A method, system, storage medium, and computer device for judging tobacco leaf quality based on multi-view latent space mapping.

Abstract

This invention discloses a method, system, storage medium, and computer device for judging tobacco leaf quality based on multi-view latent space mapping. It constructs a tobacco leaf quality grade database by acquiring near-infrared hyperspectral images of tobacco leaves of different grades and origins. Spectral images of the pure leaf surface region are obtained through masking images. The processed spectral images are divided into a predetermined number of complementary views. The multi-view latent space mapping model is trained using a dataset of complementary views of tobacco leaves to obtain a multi-view latent space mapping model with an optimal objective function. Based on the multi-view latent space mapping model, a classifier is used to finally obtain the grading result of the tobacco leaves. This method can fuse feature representations from different perspectives or modalities, integrating multi-view latent space modeling with graph structure representation for tobacco leaf quality recognition tasks. This significantly reduces annotation requirements while improving the system's modeling ability and classification robustness for complex spectral data.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of tobacco processing, and particularly relates to a tobacco leaf quality judgment method and system based on multi-view latent space composition, a storage medium and a computer device. BACKGROUND

[0002] Tobacco leaf quality evaluation is a core link in tobacco industrial production, and directly affects the sensory quality, style characteristics and market competitiveness of cigarette products. For a long time, the grading of tobacco leaf quality mainly relies on artificial sensory evaluation, which is based on the appearance characteristics of tobacco leaves such as color, oil content, structure and maturity, as well as sensory indicators such as combustibility and aroma intensity. However, this traditional method has obvious technical limitations. First, the evaluation is highly dependent on expert experience and is highly subjective, and is easily affected by factors such as the evaluator's knowledge background, emotional fluctuations and working state; second, the repeatability is poor, and even the same evaluator may have significant differences in evaluating the same sample at different times; in addition, the artificial evaluation is inefficient and difficult to meet the needs of rapid detection and large-scale processing in current industrial production; finally, there is a lack of unified and quantitative evaluation standards, which makes it difficult to compare the quality of different production areas and different batches objectively and to realize unified control and quality traceability in the national or industry range.

[0003] According to the above technical problems, a Chinese invention patent with application publication number CN110807760A discloses a tobacco leaf grading method and system, which proposes to train a model using historical labeled tobacco leaf images and a small amount of labeled tobacco leaf images of the current year to realize automatic grading of the current year's tobacco leaves. This method has made some progress in improving the grading efficiency, but it still relies on a large amount of labeled data, especially in cases where the historical data quality is not high or the production area changes greatly, the model accuracy and robustness are still limited.

[0004] In the prior art, hyperspectral imaging technology, as an advanced detection method combining imaging and spectral analysis, can simultaneously capture the morphological characteristics of the surface of tobacco leaves and the spectral response of the internal chemical composition, realizing non-contact, non-destructive and rapid detection of tobacco leaf quality. With its high spatial resolution and spectral resolution capability, this technology provides a new path for the objective evaluation and intelligent recognition of tobacco leaves. However, hyperspectral data has the natural characteristics of "high dimension, redundancy and heterogeneity", and faces many challenges in practical application. On the one hand, a single hyperspectral image often contains hundreds of bands, with extremely high feature space dimension, which easily causes "dimension disaster"; on the other hand, the high correlation between different bands leads to serious information redundancy and low effective information density. In addition, obtaining high-quality labeled samples requires relying on experienced experts, and the data collection cost is high and the period is long, which makes it difficult to support large-scale model training; furthermore, there are significant cross-domain differences in hyperspectral response between tobacco leaves from different production areas, different grades and even different years, causing obvious degradation in model generalization performance.

[0005] Therefore, a new method for judging the quality of tobacco leaves is needed to solve the above-mentioned technical problems. Summary of the Invention

[0006] The purpose of this invention is to provide a method for judging tobacco quality based on multi-view latent space composition, which solves the technical problems of low recognition accuracy and weak robustness of tobacco classification in the prior art.

