A jujube variety identification method and system based on artificial intelligence

Abstract

The application provides a jujube variety discrimination method and system based on artificial intelligence. It belongs to the field of cross technology of agricultural artificial intelligence and computer vision. The method comprises the following steps: synchronously collecting multi-modal original data of jujube samples, and constructing a multi-modal original data set; pre-processing the multi-modal original data set to obtain standardized multi-modal feature data; constructing a hyper-dimensional feature space based on the standardized multi-modal feature data, and generating a jujube hyper-dimensional feature vector set through feature cross fusion. By synchronously collecting macroscopic visual images, hyperspectral data, three-dimensional morphological scanning data and microscopic texture images, various characteristic information of the jujube samples can be comprehensively and accurately captured, a hyper-dimensional feature space is constructed, and deep fusion of cross-modal features is realized, so that the precision and robustness of variety discrimination are improved.

Description

TECHNICAL FIELD

[0001] The application provides a jujube variety discrimination method and system based on artificial intelligence, and belongs to the technical field of agricultural artificial intelligence and computer vision. BACKGROUND

[0002] As an important economic crop with both edible and medicinal values, jujubes are widely planted in a variety of varieties around the world. Traditional jujube variety discrimination relies on manual experience and is performed by observing the appearance and taste of jujubes, which is inefficient and the discrimination result is easily affected by subjective factors, and the accuracy is difficult to guarantee. With the large-scale development of the jujube industry, there is an increasing demand for accurate and efficient variety discrimination methods.

[0003] In recent years, multi-modal data acquisition technology has been continuously developed, such as high-definition cameras, hyperspectral instruments, three-dimensional scanners, and microscopic imaging systems, which can obtain rich information of jujube samples from macro to micro, from visual to spectrum, and other dimensions. Artificial intelligence has made significant progress in feature fusion and classification model construction, providing a new technical path for jujube variety discrimination. However, there is currently no mature system that can fully integrate the advantages of multi-modal data and use artificial intelligence technology to achieve high-precision and intelligent jujube variety discrimination. SUMMARY

[0004] The application provides a jujube variety discrimination method and system based on artificial intelligence to solve the problems mentioned in the background technology.

[0005] The application provides a jujube variety discrimination method based on artificial intelligence, which comprises the following steps:

[0006] S1: Synchronously collecting multi-modal original data of jujube samples to construct a multi-modal original data set; preprocessing the multi-modal original data set to obtain standardized multi-modal feature data; constructing a hyper-dimensional feature space based on the standardized multi-modal feature data, and generating a jujube hyper-dimensional feature vector set through feature cross fusion;

[0007] S2: Inputting the hyper-dimensional feature vector set into a dynamic gate fusion network, adaptively adjusting the contribution degree of each modal feature through a dynamic weight distribution module, strengthening the correlation of key features combined with a cross-modal attention mechanism, and outputting a robust fusion feature vector; performing initial variety classification based on the robust fusion feature vector to obtain a preliminary variety discrimination result and a corresponding confidence score;

[0008] S3: Introducing a jujube evolutionary pedigree prior knowledge base, performing semantic alignment of the preliminary variety discrimination result with the jujube evolutionary pedigree prior knowledge base, correcting low-confidence classification results through knowledge graph reasoning, and generating a variety feature attribution graph based on the corrected discrimination result to clearly show the contribution path of each modal feature to the classification decision.

[0009] S4: Constructing a blockchain trusted traceability system, storing the life data of the jujube sample on the chain, generating a unique digital fingerprint through a smart contract; based on the distributed verification mechanism of the blockchain node, cross-validation is performed on the cross-regional jujube variety discrimination results;

[0010] S5: Integrating the fusion data of the dynamic gate fusion network to generate a structured jujube variety discrimination report.

[0011] The jujube variety discrimination system based on artificial intelligence comprises:

[0012] One or more processors;

[0013] A memory for storing one or more programs,

[0014] When the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of the above.

[0015] The present application has the following advantages:

[0016] 1. By synchronously collecting macroscopic visual images, hyperspectral data, three-dimensional morphological scanning data and microscopic texture images, various characteristic information of jujube samples can be comprehensively and accurately captured, an ultra-dimensional feature space can be constructed, and deep fusion of cross-modal features can be realized, thereby improving the accuracy and robustness of variety discrimination.

[0017] 2. The network can effectively strengthen the relevance of key features by adaptively adjusting the contribution of each modal feature and combining a cross-modal attention mechanism, optimize the use of each modal data by the model, and further improve the accuracy of variety classification and the processing ability of low confidence classification results.

[0018] 3. By introducing a jujube evolutionary pedigree prior knowledge base and combining knowledge graph reasoning technology, low confidence classification results can be semantically aligned and corrected. This process not only improves the reliability of the discrimination results, but also provides more transparent explainability for the discrimination process.

[0019] 4. By constructing a blockchain traceability system, the collection information, feature data and discrimination results of jujube samples are stored in a trusted manner, ensuring the non-tamperability and traceability of the data. A unique digital fingerprint is generated by a smart contract, and the distributed verification mechanism of the blockchain node is used to cross-validate the cross-regional discrimination results, further improving the data credibility and system security.

[0020] 5. The method can generate a structured jujube variety discrimination report, including variety name, characteristic attribution analysis, discrimination confidence, evolutionary lineage correlation, and blockchain storage certificate, helping relevant personnel understand the discrimination results more intuitively, and providing comprehensive support for the management, protection and research of jujube varieties. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 The method steps of the present application are described. DETAILED DESCRIPTION

[0022] The preferred embodiments of the present application are described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application.

