient credibility of notarization, this invention achieves highly reliable authenticity verification through a coherent design of data acquisition, feature processing, notarization, verification, and comparison decision-making. First, it acquires multimodal data of polarized light image sequences and multi-angle visible light images, along with acquisition environment information, to lay the foundation for extracting the unique physical attributes of the goods. Then, through further analysis... The process involves high-level feature purification and fusion, eliminating interfering information and strengthening core features to generate binary visual physical feature information that represents the essential attributes of goods. This feature information is then bound to electronic warehouse receipt information to generate a registration hash value, which is jointly written into the blockchain for immutable notarization, providing a reliable benchmark for subsequent verification. When verification is triggered, data is collected and processed strictly according to the registration phase protocol to ensure the comparability of the features to be verified with the registered features. Finally, feature weights and decision thresholds are dynamically calculated based on real-time environmental information. By comparing the features to be verified with the registration information stored on the blockchain, accurate verification results are output. This entire process overcomes the limitations of easily counterfeited appearance features through multimodal data and feature processing, and solves problems such as environmental interference and data tampering by leveraging blockchain notarization and dynamic adaptation strategies, thus constructing an end-to-end reliable verification system from physical goods to digital warehouse receipts.

[0024] In one embodiment, the step of acquiring multimodal image data of the goods to be verified and collecting environmental information includes: S11, using an image acquisition device integrated with a polarizing filter, surface images of the goods are acquired at at least two different polarization angles to form a polarized light image sequence; S12, under the illumination of multiple controllable light sources with known spatial relationships, acquire visible light images of the cargo from at least two different perspectives to form a multi-angle visible light image set, and simultaneously record the on / off state and position coordinates of each light source. S13, synchronously acquires ambient light intensity and ambient temperature and humidity data at the time of data collection through environmental sensors; S14. Based on the multi-angle visible light image set, the three-dimensional point cloud of the cargo surface is calculated using a multi-view stereo vision algorithm, and the overall positioning accuracy of this acquisition is evaluated based on the reprojection error of the three-dimensional point cloud, generating a positioning accuracy evaluation value.

[0025] As described in steps S11-S14 above, by integrating specific acquisition devices and sensing equipment, multi-polarization angle images, multi-view visible light images, acquisition environment parameters, and positioning accuracy evaluation values ​​of the cargo are acquired simultaneously. This provides comprehensive, reliable, and traceable raw data support for subsequent graded feature purification and fusion and dynamic comparison, ensuring that the extracted visual physical features can truly reflect the unique physical attributes of the cargo.

[0026] The core of authenticating goods lies in their inherent, unreplicable physical properties, which cannot be fully captured by a single modality or viewpoint image. Polarized light characteristics are directly related to the material of the goods, and multi-angle visible light can reveal spatial structural information. Furthermore, the ambient light, temperature, and humidity affect image quality, while the accuracy of image acquisition and positioning directly impacts the consistency of feature acquisition. Therefore, comprehensive acquisition of this data is necessary to address the weakness in anti-counterfeiting capabilities caused by existing technologies relying solely on superficial features. Traditional solutions only use ordinary visible light for single-view image acquisition, failing to record environmental information or assess positioning accuracy. This results in data that cannot reflect the inherent physical properties of the goods, is susceptible to environmental interference, lacks reliable data, and struggles to guarantee anti-counterfeiting capabilities and verification accuracy. By employing a multi-modal acquisition scheme, combined with environmental sensing and positioning accuracy assessment, the shortcomings of existing technologies—insufficient data acquisition dimensions and lack of reliability—are specifically addressed, laying a high-quality data foundation for subsequent feature processing and verification comparison.

[0027] A series of polarized light images is formed by acquiring images of the cargo surface at at least two different polarization angles using an image acquisition device integrated with polarizing filters. This device has a built-in switchable polarizing filter with two preset fixed polarization angles of 0° and 90°. During acquisition, the filter angles are switched sequentially and images are captured synchronously, ensuring that images with different polarization states are obtained from the same acquisition location. The optical reflectivity of the cargo surface material exhibits a unique response to changes in polarization angle. For example, the intensity of reflected light from metal and plastic materials differs significantly at the same polarization angle, and this difference cannot be replicated through appearance counterfeiting. By acquiring images at multiple polarization angles, the inherent optical properties of the cargo material can be captured, providing data support for subsequent extraction of anti-counterfeiting polarization features, effectively improving the uniqueness of the features and anti-counterfeiting capabilities.

[0028] Under the illumination of multiple controllable light sources with known spatial relationships, visible light images of the cargo are acquired from at least two different perspectives, and the light source status and position coordinates are recorded. The controllable light sources consist of three evenly distributed LED light sources, fixed at preset intervals, whose spatial coordinates are pre-entered into the system. During acquisition, the light sources are controlled to turn on and off according to a set program. Simultaneously, two cameras fixed on different supports capture images from the front and side perspectives respectively. Multi-view acquisition avoids occlusion problems caused by cargo stacking or its own structure from a single perspective. For example, the texture features of the cargo's side may be obscured from the front view, but can be fully captured from the side view. The use of controllable light sources reduces image distortion caused by uneven ambient lighting. Recording the light source status and position coordinates provides a crucial reference for subsequent verification of the physical consistency of the acquired data, ensuring that the visible light images comprehensively and clearly present the cargo's surface texture and spatial structure information.

[0029] The system synchronously acquires ambient light intensity, temperature, and humidity data at the time of data collection using environmental sensors. The environmental sensors and image acquisition device share the same trigger signal, and their sampling frequency matches the image acquisition frequency, ensuring complete synchronization of data timestamps. Ambient light intensity affects image brightness and contrast; excessively high or low light levels can lead to blurred image features. Ambient temperature and humidity can affect the surface condition of goods and the operational stability of the image acquisition device. For example, condensation may occur on the surface of goods in high humidity environments, causing image texture distortion. Synchronously acquiring this environmental data provides a basis for feature correction during subsequent feature refinement, improving the adaptability of features to environmental changes and ensuring the comparability of data collected under different environments.