[0007] The present invention also aims to provide a judgment system that employs a method for judging tobacco leaf quality based on multi-view latent space composition.

[0008] Another objective of this invention is to provide a computer-readable storage medium.

[0009] Another objective of this invention is to provide a computer device.

[0010] The technical solution of this invention to solve its technical problem is as follows:

[0011] A method for judging tobacco leaf quality based on multi-view latent space composition includes the following steps:

[0012] S1: Collect near-infrared hyperspectral images of tobacco leaves of different grades and origins within a set wavelength range, and obtain the real quality grade labels corresponding to the tobacco leaf samples to establish a tobacco leaf quality grade database.

[0013] S2: A mask image is constructed by fusing a fixed band threshold and the reflectance difference. After removing the background, leaf stem, leaf veins and shadows of each tobacco leaf from the mask image to obtain the spectral image of the pure leaf surface area of ​​the tobacco leaf, abnormal samples are removed from the spectral image of the pure leaf surface area of ​​the tobacco leaf.

[0014] S3: The spectral image of the pure leaf surface region of the processed tobacco leaf is divided into a set number of complementary views according to the dimensions. Each view corresponds to a different band interval of the spectral image. Normalization is performed on each view to establish a tobacco leaf complementary view dataset. The multi-view latent space mapping model is trained using the tobacco leaf complementary view dataset to obtain a multi-view latent space mapping model with an optimal solution objective function. The objective function of the multi-view latent space mapping model includes minimizing the reconstruction error, similarity graph structure, and multi-view feature representation of label prediction.

[0015] S4: Based on the multi-view latent space mapping model, the quality of unlabeled tobacco leaves is judged, and the final quality grade is determined by a classifier, and the grading results are output.

[0016] Preferably, the objective function of the multi-view latent space mapping model in step S3 specifically includes:

[0017] The objective function term corresponding to minimizing the reconstruction error is expressed as:

[0018] ;

[0019] The objective function term corresponding to the similarity graph structure is expressed as:

[0020] ;

[0021] The objective function term corresponding to label prediction is expressed as:

[0022] ;

[0023] The constraints are as follows: , , , , R represents a real number, r represents the potential spatial dimension of the tobacco leaf, and N represents the number of tobacco leaves. This represents the time-delay characteristic dimension of tobacco leaves. Indicates the number of categories. Indicates the first The original feature matrix corresponding to each view; Indicates the first A mapping matrix for each view; The latent space representation is a low-dimensional representation common to all views; The regularization coefficient represents the reconstruction error; Indicates the first The representation of a tobacco leaf in the latent space. Indicates the first The representation of a tobacco leaf in the latent space; Indicates the first The spectral image of the first tobacco leaf and the second Similarity between spectral images of individual tobacco leaves; The regularization coefficients of similar matrices are represented. A similarity matrix representing the spectral images of tobacco leaves; Represents the trace of a matrix. express The transpose of the matrix; Represents the Laplace matrix; Represents the label prediction matrix. This indicates that the corresponding labels for the tobacco leaves have been marked; This represents the weighting coefficient that controls the intensity of label supervision.

[0024] Preferably, the specific steps for training the multi-view latent space mapping model in step S3 are as follows: following an alternating optimization strategy, the mapping matrix, latent space representation, similarity matrix, and label prediction matrix are optimized sequentially, and the optimization steps are repeated iteratively until the objective function converges or the set number of iterations is reached.

[0025] Preferably, the objective function of the optimal solution in step S3 is the sum of the objective function term for minimizing the reconstruction error, the objective function term for the similarity graph structure, and the objective function term for label prediction, specifically:

[0026] ;

[0027] The constraints are: .

[0028] Preferably, the set quantity in step S3 is 5 or 7.

[0029] Preferably, the wavelength range set in step S1 is 1000–2500 nm.

[0030] Preferably, step S4 specifically involves: determining the quality grade of unlabeled tobacco leaves based on the maximum response value of the corresponding label prediction matrix in the multi-view latent space mapping model, determining the quality grade of the tobacco leaves using a classifier, determining the quality grade of the tobacco leaves using a majority voting mechanism, and outputting the grading result.