[0023] One embodiment of the present application, as shown in Figure 1 A jujube variety discrimination method based on artificial intelligence, the method comprising:

[0024] S1: synchronously collecting multi-modal original data of jujube samples, the multi-modal original data comprising macroscopic visual images, hyperspectral data, three-dimensional morphological scanning data and microscopic texture images, and constructing a multi-modal original data set; pre-processing the multi-modal original data set to obtain standardized multi-modal feature data; constructing a hyper-dimensional feature space based on the standardized multi-modal feature data, and generating a jujube hyper-dimensional feature vector set containing space-spectrum-morphology-texture joint representation through feature cross fusion;

[0025] S2: inputting the hyper-dimensional feature vector set into a dynamic gate fusion network, adaptively adjusting the contribution degree of each modal feature through a dynamic weight distribution module, and combining a cross-modal attention mechanism to strengthen the correlation of key features, and outputting a robust fusion feature vector; performing initial variety classification based on the robust fusion feature vector to obtain a preliminary variety discrimination result and a corresponding confidence score;

[0026] S3: introducing a prior knowledge base of Ziziphus evolutionary lineage, the prior knowledge base comprising a jujube species evolution tree, genetic marker data and ecological distribution information; performing semantic alignment of the preliminary variety discrimination result and the prior knowledge base of Ziziphus evolutionary lineage, and correcting low-confidence classification results through knowledge graph reasoning; generating a variety characteristic attribution graph based on the corrected discrimination result, and explicitly defining the contribution path of each modal feature to the classification decision, so as to realize visual explanation of the discrimination process;

[0027] S4: constructing a blockchain trusted traceability system, and storing life data of jujube samples on the chain, the life data comprising collection time, geographical location, multi-modal original data, hyper-dimensional feature vector and final variety discrimination result; generating a unique digital fingerprint through a smart contract; based on a distributed verification mechanism of a blockchain node, cross- verifying the discrimination results of jujube varieties across regions;

[0028] S5: integrate the fusion data of the dynamic gate fusion network, the fusion data including robust discrimination results, enhanced explanation data of the Ziziphus evolutionary pedigree correction, and blockchain traceability verification information, and generate a structured Ziziphus variety discrimination report; the discrimination report content includes variety name, biological characteristic attribution analysis, discrimination confidence, evolutionary pedigree correlation description, and blockchain storage certificate.

[0029] The working principle and effects of the above technical solution are as follows: by synchronously collecting macroscopic, spectral, three-dimensional, and microscopic multidimensional data and constructing a hyperdimensional feature space, the problem of one-sidedness of single modal data information is avoided, the basis for variety discrimination is more comprehensive, and the utilization rate of multi-modal data is improved; the dynamic gate fusion network combined with the cross-modal attention mechanism can adaptively adjust the contribution degree of each modal feature, reduce irrelevant feature interference, make the robust fusion feature vector more suitable for variety discrimination requirements, and enhance the pertinence of feature fusion; the Ziziphus evolutionary pedigree prior knowledge base is introduced and low-confidence results are corrected through knowledge graph reasoning, the misjudgment rate caused by data noise or similar features is reduced, and the correction efficiency of the preliminary discrimination result is improved; by generating a variety characteristic attribution graph to visualize the contribution path of each modal feature, the decision-making doubts caused by the traditional AI discrimination black box problem are reduced, and the explainability of the discrimination process is enhanced; the blockchain system realizes life data storage and distributed verification, avoids data tampering risks, reduces the inconsistency of cross-regional discrimination results, and improves the credibility of the discrimination result; the structured report is generated by integrating the core data of multiple links, the biological characteristics of the variety, the evolutionary correlation, and the storage information are clear, the information query cost in subsequent tracing or application is reduced, and the practical value of the discrimination report is enhanced.

[0030] In an embodiment of the present application, the S1 comprises:

[0031] S11, screening Ziziphus samples, the samples covering Ziziphus individuals of different varieties, growth periods, and ecological environments, eliminating invalid samples damaged by pests and diseases and with incomplete morphology, and determining an effective sample set;

[0032] S12, for the effective sample set, multi-device synchronous collection of multi-modal original data is adopted, including: collecting macroscopic visual images by a high-definition camera, collecting hyperspectral data by a hyperspectral instrument, collecting three-dimensional morphological scanning data by a three-dimensional scanner, and collecting microscopic texture images by a microscopic imaging system, and integrating all the data to construct a multi-modal original data set;

[0033] S13, preprocessing the multi-modal original data set, the preprocessing including denoising and white balance correction of the macroscopic visual images, spectral baseline correction and redundant band elimination of the hyperspectral data, point cloud denoising and registration of the three-dimensional morphological scanning data, and gray scale normalization of the microscopic texture images, eliminating data interference terms, and obtaining standardized multi-modal feature data;

[0034] S14, based on the standardized multi-modal feature data, establishing a hyper-dimensional feature space coordinate system, mapping the macro visual features, hyperspectral features, three-dimensional morphological features and microscopic texture features to different dimensional axes respectively, and constructing a hyper-dimensional feature space containing multi-modal dimensions;

[0035] S15, performing feature cross fusion in the hyper-dimensional feature space, mining the correlation between the modal features (for example, the correspondence between the spectral features and the microscopic texture features, and the consistency between the three-dimensional morphology and the macro visual features) through the attention mechanism, and generating a jujube hyper-dimensional feature vector set containing space-spectrum-morphology-texture joint representation.