[0030] Based on a multi-angle visible light image set, a multi-view stereo vision algorithm is used to calculate the 3D point cloud of the cargo surface. The overall positioning accuracy of the acquisition is then evaluated based on the reprojection error of the 3D point cloud, generating a positioning accuracy assessment value. The multi-view stereo vision algorithm first detects and matches feature points in visible light images from different perspectives. Triangulation is then used to obtain the 3D coordinates of the cargo surface, forming a dense 3D point cloud. The reprojection error calculation involves reprojecting the generated 3D point cloud onto the images from each perspective, calculating the pixel deviation between the projected points and the matching feature points in the original image, and statistically analyzing the average deviation of all feature points as the reprojection error. When the average deviation is less than 0.5 pixels, the positioning accuracy assessment value is considered excellent; between 0.5 and 1 pixel is considered good; and greater than 1 pixel is considered poor. The 3D point cloud directly reflects the spatial structure and dimensions of the cargo, while the reprojection error directly reflects the accuracy of the acquisition perspective and positioning. The positioning accuracy assessment value provides a data reliability standard for subsequent feature extraction and comparison. If the assessment value is poor, re-acquisition can be triggered promptly to avoid feature deviation due to inaccurate positioning, ensuring the consistency and reliability of subsequent feature acquisition.

[0031] In one embodiment, the step of performing hierarchical feature extraction and fusion processing on the multimodal image data to generate binary visual physical feature information representing the unique physical attributes of the goods includes: S21, perform fusion processing on the polarized light image sequence to calculate the polarization degree image and polarization angle image, and extract the local binary mode feature vector from the polarization degree image as polarization feature; S22, For the multi-angle visible light image set, the normal map of the cargo surface is recovered using a photometric stereo vision algorithm, and the statistical histogram feature vector of the normal direction is calculated as a visible light feature. S23, Based on the cargo specification database, obtain the standard texture period and surface roughness range, calculate the degree of conformity between the polarization characteristics and the standard specifications, and obtain the specification conformity coefficient; S24. Based on the multi-angle visible light image set and the known light source position, the physical consistency of the collected data is evaluated by verifying whether the continuity of shadows and highlights conforms to the preset optical reflection model, and the physical consistency coefficient is obtained. The polarization characteristics and visible light characteristics are weighted and corrected by the specification conformity coefficient and the physical consistency coefficient, respectively. S25, the corrected polarization features are concatenated with the visible light features to obtain a high-dimensional fused feature vector. A pre-trained autoencoder is then used to reduce the dimensionality of the high-dimensional fused feature vector to obtain a low-dimensional feature vector. S26. For each dimension of the low-dimensional feature vector, compare it with the historical statistical mean of that dimension, and convert it into a binary value based on the comparison result to generate binary visual physical feature information.

[0032] As described in steps S21-S26 above, core features are extracted from polarized light image sequences and multi-angle visible light images respectively. After specification verification and physical consistency verification, the features are purified in a graded manner. Then, through feature fusion, dimensionality reduction and binarization processing, visual physical feature information that can accurately represent the unique physical attributes of goods is generated, providing highly reliable and interference-resistant core data support for subsequent blockchain evidence storage and authenticity comparison.

[0033] The authentication of goods relies on their inherent, non-replicable physical properties. While multimodal image data can capture diverse information such as material and structure, environmental interference, data redundancy, and abnormal components that do not conform to the goods' specifications are inevitably introduced during the acquisition process. Directly using the original features for comparison leads to a decrease in verification accuracy and a weakening of anti-counterfeiting capabilities. Therefore, it is necessary to perform hierarchical purification and fusion processing to eliminate invalid information and enhance effective features, ensuring that the final generated feature information stably reflects the inherent physical properties of the goods. Existing technologies only perform simple feature extraction on single-modal images, lacking verification and correction mechanisms for feature validity and failing to reasonably fuse the complementarity of multimodal data. This results in extracted features that are susceptible to interference and lack uniqueness, making them difficult to resist counterfeiting and fraud. By extracting features separately, performing hierarchical verification, weighted correction, fusion dimensionality reduction, and binarization, this approach specifically addresses the problems of coarse feature extraction, weak anti-interference capabilities, and insufficient uniqueness in existing technologies, achieving accurate transformation from multimodal raw data to high-value feature information.

[0034] Polarized light image sequences are fused to obtain polarization degree images and polarization angle images. Local binary pattern feature vectors are extracted from the polarization degree images as polarization features. Polarized light image fusion employs a weighted average algorithm based on pixel polarization information. The weights are determined according to the sharpness of each polarization angle image, which is evaluated by calculating the sum of image gradient magnitudes; the higher the sum of gradient magnitudes, the greater the weight. After fusion, the polarization degree and polarization angle of each pixel are calculated using the formula for the physical properties of polarized light, forming the corresponding image. Local binary pattern feature extraction uses a 3x3 neighborhood window. The gray value of the center pixel of the window is used as a threshold and compared one by one with the gray values ​​of the surrounding 8 pixels. If the gray value of the surrounding pixels is greater than or equal to that of the center pixel, it is recorded as 1; otherwise, it is recorded as 0. The 8 binary numbers are combined in a clockwise order to obtain the local binary pattern code of that pixel. The codes of the entire image are then statistically analyzed to form a feature vector. The optical reflection properties of goods materials will exhibit a unique distribution in polarization degree and polarization angle images. For example, the polarization degree values ​​of metal materials and plastic materials are significantly different, and this difference cannot be changed by appearance counterfeiting. Local binary mode features can effectively capture this microscopic material feature, and the extracted polarization feature has strong anti-counterfeiting properties and can resist fraudulent methods such as surface label counterfeiting and high appearance counterfeiting.