[0031] Preferably, the specific process of the majority voting mechanism to judge the quality of tobacco leaves is as follows: determine whether there is a set proportion of grading results that are consistent in the spectral curve of the tobacco leaves; if so, output the grading result corresponding to the set proportion; if not, output the grading result with the highest prediction confidence among the grading results.

[0032] A judgment system employing a tobacco leaf quality judgment method based on multi-view latent space mapping includes:

[0033] The image acquisition module is used to acquire near-infrared hyperspectral images of tobacco leaves of different grades and origins within a set wavelength range, and to obtain the real quality grade labels corresponding to the tobacco leaf samples to establish a tobacco leaf quality grade database.

[0034] The spectral image extraction module constructs a mask image by fusing a fixed band threshold and the reflectance difference. After removing the background, leaf stem, leaf veins, and shadows of each tobacco leaf from the mask image to obtain the spectral image of the pure leaf surface area of ​​the tobacco leaf, abnormal samples are removed from the spectral image of the pure leaf surface area of ​​the tobacco leaf.

[0035] The multi-view latent space mapping module divides the spectral image of the pure leaf surface region of the processed tobacco leaf into a set number of complementary views according to dimensions through the feature construction module. Each view corresponds to a different band interval of the spectral image. Normalization is performed on each view to establish a tobacco leaf complementary view dataset. The latent space learning module performs joint non-negative matrix factorization on each view under the guidance of unified latent variables to obtain the latent space representation. The graph structure module constructs a similarity graph structure reflecting the local relationships of samples. The label propagation module performs semi-supervised classification of tobacco leaf labels based on weight coefficient constraints. The multi-view latent space mapping model is trained using the tobacco leaf complementary view dataset to obtain a multi-view latent space mapping model with an optimal solution objective function. The objective function of the multi-view latent space mapping model includes minimizing reconstruction error, similarity graph structure, and multi-view feature representation of label prediction.

[0036] The decision output module judges the quality of unlabeled tobacco leaves based on a multi-view latent space mapping model, determines the final quality grade through a classifier, and outputs the grading results.

[0037] A computer-readable storage medium storing a computer program, characterized in that, when executed by a processor, the computer program causes the device containing the computer-readable storage medium to perform a tobacco leaf quality judgment method based on a multi-view latent space composition.

[0038] A computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit, and an input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are connected to the system bus via the input / output interface. The memory includes a computer-readable storage medium.

[0039] The beneficial effects of this invention are as follows: By acquiring near-infrared hyperspectral images of tobacco leaves of different grades and origins within a set wavelength range, the diversity of tobacco leaf samples is ensured; and the true quality grade labels corresponding to the tobacco leaf samples are obtained to establish a tobacco leaf quality grade database; a mask image is constructed by fusing fixed band thresholds and reflectance differences; after removing the background, leaf stems, veins, and shadows of the tobacco leaves from the mask image to obtain the spectral image of the pure leaf surface area, abnormal samples are removed from the spectral image of the pure leaf surface area to ensure the homogeneity and usability of the input data; the processed pure tobacco leaf... The spectral image of the leaf region is divided into a predetermined number of complementary views according to dimensions. Each view corresponds to a different band interval of the spectral image to enhance the ability to express heterogeneous information about tobacco quality. After normalization processing of each view, multi-source complementary information is extracted to establish a tobacco leaf complementary view dataset. The multi-view latent space mapping model is trained using the tobacco leaf complementary view dataset to obtain a multi-view latent space mapping model with an optimal objective function. Based on the multi-view latent space mapping model, the quality of unlabeled tobacco leaves is judged, and the final quality grade is determined by a classifier, outputting the grading result. This method can integrate feature representations from different perspectives or modalities, and extract more discriminative latent features by collaboratively modeling in multiple subspaces. At the same time, graph structure learning can construct a similarity relationship network between samples, and combined with the label propagation mechanism, it can achieve effective classification of unknown tobacco leaves in the case of scarce labels. The integration of multi-view latent space modeling and graph structure representation in the tobacco quality recognition task significantly reduces the labeling requirements while improving the system's modeling ability and classification robustness for complex spectral data, thus providing technical support for building an industrial-grade refined tobacco quality evaluation system. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating the tobacco quality judgment method based on multi-view latent space composition of the present invention.