[0036] The working principle and effect of the above technical solution are as follows: by screening jujube samples of different varieties, growth cycles and ecological environments and eliminating invalid samples, the interference of incomplete and damaged samples on subsequent data processing is reduced, more reliable basic data is provided for subsequent discrimination, and the stability of sample quality is improved; multi-device synchronous acquisition of multi-modal data is adopted to cover the feature dimensions from macro to micro and from visual to spectrum, avoiding the problem of one-sidedness of single data type information, making the data representation more in line with the real characteristics of jujube varieties, and enhancing the comprehensiveness of the original data; by targeted preprocessing to eliminate image noise, spectral redundancy and other interference terms, the interference of invalid information on feature analysis is reduced, the purity of the standardized multi-modal feature data is improved, and the influence of data interference on feature extraction is reduced; the establishment of a hyper-dimensional feature space maps different modal features to independent dimensional axes, avoiding the dimensional confusion of different types of features, making the uniqueness of each modal feature more easily embodied, and improving the discrimination of multi-modal features; the corresponding relationship between cross-modal features is mined through the attention mechanism, the blindness of feature fusion is reduced, the generated hyper-dimensional feature vector set can more accurately reflect the joint features of the varieties, more representative feature support is provided for subsequent discrimination, and the effectiveness of feature correlation is enhanced.

[0037] In an embodiment of the present application, the S14 comprises:

[0038] Extracting core feature parameters of each modality in the standardized multi-modal feature data, wherein the macro visual feature parameters include color channel mean and texture entropy value, the hyperspectral feature parameters include feature band reflectivity and spectral peak value, the three-dimensional morphological feature parameters include volume, surface area and curvature distribution, and the microscopic texture feature parameters include texture granularity and porosity, forming a modal core feature parameter set;

[0039] According to the physical meaning and data magnitude of each modal feature parameter, the modal core feature parameter set is normalized to uniformly map parameters of different dimensions to the [0, 1] interval, and a normalized feature parameter set is generated;

[0040] Based on the normalized feature parameter set, a dimension axis system of the hyper-dimensional feature space is constructed, macroscopic visual feature parameters are mapped to the X-axis cluster, hyperspectral feature parameters are mapped to the Y-axis cluster, three-dimensional morphological feature parameters are mapped to the Z-axis cluster, and microscopic texture feature parameters are mapped to the W-axis cluster, forming a four-dimensional basic coordinate system framework.

[0041] In the four-dimensional basic coordinate system framework, independent dimension sub-axes are allocated to each feature parameter, and the sub-axis included angle is adjusted through parameter correlation calculation, so that the parameter sub-axes with high correlation degree are distributed at acute angles, and the parameter sub-axes with low correlation degree are distributed at obtuse angles, thereby optimizing the coordinate system structure.

[0042] All normalized feature parameters are positioned in space according to the corresponding dimension sub-axes to form a hyper-dimensional feature space containing macroscopic visual, hyperspectral, three-dimensional morphological, and microscopic texture multi-modal dimensions.

[0043] The working principle and effects of the above technical solution are as follows: by extracting core feature parameters of each modality, the interference of redundant parameters on subsequent analysis is avoided, feature data is more focused on key information for variety identification, and the pertinence of feature parameters is improved; normalization processing uniformly maps parameters to the [0, 1] interval, reduces the imbalance of feature weights caused by data magnitude differences, makes each parameter more comparable in analysis, and reduces the conflict influence of different dimensions; constructing a four-dimensional basic coordinate system framework divides different modal parameters into independent axis clusters, avoids confusion of modal features, makes the unique properties of each modality more easily identified, and enhances the dimension distinction of multi-modal features; by adjusting the sub-axis included angle through parameter correlation, the parameter sub-axes with high correlation degree are closer, the interference of irrelevant parameters in the feature space is reduced, the feature distribution is more consistent with the actual correlation logic, and the rationality of the coordinate system structure is improved; all normalized parameters are accurately positioned to the corresponding sub-axes, feature omission is avoided, the multi-modal dimensional space formed can more comprehensively carry the feature information of jujube varieties, lays a solid foundation for subsequent fusion, and enhances the integrity of the hyper-dimensional feature space.

[0044] In an embodiment of the present application, the S15 comprises:

[0045] S151, extracting key anchor points of each modal feature in the hyper-dimensional feature space, wherein the macroscopic visual feature anchor point is the color mutation region center coordinate, the hyperspectral feature anchor point is the characteristic band reflectivity peak point, the three-dimensional morphological feature anchor point is the curvature extreme point, and the microscopic texture feature anchor point is the texture structure turning point, forming a multi-modal feature anchor point set;

[0046] S152, calculating the spatial correlation degree between the feature anchor point set, measuring the spatial proximity of different modal anchor points through Euclidean distance, evaluating the matching degree of feature attributes through cosine similarity, and generating an anchor point correlation degree matrix;

[0047] S153, input the anchor point correlation degree matrix into the attention mechanism model, assign high weight values to high correlation anchor point pairs, strengthen the internal connection between cross-modal features (for example, the correlation between the three-dimensional shape curvature anchor point and the macroscopic visual color anchor point), and generate a weighted feature correlation graph;

[0048] S154, based on the weighted feature correlation graph, cross-map and fuse the multi-modal features in the hyper-dimensional feature space, vector superposition is performed on the feature dimensions with high correlation degree, the feature dimensions with low correlation degree are kept independent representation, and a fusion feature vector preliminary draft is formed;

[0049] S155, dimension normalization processing is performed on the fusion feature vector preliminary draft, the fusion deviation caused by the order of magnitude difference of different modal features is eliminated, and finally a jujube hyper-dimensional feature vector set containing space-spectrum-shape-texture joint representation is generated.

