Tobacco blend uniformity image classification method

Abstract

This invention belongs to the field of tobacco processing technology, specifically relating to an image classification method for the uniformity of tobacco blending. First, a standard for tobacco uniformity levels is established. Then, corresponding uniformity level samples are prepared according to the proportion parameters corresponding to each uniformity level. The samples are mixed uniformly, and images of the samples entering the ground are acquired, forming a multi-level image database covering the entire range of uniformity levels. Based on the YOLOv10 deep learning framework, the model parameters are initialized using ImageNet pre-trained weights. Iterative optimization is performed through dynamic learning rate adjustment and stochastic gradient descent algorithm. Finally, the model weights with the best performance in the validation set are saved, achieving high-precision automated identification of tobacco blending uniformity. This invention uses a high-speed industrial camera to capture images of tobacco piles and improves the YOLOv10 target detection model, achieving high efficiency, high precision, and low cost in tobacco blending uniformity detection.

Description

Technical Field

[0001] This invention belongs to the field of tobacco processing technology, specifically relating to an image classification method for the uniformity of tobacco blending. Background Technology

[0002] Cigarette formulation design is the core foundation determining the quality of tobacco products. Precise analysis of the tobacco components directly impacts the product's physical properties, combustion characteristics, and sensory experience. Different components, such as leaf tobacco, stem tobacco, airflow tobacco, reconstituted tobacco, and expanded tobacco, differ fundamentally in morphology, chemical composition, and functional properties. Therefore, rapidly and accurately determining the proportions of each component in the product is crucial for ensuring consistent cigarette quality, optimizing process parameter matching, improving the precision of sensory comfort control, and promoting intelligent manufacturing upgrades in tobacco products.

[0003] Currently, the mainstream tobacco component ratio determination technologies in the tobacco industry are mainly based on the differences in material properties for differentiation and detection. One method is thermogravimetric analysis (TGA), which utilizes the different activation energies of various tobacco components during thermal decomposition. The sample tobacco is placed in a thermogravimetric analyzer, and its thermal decomposition characteristics are measured under programmed temperature control. By establishing a model corresponding to the percentage of mass loss in different temperature ranges and the content of characteristic substances such as cellulose and hemicellulose in the tobacco components, the proportion of each component can be deduced. Another method uses terahertz spectroscopy, which, based on the different molecular vibrational modes and dielectric response characteristics of various tobacco components, acquires time-domain spectral signals through terahertz beam transmission scanning. Combined with a matching algorithm between the intensity of characteristic absorption peaks and a preset material database, the quantitative inversion of component proportions is achieved. A third method, represented by near-infrared spectroscopy, establishes a chemometric model of the vibrational spectral response of molecular functional groups and the mass fraction of characteristic components such as leaf and stem filaments, achieving non-destructive and rapid detection. In addition, for the specific needs of blending processes, the industry has also developed the CO2 expanded tobacco tracer method, which indirectly evaluates the blending effect of formulated tobacco by detecting the uniformity of tracer distribution. However, existing technologies generally suffer from limitations such as long detection cycles, strong equipment dependence, and high model maintenance costs. Therefore, it is of great significance to find a way to quickly, non-destructively, and accurately identify the components of blended tobacco. Summary of the Invention

[0004] The purpose of this invention is to provide an image classification method for the uniformity of tobacco blending, in order to solve the problems of long detection cycle, strong equipment dependence, and high model maintenance cost of formulated tobacco components, and to improve the speed and accuracy of tobacco component detection.

[0005] To achieve the above objectives, this application employs the following technical solution:

[0006] A method for classifying images of tobacco blending uniformity, comprising the following steps:

[0007] S1. Determine the blending ratio of each tobacco shred in the set cigarette, then adjust the mass ratio of the set components to form several different tobacco shred blending uniformity grades, and establish tobacco shred uniformity grade standards.

[0008] S2. Prepare corresponding uniformity level samples according to the matching parameters corresponding to each uniformity level, mix the samples evenly and collect images of the samples entering the ground to form a multi-level image database covering the entire uniformity level range.

[0009] S3. Based on the YOLOv10 deep learning framework, the image data in the multi-level image database of step S2 is divided into training set, validation set and test set according to a set ratio. The model parameters are initialized with ImageNet pre-trained weights. The model is iteratively optimized by adjusting the dynamic learning rate and the stochastic gradient descent algorithm. Finally, the model weights with the best performance in the validation set are saved to achieve high-precision automated recognition of the uniformity of tobacco blending.

[0010] S4. Based on the trained YOLOv10 model, the model's performance in recognizing the uniformity level of tobacco blending was verified using an independent test dataset.

[0011] Furthermore, in step S1, the tobacco shreds are one or more of the following: leaf shreds, stem shreds, sheet shreds, air-flow shreds, reconstituted tobacco shreds, or expanded tobacco shreds.

[0012] Furthermore, in step S2, the acquired images are preprocessed and labeled to establish a database of tobacco shreds with uniformity levels.