[0041] Figure 2 This is a schematic diagram of the specific identification model corresponding to the tobacco quality judgment method based on multi-view latent space composition of the present invention;

[0042] Figure 3 This is a schematic diagram of the tobacco leaf extraction effect in Embodiment 1 of the present invention;

[0043] Figure 4 This is a similarity diagram of spectral features of a certain experiment on the Red River data in Exp02, according to Embodiment 1 of the present invention;

[0044] Figure 5 This is the label confidence matrix of a certain experiment on the Red River data in Exp02, as described in Embodiment 1 of the present invention;

[0045] Figure 6This is a similarity diagram of spectral features of a certain experiment on the Red River data in Exp03, according to Embodiment 1 of the present invention;

[0046] Figure 7 This is the label confidence matrix of a certain experiment on the Red River data in Exp03, as described in Embodiment 1 of the present invention;

[0047] Figure 8 This is a schematic diagram of the computer device of the present invention. Detailed Implementation

[0048] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0049] like Figure 1 , Figure 2 As shown, this invention discloses a method for judging tobacco leaf quality based on multi-view latent space composition, including the following steps:

[0050] S1: Acquire near-infrared hyperspectral images of tobacco leaves of different grades and origins within a set wavelength range, and obtain the corresponding true quality grade labels for the tobacco leaf samples to establish a tobacco leaf quality grade database; the set wavelength range is 1000–2500 nm. Generally, to enhance spectral stability, the samples are equilibrated under constant temperature and humidity conditions before acquiring near-infrared hyperspectral images. Constant temperature and humidity conditions generally refer to the environmental conditions of tobacco leaves after being placed in an environment at 45℃ and 80% humidity for 48 hours.

[0051] S2: A mask image is constructed by fusing a fixed band threshold and the reflectance difference. After removing the background, leaf stem, veins, and shadows of each tobacco leaf from the mask image to obtain the spectral image of the pure leaf surface area, abnormal samples are removed from the spectral image of the pure leaf surface area. By removing abnormal samples, the homogeneity and usability of the input data are ensured.

[0052] S3: The spectral image of the pure leaf surface region of the processed tobacco leaf is divided into a predetermined number of complementary views according to dimensions. Each view corresponds to a different band interval of the spectral image. Normalization is performed on each view to establish a tobacco leaf complementary view dataset. This dataset is then used to train a multi-view latent space mapping model, resulting in a multi-view latent space mapping model with an optimal objective function. The objective function of the multi-view latent space mapping model includes minimizing reconstruction error, similarity graph structure, and label prediction multi-view feature representation. Dividing the image into different band intervals enhances the ability to express heterogeneous information about tobacco leaf quality. Normalization is used to extract multi-source complementary information. The predetermined number of complementary views is 5 or 7.

[0053] The objective function of the multi-view latent space mapping model in step S3 specifically includes:

[0054] The objective function term corresponding to minimizing the reconstruction error is expressed as:

[0055] This function is designed to improve the discriminative power of latent space representations.

[0056] The objective function term corresponding to the similarity graph structure is expressed as:

[0057] The purpose of this function is to accurately model the data distribution using the geometric structure of the graph, enhance the model's ability to model complex sample distribution structures, and improve its robustness to anomalous samples.

[0058] The objective function term corresponding to label prediction is expressed as:

[0059] This function is for the expansion and dissemination of tobacco leaf labels.