[0035] A photometric stereo vision algorithm is used to reconstruct the normal map of the cargo surface from a multi-angle visible light image set. The statistical histogram feature vector of the normal direction is calculated as a visible light feature. Based on the known position coordinates and on / off states of three controllable LED light sources, the photometric stereo vision algorithm analyzes the grayscale changes of the same cargo area under different light source illuminations. Combined with the Lambertian reflection model, the surface normal vector of each pixel is calculated, and the set of normal vectors from all pixels constitutes the cargo surface normal map. During the extraction of the statistical histogram feature of the normal direction, the azimuth and elevation angles of the normal direction are divided into 36 equal intervals, and the proportion of normal vectors in each interval is statistically analyzed to form a 36-dimensional feature vector. Multi-angle visible light images can comprehensively capture the three-dimensional structural information of the cargo surface. The normal map directly reflects the concavity and convexity of the cargo surface and its spatial orientation. For example, the corners and grooves of boxed cargo will exhibit unique normal vector distributions. The statistical histogram feature of the normal direction transforms this three-dimensional structural information into quantitative features, complementing the polarization features and jointly enriching the representation dimensions of the cargo's physical properties.

[0036] Based on the cargo specification database, standard texture periods and surface roughness ranges are obtained. The degree of conformity between polarization features and standard specifications is calculated to obtain a specification conformity coefficient. The cargo specification database is a pre-established standardized data storage module containing standard texture period intervals and surface roughness ranges for various types of cargo, determined by authoritative testing. For the cargo to be verified, its corresponding standard specification parameters are retrieved through the electronic warehouse receipt number. The specification conformity coefficient is calculated using the Euclidean distance normalization method. First, the Euclidean distance between the extracted polarization features and standard specification features is calculated, and then mapped to the interval between 0 and 1. The smaller the distance, the closer the specification conformity coefficient is to 1, and vice versa. The formula for calculating the specification conformity coefficient is: ; Among them, the This represents the specification conformity coefficient (range 0~1), quantifying the degree of fit between polarization characteristics and standard cargo specifications. It is used to weight and correct polarization characteristics, eliminating non-compliant outliers. This represents the i-th dimension component of the extracted polarization feature vector, derived from the polarization feature vector extracted by LBP as described above, with dimension n. This represents the i-th dimension of the standard polarization feature vector in the cargo specification database. The source is a pre-established cargo specification database. The standard features corresponding to the standard texture period and surface roughness of the corresponding cargo are retrieved through the electronic warehouse receipt number. The validity of the polarization feature is verified by the standard specifications of the cargo itself. For example, if the standard texture period of a cargo is 0.5 to 0.8 mm, and the texture period corresponding to the extracted polarization feature is outside this range, the specification compliance coefficient will be significantly reduced, indicating that the feature may be abnormal. This provides a quantitative basis for subsequent feature correction and ensures that the extracted features are consistent with the attributes of the cargo itself.

[0037] Based on a multi-angle visible light image set and known light source locations, the physical consistency of the collected data is evaluated by verifying whether the continuity of shadows and highlights conforms to a preset optical reflection model. A physical consistency coefficient is obtained. The specification compliance coefficient and the physical consistency coefficient are used to weight and correct polarization features and visible light features, respectively. The preset optical reflection model is an extended model based on the Lambertian model, incorporating the geometric laws of shadow formation and the intensity distribution rules of highlight reflection. Shadow continuity is verified by detecting the gradient of shadow grayscale values ​​between adjacent pixels; if the gradient value is less than a preset threshold, it is considered continuous. Highlight continuity is verified by analyzing the shape and intensity distribution of highlight regions; if it conforms to the distribution rules predicted by the model, it is considered continuous. The physical consistency coefficient is calculated based on the degree of continuity between shadows and highlights; a coefficient of 1 indicates complete compliance, and a coefficient of 0 indicates severe non-compliance. Feature weighting correction uses a multiplication method: the polarization feature correction value equals the original polarization feature multiplied by the specification compliance coefficient, and the visible light feature correction value equals the original visible light feature multiplied by the physical consistency coefficient. It can effectively remove abnormal feature components caused by environmental interference or acquisition deviation. For example, the high light distortion caused by local strong light during acquisition will reduce the physical consistency coefficient. By weighted correction, the influence of the feature in this area can be weakened, and the reliability and stability of the feature can be improved.

[0038] The corrected polarization features are concatenated with visible light features to obtain a high-dimensional fused feature vector. A pre-trained autoencoder is then used to reduce the dimensionality of the high-dimensional fused feature vector to obtain a low-dimensional feature vector. Feature concatenation follows a fixed order: polarization features first, visible light features second. If the polarization features are 64-dimensional and the visible light features are 36-dimensional, then a 100-dimensional high-dimensional fused feature vector is formed after concatenation. The autoencoder consists of an input layer, an encoding layer, a decoding layer, and an output layer. The input layer dimension is the same as the high-dimensional fused feature vector dimension. The encoding layer progressively compresses the feature dimension through two convolutional layers. The first convolutional layer has 64 kernels and a size of 3x3, while the second convolutional layer has 32 kernels and a size of 3x3. Finally, the encoding layer outputs 32-dimensional low-dimensional features. The decoding layer attempts to recover the original high-dimensional features through two deconvolutional layers. The training process aims to minimize the reconstruction error between the input and output. The pre-training dataset contains multimodal feature samples of various types of goods. High-dimensional feature vectors suffer from data redundancy and high computational complexity. Autoencoder dimensionality reduction can significantly reduce data dimensionality while preserving core feature information. 32-dimensional low-dimensional features reduce the computational cost of subsequent storage and comparison, avoid the interference of redundant information on verification results, and improve system operating efficiency.

[0039] For each dimension of the low-dimensional feature vector, it is compared with the historical mean of that dimension. The comparison result is converted into a binary value to generate binary visual physical feature information. The historical mean of that dimension is the arithmetic mean of the feature values ​​of the corresponding dimension for all samples in the pre-training dataset, calculated offline and stored in the system. During the comparison, if the value of a certain dimension of the low-dimensional feature vector is greater than or equal to the corresponding historical mean, that dimension is converted to a binary value of 1; if it is less than the historical mean, it is converted to a binary value of 0. The 32-dimensional low-dimensional feature vector is converted to 32-bit binary visual physical feature information. The binary-processed feature information has the advantages of small storage space and fast comparison speed. Furthermore, binary features are not sensitive to small numerical fluctuations and are more stable. For example, slight changes in the value of a certain dimension of the low-dimensional feature within a reasonable range will not change its corresponding binary value, ensuring the consistency of feature information under different environments and laying the foundation for efficient and accurate authenticity comparison.