[0050] The working principle and effect of the above technical solution are as follows: by extracting key anchor points of each mode (such as color mutation center, spectral reflectance peak), invalid analysis of irrelevant feature regions is avoided, and subsequent correlation calculation is more focused on core features of variety discrimination, improving feature focusing; the correlation degree of anchor points is calculated by using Euclidean distance and cosine similarity in two dimensions, reducing the correlation misjudgment caused by a single measurement standard, and the generated correlation degree matrix is more consistent with the real connection between features, enhancing the accuracy of cross-modal correlation; the attention mechanism assigns high weights to high correlation anchor point pairs, strengthens the connection between key features, and weakens irrelevant features, avoiding confusion of low correlation features, making the feature correlation graph more targeted, and reducing the interference of irrelevant features; according to the high and low correlation degrees, vector superposition and independent representation are performed respectively, the cooperative value of high correlation features is retained, the unique information of low correlation features is not lost, the feature loss in the fusion process is reduced, and the rationality of feature fusion is improved; finally, the dimension normalization eliminates the order of magnitude deviation, avoids the problem that a certain type of modal feature covers other features due to large numerical value, makes the generated joint representation vector more balancedly reflect the multi-dimensional features of jujube varieties, and enhances the reliability of hyper-dimensional feature vectors.

[0051] In an embodiment of the present application, the S154 comprises:

[0052] Analyzing the weighted feature correlation graph, extracting high correlation feature dimension pairs in the graph with a correlation degree threshold (for example, the preset correlation degree is greater than or equal to 0.8), and screening low correlation feature dimensions with a correlation degree less than 0.8 to generate a high / low correlation feature dimension classification table;

[0053] For high correlation feature dimension pairs, the feature vector data corresponding to each dimension is obtained, vector dot product operation is used to realize feature superposition between dimensions, the core feature information of the superimposed vector (for example, the superimposed feature of spectral reflectance and microscopic texture porosity) is retained, and a high correlation fusion feature vector group is generated;

[0054] For low correlation feature dimensions, the feature vectors of each dimension are extracted separately, and the modal type to which each feature vector belongs is marked by a feature identifier (for example, three-dimensional shape-volume feature vector, macroscopic visual-color feature vector), to ensure that the independently represented features can be traced back, and a low correlation independent feature vector group is generated;

[0055] The high correlation fusion feature vector group and the low correlation independent feature vector group are spliced according to the dimension order of the hyper-dimensional feature space to form a combined vector sequence containing fusion features and independent features;

[0056] The combined vector sequence is subjected to integrity verification to confirm that there is no missing or repeated feature dimension, and finally a fusion feature vector preliminary draft is formed.

[0057] The working principle and effects of the above technical solution are as follows: By setting the correlation threshold to distinguish high / low correlation feature dimensions and generating a classification table, the confusion of mixing different correlation degree features together is avoided, the subsequent fusion direction is more clear, and the clarity of feature classification is improved; the high correlation dimension features are superimposed by vector dot product operation, which not only retains the core information but also strengthens the linkage effect between features, reduces the information waste when high correlation features are processed alone, and enhances the collaborative value of high correlation features; low correlation features are extracted and labeled with modal types, which avoids being hidden or misfused by high correlation features, so that the unique value of each low correlation feature can be retained, and the loss risk of low correlation features is reduced; the two types of vector groups are spliced according to the dimension order of the hyper-dimensional feature space, which avoids the subsequent processing troubles caused by vector arrangement confusion, makes the combined vector sequence structure clearer, and improves the regularity of the vector sequence; the integrity verification confirms that there is no missing or repeated feature, which avoids the influence of feature missing or redundancy on subsequent normalization processing, makes the reliability of the fusion feature vector preliminary draft higher, and reduces the error risk of the fusion preliminary draft.

[0058] In an embodiment of the present application, the S2 comprises:

[0059] S21, modal feature alignment processing is performed on the jujube hyper-dimensional feature vector set, the dimensions and data formats of each modal feature are unified, fusion deviation caused by modal differences is avoided, and an aligned hyper-dimensional feature vector set is generated;

[0060] S22, the aligned hyper-dimensional feature vector set is input into a dynamic gate fusion network, the dynamic weight distribution module built-in the network is used to adaptively adjust the contribution weight of each modal feature according to the importance of each modal feature in variety discrimination (for example, the representation advantage of hyperspectral features for internal components of varieties), and a feature vector set after weight distribution is generated;

[0061] S23, for the feature vector set after weight distribution, introduce cross-modal attention mechanism, strengthen the association between key features (for example, the association feature of "fruit shape index" in three-dimensional shape and "fruit skin color" in macroscopic vision), suppress irrelevant feature interference, and output robust fusion feature vector;

[0062] S24, input the robust fusion feature vector into a lightweight classifier (for example, an improved convolutional neural network classifier), set a classification threshold, perform initial variety classification on the jujube sample, and obtain a preliminary variety discrimination result;

[0063] S25, based on the probability distribution output by the classifier, calculate the confidence score corresponding to the preliminary variety discrimination result, set a confidence threshold (for example, 0.8), and mark the low-confidence classification results below the threshold.

[0064] The working principle and effect of the above technical solution are as follows: by aligning the dimensions and formats of the features of each mode, the fusion confusion caused by the non-uniformity of the data form is avoided, the subsequent feature processing is smoother, and the fusion deviation caused by the mode difference is reduced;

[0065] The dynamic gating network adjusts the weight according to the actual role of each mode in discrimination, for example, highlights the representation advantage of hyperspectrum for internal composition, reduces the irrationality of one-size-fits-all weight allocation, and improves the rationality of feature weight allocation;

[0066] The cross-modal attention mechanism strengthens the core associated features such as fruit shape index and fruit skin color, while suppressing irrelevant information interference, so that the output fusion feature vector is more focused on variety discrimination key, and the relevance of key features is enhanced;

[0067] Processing the robust fusion features with a lightweight classifier reduces the computational burden while ensuring discrimination effect, avoids the time-consuming problem caused by complex models, and improves the efficiency of initial classification;

[0068] By calculating the confidence and marking the results below the threshold, a clear direction is provided for subsequent correction, avoiding the use of unreliable results to affect the final discrimination accuracy, and reducing the misjudgment risk of low-confidence results.