[0060] The constraints are as follows: , , , , R represents a real number, r represents the potential spatial dimension of the tobacco leaf, and N represents the number of tobacco leaves. This represents the time-delay characteristic dimension of tobacco leaves. Indicates the number of categories. Indicates the first The original feature matrix corresponding to each view; Indicates the first A mapping matrix for each view; The latent space representation is a low-dimensional representation common to all views; The regularization coefficient represents the reconstruction error; Indicates the first The representation of a tobacco leaf in the latent space. Indicates the first The representation of a tobacco leaf in the latent space; Indicates the first The spectral image of the first tobacco leaf and the second Similarity between spectral images of individual tobacco leaves; The regularization coefficients of similar matrices are represented. A similarity matrix representing the spectral images of tobacco leaves; Represents the trace of a matrix. express The transpose of the matrix; Represents the Laplace matrix; Represents the label prediction matrix. This indicates that the corresponding labels for the tobacco leaves have been marked; This represents the weighting coefficient that controls the intensity of label supervision.

[0061] The specific steps for training the multi-view latent space mapping model are as follows: following an alternating optimization strategy, the mapping matrix, latent space representation, similarity matrix, and label prediction matrix are optimized sequentially. This optimization process is repeated iteratively until the objective function converges or the set number of iterations is reached. Specifically, the optimization of the mapping matrix involves using non-negative matrix factorization to linearly map the latent space representation back to the original view space. The optimization of the latent space representation involves reconstructing the data from each viewpoint using their respective matrices and applying norm regularization to ensure a smooth and stable representation. The optimization of the similarity matrix involves obtaining a similarity matrix reflecting the inherent geometric structure of the data based on the latent space representation and the regularization coefficients of the similarity matrix, using the representations of the two tobacco leaf latent spaces. The optimization of the label prediction matrix involves using the constructed similarity matrix to propagate the labels of labeled tobacco leaves to unlabeled tobacco leaves, resulting in a label prediction matrix for the tobacco leaves.

[0062] The objective function of the optimal solution is the sum of the objective function terms for minimizing reconstruction error, similarity graph structure, and label prediction, specifically:

[0063] ;

[0064] The constraints are: .

[0065] S4: The quality of unlabeled tobacco leaves is assessed based on a multi-view latent space mapping model, and the final quality grade is determined using a classifier, outputting the grading results. Specifically, step S4 involves: determining the quality grade of unlabeled tobacco leaves using a classifier based on the maximum response value of the corresponding label prediction matrix in the multi-view latent space mapping model; using a majority voting mechanism to determine the quality grade of the tobacco leaves; and outputting the grading results. The specific process of the majority voting mechanism for judging tobacco leaf quality is as follows: It is determined whether a predetermined proportion of the grading results in the spectral curve of the tobacco leaves are consistent. If so, the grading result corresponding to that predetermined proportion is output; otherwise, the grading result with the highest prediction confidence among the grading results is output.

[0066] Example 1:

[0067] Taking the four main production areas of Honghe, Chuxiong, Kunming, and Baoshan as an example, and covering four grades (B2F, C2F, C3FA, and C4F), with each grade containing three quality categories (good, medium, and poor), 20 pieces of each category, for a total of 960 samples, the sample list is shown in Table 1.

[0068] Table 1 Classification of Tobacco Leaf Sample Grades

[0069]

[0070] Before image acquisition, all tobacco leaf samples were placed in a constant temperature and humidity environment for 48 hours to equilibrate. The leaves were then unfolded and images were acquired using a 1000–2500 nm band hyperspectral imaging system.

[0071] In the image preprocessing stage, a masking method and a thresholding method are used to extract the effective region to remove invalid information such as leaf stems and background. The extraction effect is as follows: Figure 3 As shown in the figure. After outlier removal, all tobacco leaf samples were used to construct the final dataset for training and validation.

[0072] This method can classify tobacco leaf samples of the same origin and grade, as well as tobacco leaves of different grades from the same origin. First, it focuses on the quality classification of tobacco leaf samples of the same origin and grade, selecting a specific grade from a particular origin (e.g., B2F grade from the Honghe origin) and obtaining the corresponding true quality grade labels for the tobacco leaf samples to establish a tobacco leaf quality grade database. The samples are randomly divided into training and testing sets according to quality categories to train a multi-view latent space graph model. During training, the latent space representation and graph structure relationship are alternately optimized to capture the latent semantics between high-dimensional spectral features. The classifier uses a graph structure label propagation algorithm, driving model generalization with a small number of labeled samples.