[0040] In one embodiment, the step of writing the visual physical feature information and the registration hash value together into the blockchain for storage to complete the registration and notarization of the cargo features includes: S31, Obtain the electronic warehouse receipt number, collection timestamp, and image acquisition device identifier based on the collection environment information; S32, which concatenates visual physical feature information, electronic warehouse receipt number, acquisition timestamp, and image acquisition device identifier; S33, apply a cryptographic hash algorithm to the concatenated data stream to generate a unique registration hash value; S34, the visual physical feature information and the registration hash value are encapsulated together into a transaction data, uploaded to a pre-built blockchain network for storage and consensus verification, and the immutable binding of the physical features of the goods and the electronic warehouse receipt is completed.

[0041] As described in steps S31-S34 above, by splicing the visual physical feature information of the goods with the core data related to the electronic warehouse receipt, encrypting and hashing the data, and then uploading the feature information and hash value together to the blockchain for storage, the physical attributes of the goods and the electronic warehouse receipt are bound together in an immutable manner, and the registration and storage of the goods features are completed, providing a reliable and traceable original feature benchmark for subsequent verification.

[0042] The core value of electronic warehouse receipts lies in their unique correspondence with physical goods. This correspondence needs to be solidified in an immutable way to resist fraudulent activities such as data tampering and goods substitution. At the same time, the verification process requires original features as a basis for comparison. Simply storing hash values ​​cannot meet the needs of direct comparison. Therefore, it is necessary to establish a credible association between feature information and warehouse receipt information and to fully preserve the evidence to solve the problems of unreliable data evidence and lack of effective comparison benchmarks in supply chain finance.

[0043] The system obtains the electronic warehouse receipt number, collection timestamp, and image acquisition device identifier based on the collected environment information. The electronic warehouse receipt number is obtained through the interface between the system and the warehouse management platform, corresponding one-to-one with the goods to be verified. The collection timestamp is generated synchronously with the system clock of the image acquisition device, accurate to the millisecond level, ensuring complete consistency with the time of multimodal image data acquisition. The image acquisition device identifier is a unique factory code for each acquisition device, pre-entered into the system and bound to the collection environment information for storage, and automatically retrieved during collection. This data is key information for associating goods characteristics with electronic warehouse receipts and tracing the collection process. The electronic warehouse receipt number establishes the correspondence between characteristics and the warehouse receipt, the collection timestamp ensures the time uniqueness of the data, and the device identifier facilitates tracing the operating status of the acquisition device, providing a foundation for the traceability of the subsequent verification process.

[0044] The visual physical feature information, electronic warehouse receipt number, acquisition timestamp, and image acquisition device identifier are concatenated in a fixed order: electronic warehouse receipt number, acquisition timestamp, image acquisition device identifier, and visual physical feature information. During concatenation, specific delimiters are used to distinguish different data segments. These delimiters are system-preset combinations of meaningless characters to avoid conflicts with the data content itself. This fixed-order concatenation method ensures that the same data combination generates only a unique data stream, preventing hash value generation errors due to disordered concatenation. Furthermore, the use of delimiters allows for reverse data splitting when needed later, ensuring data integrity and parsability, and providing a unified format of input data for subsequent cryptographic hashing processing.

[0045] A cryptographic hash algorithm is applied to the concatenated data stream to generate a unique registration hash value. The cryptographic hash algorithm uses SHA-256, which transforms an input data stream of arbitrary length into a fixed-length 256-bit hash value. During the calculation process, the data is processed multiple times using a non-linear compression function to ensure that even small changes in the input data will result in a significantly different hash value. For example, if a single bit in the visual physical feature information changes, the generated registration hash value will be completely different, and the original data stream cannot be deduced from the hash value. The registration hash value serves as a verification identifier for data integrity, verifying whether the stored feature information has been tampered with during subsequent verification, ensuring the authenticity of the data. Furthermore, its uniqueness further strengthens the unique association between cargo features and warehouse receipt information.

[0046] Visual physical feature information and registration hash value are jointly encapsulated into a transaction data entry, which is then uploaded to a pre-built blockchain network for storage and consensus verification. The transaction data encapsulation adopts a standard data format supported by the blockchain network, including a data header and a data body. The header records metadata such as the transaction initiator's identifier and timestamp, while the body stores the visual physical feature information and registration hash value. The pre-built blockchain network adopts a consortium blockchain architecture, with relevant parties such as warehousing providers, financial institutions, and regulatory agencies participating as nodes in the consensus process. The consensus mechanism employs a practical Byzantine fault-tolerant algorithm to ensure that transaction data is stored synchronously across multiple nodes and is tamper-proof. After the data is uploaded, each node verifies the transaction data using the consensus algorithm. Once verification is successful, the data is written into the blockchain ledger, completing the immutable binding of the goods' physical characteristics and the electronic warehouse receipt. The distributed storage characteristic of blockchain avoids the risk of data loss or tampering on a single node, and the consensus verification mechanism ensures the credibility of the data, making the stored visual physical feature information a reliable benchmark for subsequent verification, thus constructing an end-to-end trusted closed loop from physical goods to digital warehouse receipts.

[0047] In one embodiment, the step of collecting multimodal image data of the goods according to the same protocol and performing the hierarchical feature purification and fusion processing to generate visual physical feature information to be verified when verification is triggered includes: S41, receive a verification instruction for a specific electronic warehouse receipt, and according to the cargo identifier and acquisition protocol identifier carried in the instruction, control the image acquisition device to re-acquire the polarized light image sequence and multi-angle visible light image of the cargo according to the same parameter settings as in the registration stage, and simultaneously acquire the current acquisition environment information. S41, For the re-acquired multimodal image data, perform the same feature purification and fusion processing flow as in the registration stage to generate visual physical feature information to be verified; S41. Based on the re-acquired environmental information and multimodal image data, recalculate the physical consistency coefficient and positioning accuracy assessment value of the current acquisition, which will serve as the basis for evaluating the reliability of the acquired data in this verification.