[0069] In an embodiment of the present application, the S22 comprises:

[0070] Parse the super-dimensional feature vector set after alignment, extract the feature sub-vectors of the four modes, the four modes include macroscopic vision, hyperspectrum, three-dimensional shape and microscopic texture, determine the core representation direction of each mode feature in variety discrimination (for example, hyperspectrum corresponds to internal composition, three-dimensional shape corresponds to appearance structure), and generate a mode feature classification list;

[0071] The modal feature classification list is input into the feature importance evaluation submodule of the dynamic gating fusion network. Based on the feature contribution statistics of historical discrimination samples (e.g., the accuracy of hyperspectral features in the discrimination of variety A is improved by 35%), the basic weight values ​​of each modal feature are initially calculated to form a basic weight matrix.

[0072] The real-time adjustment algorithm of the dynamic weight allocation module is invoked, and the basic weight matrix is ​​dynamically corrected by taking into account the special characteristics of the current jujube samples (e.g., the microscopic texture of some samples is blurred and the weight of this mode needs to be reduced) to generate real-time weight adjustment coefficients.

[0073] The real-time weight adjustment coefficients are weighted dimension-wise with each modal feature sub-vector to ensure that high-importance modal features (such as hyperspectral features) have a higher proportion in the fusion, while the proportion of low-importance modal features is reasonably reduced, resulting in weighted modal feature sub-vectors.

[0074] The weighted feature vectors of each modality are reorganized according to the order of the hyperdimensional feature space to form a set of feature vectors with complete structure and reasonable weight allocation.

[0075] The working principle and effects of the above technical solution are as follows: By analyzing the vector set, the core representation direction of each modality is clarified and a classification list is generated, avoiding a vague understanding of the role of modal features, making subsequent weight allocation more in line with the discrimination requirements, and improving the pertinence of modal feature analysis; the basic weights are calculated by combining the feature contribution statistics of historical samples, reducing the subjectivity of setting weights based on experience, making the initial weight allocation more data-supported, and enhancing the reliability of the basic weights; the weights are dynamically corrected according to the current sample features (e.g., blurred microscopic texture), avoiding the rigidity of using fixed weights to deal with all samples, making weight adjustment more flexible and adaptable to the actual situation, and reducing the weight bias caused by sample particularity; the weighted operation is performed dimension by dimension to ensure the proportion of high-importance modalities (e.g., hyperspectral), avoiding the imbalance of local feature weights caused by overall weighting, making the contribution of each modal feature more reasonable, and improving the accuracy of feature weighting; the weighted sub-vectors are re-integrated according to the order of the hyperdimensional space dimensions, avoiding the impact of disordered vector arrangement on subsequent processing, making the final generated feature vector set structure more regular and more usable, and reducing the structural chaos of the integrated vector set.

[0076] In one embodiment of the present invention, step S3 includes:

[0077] S31. Load the prior knowledge base of the genus *Ziziphus* evolutionary lineage, perform structured analysis on the evolutionary tree of *Ziziphus* species, genetic marker data (e.g., SSR molecular markers), and ecological distribution information stored in the base, extract key knowledge that can be used for variety identification correction (e.g., genetic differences of closely related varieties, ecological distribution range of specific varieties), and generate a structured prior knowledge set.

[0078] S32, the preliminary variety discrimination result is semantically aligned with the structured prior knowledge set, entity matching (for example, variety name standardization matching), feature mapping (for example, matching the fruit skin thickness in the discrimination result with the genetic marker associated feature in the knowledge base), unified data semantic standard, and an aligned discrimination-knowledge data set is generated;

[0079] S33, for the low confidence classification result in the aligned discrimination-knowledge data set, a knowledge graph reasoning engine is called, and reasoning correction is performed based on the evolutionary tree kinship rule and the genetic marker matching degree rule (for example, genetic similarity >= 95% is determined as a close relative variety), and a corrected variety discrimination result is obtained;

[0080] S34, based on the corrected variety discrimination result, the role path of each modality feature on the classification decision (for example, specific band reflectivity in hyperspectrum -> determine variety sugar content -> associate variety A) is traced, a variety feature attribution graph is generated, and the contribution proportion of each modality feature is determined;

[0081] S35, a visual tool (for example, a heat map, a feature contribution line chart) is used to convert the variety feature attribution graph and the discrimination process into a visual interface, and the influence of different modality features on the final discrimination result is intuitively displayed, and the visual explanation of the discrimination process is realized.

[0082] The working principle and effect of the above technical solution are as follows: key correction information is extracted from the structured analysis knowledge base, application difficulties caused by knowledge fragmentation are avoided, evolutionary pedigree, genetic markers and other data can accurately support discrimination correction, and the utilization rate of prior knowledge is improved; entity matching and feature mapping are used to unify the semantic standard, knowledge association failure caused by non-standard variety names and inconsistent feature expressions is avoided, the discrimination result is more smoothly connected with the knowledge base, and the matching error caused by data semantic deviation is reduced; based on the evolutionary tree kinship rule and the genetic matching degree rule, the reasoning correction is performed, the one-sidedness of the judgment based on only data features is reduced, the probability of misjudgment of close relative varieties is reduced, and the correction accuracy of low confidence results is enhanced; the variety feature attribution graph is generated to determine the contribution path of each modality, the problem of knowing the phenomenon but not knowing the reason in the traditional discrimination is avoided, the basis of each discrimination result is clear and traceable, and the traceability of the discrimination process is improved; the process is displayed by using a visual tool such as a heat map and a line chart, understanding difficulties caused by complex data and reasoning logic are avoided, non-professionals can also intuitively grasp the influence of features on discrimination, and the understanding threshold of discrimination results is reduced.