[0073] The experiment was conducted in two scenarios: outlier removal and outlier not removal. Ten tobacco leaves were randomly selected from each category as training samples. Each experiment was run 10 times, and the accuracy results were averaged. The experimental results are shown in Table 2. Figure 4 and Figure 5 This is the result of an experiment on the Red River data in Exp02.

[0074] Table 2. Accuracy of tobacco leaf quality judgment in Experiment 1

[0075]

[0076] As can be seen, for quality identification within the same level, this method performs better with global data, achieving an accuracy rate of over 85% across all production areas. Furthermore, the model demonstrates significant performance in identifying medium-quality samples, reflecting the advantages of multi-view latent spatial structure in fine-grained quality identification. In addition, the model exhibits strong robustness to changes in illumination and surface disturbances on tobacco leaves, making it suitable for the rapid assessment of tobacco leaf quality in real-world industrial environments.

[0077] In a joint classification experiment of tobacco leaves of different grades from the same production area, a multi-grade and multi-quality composite classification task was constructed based on the above data. The samples were divided into 12 categories (4 grades × 3 qualities). A semi-supervised learning model was constructed by predicting labels, with only a subset of samples having labels. This model accurately models the intrinsic relationships between complex categories through multi-view feature representation and latent space representation learning. To improve model robustness, some outlier samples were removed, and a class balancing strategy was introduced to avoid training bias. During training, samples of each class were randomly divided into training and validation sets. The experimental results are shown in Table 3. Figure 6 and Figure 7 This is the result of an experiment on the Red River data in Exp03.

[0078] Table 3. Accuracy of tobacco leaf quality judgment in Experiment 2

[0079]

[0080] The classification results show that the proposed method maintains high accuracy in multi-level and multi-quality mixed classification tasks, demonstrating good generalization performance.

[0081] In summary, this method ensures that the final tobacco leaf quality classification results maintain accuracy and stability. It achieves accurate grade identification of a large number of unlabeled tobacco leaf samples while relying on only a small number of labeled samples, significantly reducing manual labeling costs and improving model generalization efficiency. This method is applicable to tobacco leaf sample identification tasks from different production areas and batches, possessing good cross-domain adaptability. It can be embedded in intelligent tobacco leaf identification systems for rapid grading, demonstrating broad engineering application prospects and industrial value.

[0082] A judgment system employing a tobacco leaf quality judgment method based on multi-view latent space mapping includes:

[0083] The image acquisition module is used to acquire near-infrared hyperspectral images of tobacco leaves of different grades and origins within a set wavelength range, and to obtain the real quality grade labels corresponding to the tobacco leaf samples to establish a tobacco leaf quality grade database.

[0084] The spectral image extraction module constructs a mask image by fusing a fixed band threshold and the reflectance difference. After removing the background, leaf stem, leaf veins, and shadows of each tobacco leaf from the mask image to obtain the spectral image of the pure leaf surface area of ​​the tobacco leaf, abnormal samples are removed from the spectral image of the pure leaf surface area of ​​the tobacco leaf.

[0085] The multi-view latent space mapping module divides the spectral image of the pure leaf surface region of the processed tobacco leaf into a set number of complementary views according to dimensions through the feature construction module. Each view corresponds to a different band interval of the spectral image. Normalization is performed on each view to establish a tobacco leaf complementary view dataset. The latent space learning module performs joint non-negative matrix factorization on each view under the guidance of unified latent variables to obtain the latent space representation. The graph structure module constructs a similarity graph structure reflecting the local relationships of samples. The label propagation module performs semi-supervised classification of tobacco leaf labels based on weight coefficient constraints. The multi-view latent space mapping model is trained using the tobacco leaf complementary view dataset to obtain a multi-view latent space mapping model with an optimal solution objective function. The objective function of the multi-view latent space mapping model includes minimizing reconstruction error, similarity graph structure, and multi-view feature representation of label prediction.