[0048] As described in steps S41-S43 above, by strictly following the collection protocol and feature processing process of the registration stage during the verification stage, multimodal image data of goods and current environmental information are collected simultaneously, visual and physical feature information to be verified that is of the same origin as the registered features is generated, and the reliability of the collected data is re-evaluated, providing an accurate, comparable and traceable verification benchmark for subsequent dynamic comparison.

[0049] The accuracy of verification results depends on the comparability of the features to be verified and the registered features. However, differences in data collection parameters and processing procedures can lead to feature distortion. Furthermore, environmental changes during verification can affect data quality. Directly using features that have not undergone consistency verification for comparison will significantly reduce verification accuracy. Therefore, it is necessary to ensure consistency between data collection and processing, and to assess data reliability to address the issues of incomparable features and unreliable data quality during the verification phase. Existing technologies often use different data collection parameters or simplified processing procedures during verification compared to the registration phase, and do not assess the reliability of the collected data. This results in a lack of comparability between the features to be verified and the registered features, failing to accurately reflect the true state of the goods and leading to low credibility of the verification results.

[0050] The system receives verification instructions for specific electronic warehouse receipts. Based on the cargo identifier and data acquisition protocol identifier carried in the instructions, it controls the image acquisition device to re-acquire polarized light image sequences and multi-angle visible light images of the cargo according to the same parameter settings as during the registration phase. Simultaneously, it acquires current environmental information. The cargo identifier corresponds one-to-one with the electronic warehouse receipt number, and the data acquisition protocol identifier is associated with the data acquisition parameter configuration file from the registration phase and stored in the system database. Acquisition parameters include the 0° and 90° polarization angles of the polarization filter, the position coordinates and on / off sequence of the three controllable LED light sources, and the angle parameters of the front and side acquisition viewpoints. The image acquisition device automatically matches the parameters from the registration phase by reading the configuration file. Current environmental information is acquired through environmental sensors triggered synchronously with the image acquisition device, including ambient light intensity and temperature / humidity data, ensuring that the data timestamp is consistent with the image acquisition time. Maintaining complete consistency in the acquisition parameters can avoid feature deviations caused by differences in polarization angle, light source position, acquisition angle, etc. For example, if 0° and 90° polarization angles are used for acquisition during registration, and other angles are used during verification, the polarization characteristics will be completely different, making effective comparison impossible. Synchronously acquiring current environmental information provides data support for subsequent dynamic adjustment of the comparison strategy.

[0051] For the re-acquired multimodal image data, the same feature purification and fusion process as in the registration phase is executed to generate visual-physical feature information to be verified. This process includes polarized light image fusion, local binary pattern feature extraction, photometric stereo vision algorithm to recover the normal map, normal direction statistical histogram feature extraction, calculation of specification compliance coefficient and physical consistency coefficient, feature weighting correction, high-dimensional feature stitching, autoencoder dimensionality reduction, and binarization. The autoencoder uses the model parameters pre-trained in the registration phase, including a 3x3 convolutional kernel encoding layer and deconvolutional layer structure, and a 32-dimensional low-dimensional feature output setting. The same historical statistical mean is used as the comparison benchmark for binarization, as in the registration phase. This identical process ensures that the generation logic of the features to be verified is consistent with that of the registered features, and that the quantization dimension and representation method are completely matched. For example, if a 3x3 neighborhood window is used to extract local binary pattern features during registration, using a different window size during verification would cause a change in the feature vector dimension, making direct comparison impossible. Process reuse ensures feature comparability and provides a prerequisite for subsequent similarity calculations.

[0052] Based on the re-acquired environmental information and multimodal image data, the physical consistency coefficient and positioning accuracy evaluation value of the current acquisition are recalculated as the basis for assessing the reliability of the data acquired in this verification. The physical consistency coefficient is calculated by verifying whether the continuity of shadows and highlights in the re-acquired multi-angle visible light images conforms to the preset optical reflection model. The coefficient is 1 when it fully conforms and 0 when it does not conform significantly. The positioning accuracy evaluation value is obtained by solving the 3D point cloud using a multi-view stereo vision algorithm and calculating the reprojection error. An average deviation of less than 0.5 pixels is considered excellent, between 0.5 and 1 pixel is considered good, and greater than 1 pixel is considered poor. The environment during verification may differ from that during registration. For example, changes in light intensity may cause image shadow distortion, which will reduce the physical consistency coefficient. If the positioning accuracy evaluation value is poor, it indicates that the acquisition viewpoint positioning is inaccurate. These situations will affect the feature quality. Recalculating the evaluation coefficient can intuitively reflect the reliability of the data acquired in this case. If the evaluation result is not good, re-acquisition can be triggered in time to avoid misjudgment caused by comparing with low-quality data and improve the credibility of the verification results.

[0053] In one embodiment, the step of comparing the visual physical feature information to be verified with the registration information stored in the blockchain and outputting the authenticity verification result includes: S51, Based on the physical consistency coefficient and the positioning accuracy evaluation value, dynamically allocate the comparison weights of polarization features and visible light features, and obtain the polarization feature weights and visible light feature weights. The higher the physical consistency coefficient and the higher the positioning accuracy evaluation value, the greater the weight allocated to the polarization feature. At the same time, ensure that the sum of the polarization feature weights and the visible light feature weights is a fixed value. S52, calculate the similarity between the visual physical feature information to be verified and the registered visual physical feature information retrieved from the blockchain in the polarization feature segment and the visible light feature segment, respectively; S53, based on the polarization feature weight and the visible light feature weight, the similarity of the polarization feature sub-segment and the visible light feature sub-segment is weighted and summed to obtain the weighted comprehensive similarity; S54. Based on the preset risk level of this verification business and the current overall collection quality, the judgment threshold is dynamically adjusted. The higher the risk level or the lower the overall collection quality, the higher the judgment threshold will be. S55. Compare the weighted comprehensive similarity with the dynamically adjusted judgment threshold. If the weighted comprehensive similarity is greater than or equal to the judgment threshold, the goods are determined to be genuine and the verification is passed. If the weighted comprehensive similarity is less than the judgment threshold, the goods are determined to be fake and the verification is not passed. The key data and decision results of this verification process are then used to generate a new hash value and stored on the blockchain.