[0083] In an embodiment of the present application, the S4 comprises:

[0084] S41, a blockchain credible traceability system is constructed, a storage structure of the blockchain is designed, storage fields (collection time, geographic position, multi-modal original data hash value, hyper-dimensional feature vector abstract, variety discrimination result) of life data are determined, and a data access permission is set by using an asymmetric encryption algorithm;

[0085] S42, complete life data of jujube samples is collected, the data is formatted according to the blockchain storage structure, a standardized life data set is generated, and a sample unique identifier (for example, a sample number) is associated;

[0086] S43, the standardized life data set is uploaded to the blockchain system, data storage is completed through a consensus mechanism (for example, PoS proof of stake), and an unalterable life data chain is formed;

[0087] S44, a blockchain smart contract is called, a sample unique digital fingerprint is generated based on a hash value of the standardized life data, the digital fingerprint is bound with the sample identifier, and serves as a credible certificate of the sample identity;

[0088] S45, a distributed verification mechanism of the blockchain node is started, cross-regional jujube variety discrimination results and corresponding life data are loaded by different regional blockchain nodes, cross-regional discrimination results are cross-verified through comparison of digital fingerprint consistency and discrimination logic compliance, and a unified discrimination result that passes verification is output.

[0089] The working principle and effect of the above technical solution are as follows: through the design of the blockchain structure and the setting of the access permission by the asymmetric encryption, the risk of life data being randomly tampered with or accessed beyond authority is avoided, the sample full-process data is more reliable, and the security of data storage is enhanced; the life data is formatted according to the blockchain storage structure and is associated with the unique identifier, the difficulty of traceability caused by data format disorder is reduced, the data of each sample can be accurately corresponded, and the uniformity of the data format is improved; the data storage is completed through the consensus mechanism to form an unalterable life data chain, the problem that the discrimination result is affected by human modification of data is avoided, the originality of the data is ensured, and the possibility of data tampering is reduced; the unique digital fingerprint is generated based on the data hash value and is bound with the sample identifier, the risk of sample confusion or identity forgery is reduced, each sample has a credible digital identity certificate, and the uniqueness of the sample identity is enhanced; the fingerprint and the logic are verified and compared through the distributed node verification, the result conflict caused by different regional discrimination standards is avoided, the dispute in cross-regional application is reduced, and the consistency of the cross-regional discrimination result is improved.

[0090] In an embodiment of the present application, the S45 comprises:

[0091] Extracting the variety identification results and corresponding life data (including digital fingerprints and identification logic explanations) of cross-regional jujube samples from the blockchain system, grouping the data according to regions, and generating regionalized identification-data correlation packages;

[0092] Sending verification instructions to each regional node of the blockchain and distributing the corresponding regionalized identification-data correlation packages. After receiving the data packages, the nodes parse the package structure, confirm the data integrity (e.g., no missing fields, and identifiable hash values), and generate node data reception confirmation information;

[0093] Each regional node first performs digital fingerprint consistency verification. It calculates the hash value of the life data received locally and compares it with the digital fingerprint attached in the package. If they are consistent, it proceeds to the next step. If they are not consistent, it marks the fingerprint as abnormal and records the abnormal reason.

[0094] For data packages that pass the fingerprint verification, the node starts the identification logic compliance verification. It checks the logical correlation between the identification results and the life data (e.g., whether the hyperspectral characteristics support the determination results) against the prior knowledge of the Ziziphus evolutionary pedigree (e.g., the genetic feature matching degree and ecological distribution adaptability) and generates a logic verification report.

[0095] Summarizing the verification results of each regional node (including fingerprint verification results and logic verification reports), calculating the proportion of nodes that pass the verification (e.g., if ≥80% of the nodes pass, it is considered valid), and outputting the unified identification results if the verification passes or triggering a secondary verification process if it does not.

[0096] The working principle and effects of the above technical solutions are as follows: The correlation packages are divided according to regions and distributed to corresponding nodes, avoiding verification confusion caused by data mixing, making the verification target of each node more clear, and improving the organization of cross-regional data verification; After receiving the data, the nodes first confirm the data integrity, avoiding the impact of problems such as missing fields and hash value errors on subsequent verification, reducing the waste of invalid verification, and reducing the risk of data transmission loss; Through digital fingerprint consistency comparison, it is possible to investigate whether the data has been tampered with from the source, reducing the possibility of false data participating in verification and enhancing the verification strength of data authenticity; After the fingerprint verification passes, it is checked against prior knowledge to check the logical correlation, avoiding the problem of only looking at the surface consistency of the data and ignoring the internal rationality, improving the logical reliability of the identification results; The final result is determined according to the node passing rate, avoiding the one-sidedness of a single region deciding, making the unified identification result more credible, and reducing the controversy of cross-regional results.