[0086] The decision output module judges the quality of unlabeled tobacco leaves based on a multi-view latent space mapping model, determines the final quality grade through a classifier, and outputs the grading results.

[0087] A computer-readable storage medium storing a computer program, characterized in that, when executed by a processor, the computer program causes the device containing the computer-readable storage medium to perform a tobacco leaf quality judgment method based on a multi-view latent space composition.

[0088] like Figure 8 As shown, a computer device is described. This device, which can be a terminal device, includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus. The communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The memory includes a computer-readable storage medium, specifically a non-volatile storage medium. In practical applications, the memory includes the non-volatile storage medium and internal memory. The processor of this computer device provides computing and control capabilities. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface of this computer device is used for exchanging information between the processor and external devices. The communication interface of this computer device is used for wired or wireless communication with external terminals. Wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies.

[0089] When executed by a processor, the computer program implements the automatic tobacco quality identification method described in Embodiment 1. The display unit of the computer device is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an e-ink display screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, a trackball, or a touchpad provided on the computer device casing, or an external keyboard, touchpad, or mouse, etc.

[0090] Those skilled in the art will understand that the structure shown in the figure is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0091] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

Claims

1. A method for judging tobacco leaf quality based on multi-view latent space composition, characterized in that, Includes the following steps: S1: Collect near-infrared hyperspectral images of tobacco leaves of different grades and origins within a set wavelength range, and obtain the real quality grade labels corresponding to the tobacco leaf samples to establish a tobacco leaf quality grade database. S2: A mask image is constructed by fusing a fixed band threshold and the reflectance difference. After removing the background, leaf stem, leaf veins and shadows of each tobacco leaf from the mask image to obtain the spectral image of the pure leaf surface area of ​​the tobacco leaf, abnormal samples are removed from the spectral image of the pure leaf surface area of ​​the tobacco leaf. S3: The spectral image of the pure leaf surface region of the processed tobacco leaf is divided into a set number of complementary views according to the dimensions. Each view corresponds to a different band interval of the spectral image. Normalization is performed on each view to establish a tobacco leaf complementary view dataset. The multi-view latent space mapping model is trained using the tobacco leaf complementary view dataset to obtain a multi-view latent space mapping model with an optimal solution objective function. The objective function of the multi-view latent space mapping model includes minimizing the reconstruction error, similarity graph structure, and multi-view feature representation of label prediction. S4: Based on the multi-view latent space mapping model, the quality of unlabeled tobacco leaves is judged, and the final quality grade is determined by a classifier, and the grading results are output.

2. The method for judging tobacco quality based on multi-view latent space composition according to claim 1, characterized in that, The objective function of the multi-view latent space mapping model in step S3 specifically includes: The objective function term corresponding to minimizing the reconstruction error is expressed as: ； The objective function term corresponding to the similarity graph structure is expressed as: ； The objective function term corresponding to label prediction is expressed as: ； The constraints are as follows: , , , , R represents a real number, r represents the potential spatial dimension of the tobacco leaf, and N represents the number of tobacco leaves. This represents the time-delay characteristic dimension of tobacco leaves. Indicates the number of categories. Indicates the first The original feature matrix corresponding to each view; Indicates the first A mapping matrix for each view; The latent space representation is a low-dimensional representation common to all views; The regularization coefficient represents the reconstruction error; Indicates the first The representation of a tobacco leaf in the latent space. Indicates the first The representation of a tobacco leaf in the latent space; Indicates the first The spectral image of the first tobacco leaf and the second Similarity between spectral images of individual tobacco leaves; The regularization coefficients of similar matrices are represented. A similarity matrix representing the spectral images of tobacco leaves; Represents the trace of a matrix. express The transpose of the matrix; Represents the Laplace matrix; Represents the label prediction matrix. This indicates that the corresponding labels for the tobacco leaves have been marked; This represents the weighting coefficient that controls the intensity of label supervision.