[0054] As described in steps S51-S55 above, by dynamically allocating feature comparison weights, accurately calculating feature similarity, and adaptively adjusting the judgment threshold, the visual and physical feature information to be verified is compared with the registration information stored on the blockchain, and a reliable verification result of the authenticity of the goods is output. At the same time, the verification process data is stored on the blockchain to complete the closed-loop verification from feature comparison to result traceability.

[0055] The accuracy of verification results depends on the adaptability of the comparison strategy to the data collection environment and business risks. Physical consistency and positioning accuracy during data collection directly determine feature reliability. Different business scenarios require varying degrees of verification rigor due to their different risk levels. Fixed weights and thresholds cannot cope with complex environmental changes and differentiated risk needs, easily leading to misjudgments or omissions. Therefore, it is necessary to achieve accurate comparison through dynamic strategy adjustment to solve the problems of poor adaptability and unreliable verification results in traditional comparison methods. Existing technologies use fixed feature weights and judgment thresholds for comparison, without considering the reliability differences of collected data or distinguishing business risk levels, resulting in low verification accuracy in complex environments and insufficient anti-counterfeiting capabilities in high-risk scenarios. By designing a complete process including dynamic weight allocation, multi-dimensional similarity calculation, weighted comprehensive evaluation, dynamic threshold adjustment, result output, and evidence storage, the system is specifically adapted to changes in the data collection environment and differences in business risks, significantly improving the accuracy and reliability of verification results.

[0056] The comparison weights of polarization features and visible light features are dynamically allocated based on the physical consistency coefficient and positioning accuracy assessment value, ensuring that the sum of their weights is a fixed value of 1. The physical consistency coefficient is obtained by verifying whether the continuity of shadows and highlights conforms to the preset optical reflection model, and the positioning accuracy assessment value is obtained by calculating the reprojection error of the 3D point cloud. Both are divided into three levels: excellent, good, and poor. The weight allocation adopts a linear mapping rule: when both the physical consistency coefficient and the positioning accuracy assessment value are excellent, the weight allocation for polarization features is 0.6, and the weight allocation for visible light features is 0.4; when both are good, the weight allocation for polarization features is 0.5, and the weight allocation for visible light features is 0.5; when both are poor, the weight allocation for polarization features is 0.4, and the weight allocation for visible light features is 0.6. Polarization features reflect the inherent properties of the cargo material, and their reliability is more affected by the quality of data collection. The higher the physical consistency and positioning accuracy, the more reliable the polarization features are, and assigning them higher weights can strengthen the role of core anti-counterfeiting features. Conversely, the weight of visible light features is increased to compensate for the deficiencies of polarization features, ensuring that the weight allocation matches the reliability of the features and improving the targeting of the comparison.

[0057] The similarity between the visual physical feature information to be verified and the registered visual physical feature information retrieved from the blockchain is calculated separately in the polarization feature segment and the visible light feature segment. The polarization feature segment is the binary fragment corresponding to the local binary mode feature vector, and the visible light feature segment is the binary fragment corresponding to the normal direction statistical histogram feature vector. The similarity calculation uses the Hamming distance algorithm, counting the number of different corresponding binary bits in the two feature segments, and converting the degree of difference into a similarity value between 0 and 1 using the formula "Similarity = 1 - (Hamming distance / Feature segment length)". For example, if the polarization feature segment length is 64 bits and the Hamming distance between the feature to be verified and the registered feature is 3, then the similarity of the polarization feature segment is 1 - 3 / 64 ≈ 0.953; if the visible light feature segment length is 36 bits and the Hamming distance is 2, then the similarity of the visible light feature segment is 1 - 2 / 36 ≈ 0.944. Calculating similarity in multiple dimensions can accurately capture the matching degree of the two types of features, providing a quantitative basis for subsequent weighted comprehensive evaluation and avoiding information omissions caused by single-dimensional feature comparison.

[0058] Based on the weights of polarization features and visible light features, a weighted comprehensive similarity is obtained by weighted summation of the similarities of the two types of feature segments. The weighted summation formula is: "Weighted Comprehensive Similarity = Polarization Feature Weight × Polarization Feature Segment Similarity + Visible Light Feature Weight × Visible Light Feature Segment Similarity". For example, if the polarization feature weight is 0.6 and the similarity is 0.953, and the visible light feature weight is 0.4 and the similarity is 0.944, then the weighted comprehensive similarity = 0.6 × 0.953 + 0.4 × 0.944 ≈ 0.950. By assigning differentiated influence weights to different feature segments, the comprehensive similarity reflects both the matching of core anti-counterfeiting features and the supplementary role of auxiliary features, comprehensively reflecting the feature fit between the goods to be verified and the registered goods, providing a precise quantitative indicator for authenticity determination.

[0059] Based on the pre-set risk level of this verification business and the current overall data collection quality, the judgment threshold is dynamically adjusted. The business risk level is pre-set by the user and is divided into three levels: high, medium, and low. High risk corresponds to scenarios such as large-scale financing, medium risk corresponds to regular transaction scenarios, and low risk corresponds to internal inventory scenarios. The overall data collection quality is determined by a combination of the physical consistency coefficient and the positioning accuracy assessment value. If both are excellent, the data collection quality is excellent; if both are good, the data collection quality is good; if either is poor, the data collection quality is poor. The judgment threshold adjustment rules are as follows: for low risk and excellent data collection quality, the threshold is set to 0.85; for medium risk or good data collection quality, the threshold is set to 0.90; for high risk or poor data collection quality, the threshold is set to 0.95. High-risk scenarios require stricter judgment standards to prevent counterfeit goods from passing through verification and causing financial risks. Increasing the threshold when the data collection quality is poor can reduce the probability of misjudgment due to feature distortion, ensuring that the judgment threshold is adapted to business needs and data quality.