[0097] In one embodiment of the present application, the S5 comprises:

[0098] S51, collect core data at each link, dynamically gate the robustness discrimination results output by the fusion network, the enhanced explanation data of the Ziziphus evolutionary phylogeny correction (for example, key information of attribution graph, correction reasoning process), and the blockchain traceability verification information (digital fingerprint, node verification record), unify the data format and remove redundant information, and generate integrated discrimination data set;

[0099] S52, structure analysis of the integrated discrimination data set, extraction of core information, including variety name, biological attribution of each modal feature (for example, flavonoid content corresponding to hyperspectral feature of variety), final discrimination confidence, and genetic relationship with target variety in evolutionary phylogeny (for example, genetic similarity with variety B is 92%, third branch of evolutionary tree of the same genus);

[0100] S53, call the preset discrimination report template, fill the structured analysis core information into the template according to the logical order of variety basic information-feature attribution analysis-discrimination reliability explanation-traceability certificate, and generate the preliminary structured jujube variety discrimination report;

[0101] S54, check the preliminary discrimination report, check the data accuracy (for example, the variety name is consistent with the blockchain storage), the logical coherence (for example, the attribution analysis matches the discrimination result), and the information integrity (for example, there is no missing storage certificate), and correct the deviation and omission in the report;

[0102] S55, output the structured jujube variety discrimination report that passes the verification, which contains variety name, biological feature attribution analysis (including contribution ratio of each modal feature), discrimination confidence (with confidence calculation basis), evolutionary phylogeny correlation description (including evolutionary tree branch diagram), and blockchain storage certificate (including digital fingerprint query link).

[0103] The working principle and effect of the above technical solution are as follows: by collecting core data at each link, unifying the format, and removing redundancy, the scattered and disordered data is avoided, the integrated data set generated is more refined and usable, and the efficiency of data integration is improved; structure analysis extracts key information, reduces irrelevant information interference to core conclusion, highlights report focus, and enhances core focus of report content; fill in the preset template according to the fixed logical order, avoid reading difficulty caused by unordered content arrangement, make the preliminary report structure more regular, and reduce the format disorder problem of report generation; through verification, ensure data accuracy, logical coherence and information integrity, reduce the report reliability caused by omission or error, make the final output report more rigorous, and improve the reliability of the report; including feature contribution ratio, confidence basis, evolutionary tree diagram and other details, avoiding the hollow and general report content, making the user can fully understand the discrimination process and basis, facilitating subsequent application and traceability, and enhancing the practical value of the report.

[0104] One embodiment of the present application is a jujube variety discrimination system based on artificial intelligence, comprising:

[0105] one or more processors;

[0106] a memory for storing one or more programs,

[0107] wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the method of any one of the above.

[0108] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A method for identifying jujube varieties based on artificial intelligence, characterized in that, The method includes: S1: Synchronously collect multimodal raw data of jujube samples to construct a multimodal raw dataset; preprocess the multimodal raw dataset to obtain standardized multimodal feature data; construct a hyperdimensional feature space based on the standardized multimodal feature data, and generate a hyperdimensional feature vector set of jujubes through feature cross-fusion; including: S14. Based on standardized multimodal feature data, establish a hyperdimensional feature space coordinate system, and map macroscopic visual features, hyperspectral features, three-dimensional morphological features, and microscopic texture features to different dimensional axes to construct a hyperdimensional feature space. S2: Input the hyperdimensional feature vector set into the dynamic gated fusion network, and adaptively adjust the contribution of each modality feature through the dynamic weight allocation module. Combine the cross-modal attention mechanism to strengthen the association of key features and output a robust fusion feature vector. Based on the robust fusion feature vector, perform initial variety classification to obtain preliminary variety discrimination results and corresponding confidence scores. S3: Introduce a prior knowledge base of the genus *Ziziphus* evolutionary lineage, semantically align the preliminary variety identification results with the prior knowledge base of the genus *Ziziphus* evolutionary lineage, and correct low-confidence classification results through knowledge graph reasoning; based on the corrected results, generate a variety feature attribution graph to clarify the contribution path of each modality feature to the classification decision. S4: Construct a blockchain-based trusted traceability system to store the life data of jujube samples on the blockchain and generate unique digital fingerprints through smart contracts; based on the distributed verification mechanism of blockchain nodes, cross-verify the identification results of jujube varieties across regions; S5: Integrate the fusion data from the dynamic gating fusion network to generate a structured jujube variety identification report; The S4 includes: S41. Construct a blockchain-based trusted traceability system, design the blockchain's storage structure, and use asymmetric encryption algorithms to set data access permissions; S42. Collect complete life data of jujube samples, format the data according to the blockchain storage structure, generate a standardized life dataset, and associate it with a unique sample identifier; S43. Upload the standardized life dataset to the blockchain system, complete the on-chain storage of data through the consensus mechanism, and form an immutable life data chain. S44. Call the blockchain smart contract to generate a unique digital fingerprint for the sample based on the hash value of the standardized life data, and bind the digital fingerprint with the sample identifier as a trusted credential for the sample's identity. S45. Activate the distributed verification mechanism of blockchain nodes. Blockchain nodes in different regions load the cross-regional jujube variety identification results and corresponding life data respectively, cross-verify the cross-regional identification results, and output a unified identification result that has passed the verification.

2. The method for identifying jujube varieties based on artificial intelligence according to claim 1, characterized in that, S1 includes: S11. Screen the jujube samples, remove invalid samples damaged by pests and diseases or with incomplete morphology, and determine the set of valid samples. S12. For the effective sample set, use multiple devices to synchronously collect multimodal raw data and integrate all data to construct a multimodal raw dataset; S13. Preprocess the original multimodal dataset to obtain standardized multimodal feature data; S15. Perform feature cross-fusion in the hyperdimensional feature space, and use the attention mechanism to explore the correlation between features of each modality to generate a hyperdimensional feature vector set of jujube.