3. The method for judging tobacco quality based on multi-view latent space composition according to claim 2, characterized in that: The specific steps for training the multi-view latent space mapping model in step S3 are as follows: following an alternating optimization strategy, the mapping matrix, latent space representation, similarity matrix, and label prediction matrix are optimized sequentially, and the optimization steps are repeated iteratively until the objective function converges or the set number of iterations is reached.

4. The method for judging tobacco quality based on multi-view latent space composition according to claim 2, characterized in that: The objective function for the optimal solution in step S3 is the sum of the objective function term for minimizing the reconstruction error, the objective function term for the similarity graph structure, and the objective function term for label prediction. Specifically: ； The constraints are: .

5. The method for judging tobacco quality based on multi-view latent space composition according to claim 1, characterized in that: The set quantity in step S3 is 5 or 7.

6. The method for judging tobacco quality based on multi-view latent space composition according to claim 1, characterized in that: The wavelength range set in step S1 is 1000–2500 nm.

7. The method for judging tobacco quality based on multi-view latent space composition according to claim 1, characterized in that: Step S4 specifically involves: determining the quality grade of unlabeled tobacco leaves based on the maximum response value of the corresponding label prediction matrix in the multi-view latent space mapping model, determining the quality grade of the tobacco leaves using a classifier, determining the quality grade of the tobacco leaves using a majority voting mechanism, and outputting the grading results.

8. The method for judging tobacco quality based on multi-view latent space composition according to claim 7, characterized in that: The specific process of the majority voting mechanism in judging the quality of tobacco leaves is as follows: it is determined whether there is a set proportion of consistent grading results in the spectral curve of the tobacco leaves. If so, the grading result corresponding to the set proportion is output; if not, the grading result with the highest prediction confidence among the grading results is output.

9. A judgment system employing the tobacco leaf quality judgment method based on multi-view latent space composition as described in any one of claims 1-8, characterized in that, include: The image acquisition module is used to acquire near-infrared hyperspectral images of tobacco leaves of different grades and origins within a set wavelength range, and to obtain the real quality grade labels corresponding to the tobacco leaf samples to establish a tobacco leaf quality grade database. The spectral image extraction module constructs a mask image by fusing a fixed band threshold and the reflectance difference. After removing the background, leaf stem, leaf veins, and shadows of each tobacco leaf from the mask image to obtain the spectral image of the pure leaf surface area of ​​the tobacco leaf, abnormal samples are removed from the spectral image of the pure leaf surface area of ​​the tobacco leaf. The multi-view latent space mapping module, through the feature construction module, divides the spectral image of the pure leaf surface region of the processed tobacco leaf into a set number of complementary views according to dimensions. Each view corresponds to a different band interval of the spectral image. Normalization processing is performed on each view to establish a complementary view dataset of tobacco leaves. The latent space learning module performs joint nonnegative matrix factorization on each view under the guidance of unified latent variables to obtain the latent space representation. The graph structure module constructs a similarity graph structure that reflects the local relationships of the samples. The label propagation module performs semi-supervised classification of tobacco labels based on weight coefficient constraints. A multi-view latent space mapping model was trained using a tobacco leaf complementary view dataset to obtain a multi-view latent space mapping model with an optimal objective function. The objective function of the multi-view latent space mapping model includes minimizing the reconstruction error, similarity graph structure, and multi-view feature representation of label prediction. The decision output module judges the quality of unlabeled tobacco leaves based on a multi-view latent space mapping model, determines the final quality grade through a classifier, and outputs the grading results.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it causes the device containing the computer-readable storage medium to perform the method of claims 1-8.

11. A computer device, characterized in that: It includes a processor, a memory, an input / output interface, a communication interface, a display unit, and an input device. The processor, memory, and input / output interface are connected via a system bus. The communication interface, display unit, and input device are connected to the system bus via the input / output interface. The memory includes the storage medium described in claim 10.

Images