[0060] The verification process compares the weighted overall similarity with a dynamically adjusted threshold, outputs the authenticity verification result, and stores key data on the blockchain. If the weighted overall similarity is greater than or equal to the threshold, the goods are deemed genuine, and the verification passes. If it is less than the threshold, the goods are deemed fake, and the verification fails. Key data in the verification process includes current collection environment information, feature comparison weights, similarity of each segment, weighted overall similarity, threshold, and decision result. These data are concatenated and then the SHA-256 encryption hash algorithm is applied to generate a new hash value, which is then packaged as transaction data and uploaded to a pre-built blockchain network for storage. For example, if the weighted overall similarity is 0.950 and the threshold is 0.95, the verification passes, and the relevant data is hashed and uploaded to the blockchain. If the weighted overall similarity is 0.940 and the threshold is 0.95, the verification fails, but the data is still uploaded to the blockchain. By comparing specific thresholds, the authenticity of the verification results can be accurately determined. Blockchain evidence storage ensures that the verification process is traceable and tamper-proof, further enhancing the credibility of the verification results and building a complete and trustworthy verification loop.

[0061] like Figure 2 As shown, this invention also discloses an electronic warehouse receipt goods authenticity verification system based on AI image recognition, including: Acquisition module 1 is used to acquire multimodal image data of the goods to be verified and information on the acquisition environment. The multimodal image data includes polarized light image sequences and multi-angle visible light images. Generation module 2 is used to perform hierarchical feature purification and fusion processing on the multimodal image data to generate binary visual physical feature information representing the unique physical attributes of the goods. Registration module 3 is used to bind the visual physical feature information with the corresponding electronic warehouse receipt information, generate a registration hash value, and write the visual physical feature information and the registration hash value together into the blockchain for storage, thereby completing the registration and evidence storage of the cargo features. The fusion module 4 is used to collect multimodal image data of the goods according to the same protocol when the verification is triggered, and to perform the hierarchical feature purification and fusion processing to generate visual physical feature information to be verified. Output module 5 is used to dynamically calculate feature weights and decision thresholds based on real-time collected environmental information, compare the visual physical feature information to be verified with the registration information stored in the blockchain, and output the authenticity verification result.

[0062] In one embodiment, the registration module includes: The acquisition unit is used to acquire the electronic warehouse receipt number, acquisition timestamp, and image acquisition device identifier based on the acquisition environment information. The splicing unit is used to splice visual physical feature information, electronic warehouse receipt number, acquisition timestamp, and image acquisition device identifier; The generation unit is used to apply a cryptographic hash algorithm to the concatenated data stream to generate a unique registration hash value. The binding unit is used to encapsulate the visual physical feature information and the registration hash value into a transaction data, upload it to the pre-built blockchain network for storage and consensus verification, and complete the tamper-proof binding of the physical features of the goods with the electronic warehouse receipt.

[0063] This application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method for verifying the authenticity of goods in electronic warehouse receipts based on AI image recognition.

[0064] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method for verifying the authenticity of goods in electronic warehouse receipts based on AI image recognition.

[0065] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0066] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0067] The above description is merely a preferred embodiment of the present invention and does not limit the scope of this application. Any equivalent results or equivalent process transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of protection of this application.