3. The method for identifying jujube varieties based on artificial intelligence according to claim 2, characterized in that, S15 includes: S151. Extract key anchor points of each modality feature in the hyperdimensional feature space to form a multimodal feature anchor point set; S152. Calculate the spatial correlation between feature anchor sets, measure the spatial proximity of different modal anchors by Euclidean distance, evaluate the matching degree of feature attributes by cosine similarity, and generate an anchor correlation matrix. S153. Input the anchor point correlation matrix into the attention mechanism model, assign high weight values ​​to high correlation anchor point pairs, and generate a weighted feature correlation map. S154. Based on the weighted feature association map, cross-mapping and fusion of multimodal features in the hyperdimensional feature space are performed. The feature dimensions with high correlation are vector superimposed, while the feature dimensions with low correlation are kept independently represented to form a draft of the fused feature vector. S155. Perform dimensionality normalization on the initial draft of the fused feature vector to eliminate the fusion bias caused by the difference in the magnitude of features of different modalities, and finally generate a set of jujube-type hyperdimensional feature vectors.

4. The method for identifying jujube varieties based on artificial intelligence according to claim 3, characterized in that, S154 includes: The weighted feature association graph is analyzed, and high-association feature dimension pairs with an association degree above the threshold are extracted. At the same time, low-association feature dimensions with an association degree <0.8 are filtered out, and a high / low-association feature dimension classification table is generated. For highly correlated feature dimension pairs, the feature vector data corresponding to each dimension is obtained, and the vector dot product operation is used to realize the feature superposition between dimensions. The core feature information of the superimposed vector is retained to generate a highly correlated fused feature vector group. For low-association feature dimensions, extract the feature vectors of each dimension separately, and generate low-association independent feature vector groups by labeling their modality types with feature identifiers; The highly correlated fused feature vector group and the low-correlation independent feature vector group are concatenated according to the dimensional order of the hyperdimensional feature space to form a combined vector sequence containing fused features and independent features; The integrity of the combined vector sequence is verified, and the initial draft of the fused feature vector is finally formed.

5. The method for identifying jujube varieties based on artificial intelligence according to claim 1, characterized in that, S2 includes: S21. Perform modal feature alignment processing on the jujube hyperdimensional feature vector set to generate an aligned hyperdimensional feature vector set. S22. Input the aligned hyperdimensional feature vector set into the dynamic gated fusion network. Through the network's built-in dynamic weight allocation module, the contribution weight of each modality feature is adaptively adjusted according to the importance of each modality feature in variety discrimination, and a weighted feature vector set is generated. S23. For the feature vector set after weight allocation, a cross-modal attention mechanism is introduced to strengthen the correlation between key features, suppress interference from irrelevant features, and output a robust fused feature vector. S24. Input the robust fusion feature vector into the lightweight classifier, set the classification threshold, perform initial variety classification on the jujube samples, and obtain preliminary variety discrimination results; S25. Based on the probability distribution output by the classifier, calculate the confidence score corresponding to the preliminary variety identification result, set a confidence threshold, and mark the low confidence classification results below the threshold.

6. The method for identifying jujube varieties based on artificial intelligence according to claim 5, characterized in that, S22 includes: After parsing the aligned hyperdimensional feature vector set, feature sub-vectors of the four modalities are extracted, the core representation direction of each modal feature in variety discrimination is determined, and a list of modal feature classifications is generated. The modal feature classification list is input into the feature importance evaluation submodule of the dynamic gating fusion network. Based on the feature contribution statistics of historical discrimination samples, the basic weight values ​​of each modal feature are initially calculated to form a basic weight matrix. The real-time adjustment algorithm of the dynamic weight allocation module is invoked, and the basic weight matrix is ​​dynamically corrected based on the specific characteristics of the current jujube samples to generate real-time weight adjustment coefficients. The real-time weight adjustment coefficients are weighted dimension by dimension with the feature vectors of each modality to obtain the weighted feature vectors of each modality. The weighted feature vectors of each modality are reorganized according to the dimensional order of the hyperdimensional feature space to form a set of feature vectors after weight allocation.

7. The method for identifying jujube varieties based on artificial intelligence according to claim 1, characterized in that, The S3 includes: S31. Load the genus *Ziziphus* evolutionary phylogenetic prior knowledge base, perform structured analysis on the phylogenetic tree, genetic marker data and ecological distribution information of *Ziziphus* species stored in the base, extract key knowledge that can be used for variety identification and correction, and generate a structured prior knowledge set. S32. Semantically align the preliminary variety identification results with the structured prior knowledge set. Through entity matching and feature mapping, unify the data semantic standards and generate the aligned discrimination-knowledge dataset. S33. For the low-confidence classification results in the aligned discriminant-knowledge dataset, call the knowledge graph reasoning engine to perform reasoning correction based on the kinship rules of the evolutionary tree and the matching degree rules of genetic markers to obtain the corrected variety discrimination results; S34. Based on the corrected variety discrimination results, trace the role path of each modal feature in the classification decision, generate a variety feature attribution map, and clarify the contribution ratio of each modal feature. S35. Use visualization tools to transform the variety characteristic attribution map and discrimination process into a visualization interface, intuitively showing the impact of different modal characteristics on the final discrimination result, and realizing a visual explanation of the discrimination process.

8. The method for identifying jujube varieties based on artificial intelligence according to claim 1, characterized in that, The S5 includes: S51. Collect core data from each stage, standardize the data format and remove redundant information to generate an integrated discrimination dataset. S52. Perform structured analysis on the integrated discriminant dataset to extract core information; S53. Call the preset discrimination report template, fill the core information after structured analysis into the template in the logical order of variety basic information - characteristic attribution analysis - discrimination reliability description - traceability certificate, and generate a preliminary structured jujube variety discrimination report; S54. Verify the preliminary judgment report, check the accuracy of the data, the logical coherence, and the completeness of the information, and correct any deviations or omissions in the report; S55. Output a structured jujube variety identification report that has passed verification.

9. An artificial intelligence-based jujube variety identification system, comprising: One or more processors; Memory, used to store one or more programs. Wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1 to 8.

Images