Claims

1. An AI image recognition-based electronic warehouse receipt cargo authenticity verification method, characterized in that, Includes the following steps: Acquire multimodal image data of the goods to be verified and information on the acquisition environment. The multimodal image data includes polarized light image sequences and multi-angle visible light images. The multimodal image data is subjected to hierarchical feature purification and fusion processing to generate binary visual physical feature information representing the unique physical attributes of the goods; The visual physical feature information is bound to the corresponding electronic warehouse receipt information to generate a registration hash value. The visual physical feature information and the registration hash value are then written into the blockchain for storage, thus completing the registration and evidence storage of the cargo features. When verification is triggered, multimodal image data of the goods are collected according to the same protocol, and the hierarchical feature purification and fusion processing is performed to generate visual physical feature information to be verified. Based on the real-time collected environmental information, the feature weights and decision thresholds are dynamically calculated. The visual physical feature information to be verified is compared with the registration information stored in the blockchain, and the authenticity verification result is output. 2.The AI image recognition-based electronic warehouse warrant cargo authenticity verification method of claim 1, wherein, The steps for acquiring multimodal image data of the goods to be verified and collecting environmental information include: By using an image acquisition device integrated with a polarizing filter, surface images of the goods are acquired at at least two different polarization angles to form a polarized light image sequence. Under the illumination of multiple controllable light sources with known spatial relationships, visible light images of the cargo are acquired from at least two different perspectives to form a multi-angle visible light image set, and the on / off state and position coordinates of each light source are recorded simultaneously. The ambient light intensity, ambient temperature and humidity data are acquired synchronously at the time of data collection using environmental sensors. Based on the multi-angle visible light image set, the three-dimensional point cloud of the cargo surface is calculated using a multi-view stereo vision algorithm, and the overall positioning accuracy of this acquisition is evaluated based on the reprojection error of the three-dimensional point cloud, generating a positioning accuracy evaluation value. 3.The AI image recognition-based electronic warehouse warrant cargo authenticity verification method according to claim 2, characterized in that, The step of performing hierarchical feature extraction and fusion processing on the multimodal image data to generate binary visual physical feature information representing the unique physical attributes of the goods includes: The polarized light image sequence is fused to calculate the degree of polarization image and the angle of polarization image, and the local binary mode feature vector is extracted from the degree of polarization image as the polarization feature. For the multi-angle visible light image set, the normal map of the cargo surface is recovered using a photometric stereo vision algorithm, and the statistical histogram feature vector of the normal direction is calculated as a visible light feature. Based on the cargo specification database, the standard texture period and surface roughness range are obtained, the degree of conformity between the polarization characteristics and the standard specifications is calculated, and the specification conformity coefficient is obtained. Based on the multi-angle visible light image set and the known light source position, the physical consistency of the collected data is evaluated by verifying whether the continuity of shadows and highlights conforms to the preset optical reflection model, and the physical consistency coefficient is obtained. The polarization characteristics and visible light characteristics are then weighted and corrected using the specification compliance coefficient and the physical consistency coefficient, respectively. The corrected polarization features are concatenated with the visible light features to obtain a high-dimensional fused feature vector. A pre-trained autoencoder is then used to reduce the dimensionality of the high-dimensional fused feature vector to obtain a low-dimensional feature vector. For each dimension of the low-dimensional feature vector, it is compared with the historical statistical mean of that dimension. Based on the comparison result, it is converted into a binary value to generate binary visual physical feature information. 4.The AI image recognition-based electronic warehouse warrant cargo authenticity verification method of claim 3, wherein, The step of writing the visual physical feature information and the registration hash value together into the blockchain for storage, thereby completing the registration and notarization of the cargo features, includes: Based on the collected environment information, obtain the electronic warehouse receipt number, the collection timestamp, and the image acquisition device identifier; The visual physical feature information, electronic warehouse receipt number, collection timestamp, and image acquisition device identifier are concatenated together; Apply a cryptographic hash algorithm to the concatenated data stream to generate a unique registration hash value; The visual physical feature information and the registration hash value are encapsulated together into a transaction data, uploaded to a pre-built blockchain network for storage and consensus verification, and the immutable binding of the physical features of the goods with the electronic warehouse receipt is completed. 5.The AI image recognition-based electronic warehouse warrant cargo authenticity verification method according to claim 4, characterized in that, The steps of collecting multimodal image data of the goods according to the same protocol when verification is triggered, and performing the hierarchical feature purification and fusion processing to generate visual physical feature information to be verified include: Receive verification instructions for a specific electronic warehouse receipt, and according to the cargo identifier and collection protocol identifier carried in the instructions, control the image acquisition device to re-acquire the polarized light image sequence and multi-angle visible light image of the cargo according to the same parameter settings as in the registration stage, and simultaneously acquire the current acquisition environment information. For the re-acquired multimodal image data, perform the same feature purification and fusion processing flow as in the registration phase to generate visual physical feature information to be verified; Based on the re-acquired environmental information and multimodal image data, the physical consistency coefficient and positioning accuracy assessment value of the current acquisition are recalculated as the basis for evaluating the reliability of the acquired data in this verification. 6.The AI image recognition-based electronic warehouse warrant cargo authenticity verification method according to claim 5, characterized in that, The step of comparing the visual physical feature information to be verified with the registration information stored in the blockchain and outputting the authenticity verification result includes: Based on the physical consistency coefficient and the positioning accuracy evaluation value, the comparison weights of polarization features and visible light features are dynamically allocated, and the weights of polarization features and visible light features are obtained. The higher the physical consistency coefficient and the higher the positioning accuracy evaluation value, the greater the weight allocated to the polarization features. At the same time, it is ensured that the sum of the weights of polarization features and visible light features is a fixed value. Calculate the similarity between the visual physical feature information to be verified and the registered visual physical feature information retrieved from the blockchain in the polarization feature segment and the visible light feature segment, respectively. Based on the polarization feature weights and visible light feature weights, the similarities of the polarization feature sub-segments and visible light feature sub-segments are weighted and summed to obtain a weighted comprehensive similarity. The judgment threshold will be dynamically adjusted based on the preset risk level of this verification business and the current overall collection quality. The higher the risk level or the lower the overall collection quality, the higher the judgment threshold will be. The weighted comprehensive similarity is compared with the dynamically adjusted judgment threshold. If the weighted comprehensive similarity is greater than or equal to the judgment threshold, the goods are judged to be genuine and the verification is passed. If the weighted comprehensive similarity is less than the judgment threshold, the goods are judged to be fake and the verification is failed. The key data and decision results of this verification process are then used to generate a new hash value and stored on the blockchain.

7. An AI image recognition-based electronic warehouse receipt cargo authenticity verification system, characterized in that, include: The acquisition module is used to acquire multimodal image data of the goods to be verified and information on the acquisition environment. The multimodal image data includes polarized light image sequences and multi-angle visible light images. The generation module is used to perform hierarchical feature purification and fusion processing on the multimodal image data to generate binary visual physical feature information representing the unique physical attributes of the goods. The registration module is used to bind the visual physical feature information with the corresponding electronic warehouse receipt information, generate a registration hash value, and write the visual physical feature information and the registration hash value together into the blockchain for storage, thereby completing the registration and evidence storage of the cargo features. The fusion module is used to collect multimodal image data of the goods according to the same protocol when verification is triggered, and to perform the hierarchical feature purification and fusion processing to generate visual physical feature information to be verified. The output module is used to dynamically calculate feature weights and decision thresholds based on real-time collected environmental information, compare the visual physical feature information to be verified with the registration information stored in the blockchain, and output the authenticity verification result. 8.The AI image recognition-based electronic warehouse warrant cargo authenticity verification system according to claim 7, characterized in that, The registration module includes: The acquisition unit is used to acquire the electronic warehouse receipt number, acquisition timestamp, and image acquisition device identifier based on the acquisition environment information. The splicing unit is used to splice visual physical feature information, electronic warehouse receipt number, acquisition timestamp, and image acquisition device identifier; The generation unit is used to apply a cryptographic hash algorithm to the concatenated data stream to generate a unique registration hash value. The binding unit is used to encapsulate the visual physical feature information and the registration hash value into a transaction data, upload it to the pre-built blockchain network for storage and consensus verification, and complete the tamper-proof binding of the physical features of the goods with the electronic warehouse receipt. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images