[0013] Furthermore, step S3 also includes introducing RepGFPN and PSAMLCA modules to improve the detection model.

[0014] Furthermore, in step S3, the image data in the multi-level image database is divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0015] Furthermore, in step S3, the localization loss, classification loss, and average accuracy of the validation set are monitored in real time during the training process. If the validation loss does not improve for 10 consecutive rounds, an early stopping mechanism is triggered.

[0016] Furthermore, in step S4, by inputting the preprocessed test set image, the model outputs the prediction level and the corresponding confidence level.

[0017] The beneficial effects of this invention are:

[0018] This invention replaces traditional detection methods with image recognition, enabling rapid detection of the uniformity of tobacco blending. Compared to traditional methods that require significant labor costs and time, this invention uses a high-speed industrial camera to capture images of tobacco piles and improves the YOLOv10 target detection model, achieving high efficiency, high accuracy, and low cost in detecting the uniformity of tobacco blending.

[0019] High efficiency: The RepGFPN module is introduced to merge multi-branch convolutions into a single-path inference structure, reducing computational redundancy and improving real-time detection speed (FPS).

[0020] High precision: The PSAMLCA module enhances the fine-grained feature extraction of tobacco texture and distribution density. Combined with a multi-gradient uniformity level database covering ±1% to ±5% error, the model becomes sensitive to subtle proportion deviations, and mAP@0.5 is improved to 0.95.

[0021] Low cost: Industrial cameras replace specialized equipment such as thermogravimetric analyzers and terahertz spectrometers, significantly reducing hardware investment. The RepGFPN dynamic channel allocation strategy (such as 1 / 8 resolution feature map with only 96 channels) compresses the number of model parameters, adapts to edge device deployment, and reduces reliance on computing power. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0023] Figure 2 This is a schematic diagram of the image acquisition device of the present invention, wherein 1 is a POE power supply line, 2 is a light source power supply line, 3 is an industrial camera, 4 is a Gigabit Ethernet cable, 5 is a lens, 6 is a ring LED light source, 7 is a computer, 8 is a sample, 9 is a camera bracket, and 10 is a power supply.

[0024] Figure 3 A diagram illustrating the comparison of the effects of data augmentation operations.

[0025] Figure 4 This is a diagram of the YOLOv10-RepGFPN-PSAMLCA network structure.

[0026] Figure 5 The image shows the recall rate curve after the model's algorithm was improved.

[0027] Figure 6 This is a graph showing the accuracy of the algorithm after the tobacco processing was improved.

[0028] Figure 7 This is a comparison chart showing the results of the tobacco uniformity grade test. Detailed Implementation

[0029] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings. The following embodiments are merely exemplary and can only be used to explain and illustrate the technical solution of the present invention, and should not be construed as limiting the technical solution of the present invention.

[0030] The terms used in this embodiment are explained as follows:

[0031] 1. Uniformity Grade: This is a quantitative evaluation standard based on the actual error range of the distribution ratio of each group during tobacco blending. Each grade corresponds to a combination of ratio deviations, used to visually reflect the uniformity of the tobacco distribution after blending.

[0032] 2. Data Augmentation: This is a technique that uses algorithms to perform supervised transformations on the original training data to generate diverse derived samples. Its core objective is to expand the size and diversity of the dataset, thereby improving the generalization ability and robustness of machine learning models.

[0033] 3. YOLOv10: This is an object detection model specifically designed to quickly and accurately identify objects in images and determine their location and category. It is one version of the YOLO series, offering faster speed and higher accuracy, making it particularly suitable for industrial inspection scenarios.

[0034] 4. Dynamic label allocation strategy: This is a technique in which the object detection model intelligently adjusts the matching relationship between each predicted box and the real target based on the current predictive ability of the model during the training process.

[0035] 5. mAP: This is a core metric for evaluating the overall performance of a model in object detection tasks, reflecting the balance between accuracy and recall for multi-class object recognition. It is calculated by statistically analyzing detection results at different confidence thresholds and then taking the average area under the precision-recall curves for each class. mAP@0.5 represents the average accuracy with a 50% Intersection over Union (IoU) threshold, while mAP@0.5:0.95 comprehensively evaluates the model's fine-grained recognition capability under strict localization requirements.

[0036] 6. FPS: Represents the number of images that the model can process per second, and is a key parameter for evaluating real-time detection efficiency.

[0037] 7. Recall: Measures the model's ability to cover true positive samples. It is calculated as the proportion of correctly identified positive samples out of all positive samples.

[0038] 8. Precision: This measures the proportion of true positive samples among those predicted as positive by the model. The formula is the percentage of correctly identified positive samples out of all samples predicted as positive.

[0039] 9. TS: Leaf shreds; 10. ETS: Expanded tobacco shreds; 11. RTS: Reconstituted tobacco shreds; 12. CS: Stem shreds.

[0040] 13. RepGFPN module: It is a feature fusion network for target detection. It enhances feature fusion capability by introducing residual structure. It consists of three parts: GF module, PA module and feature cascade.

[0041] 14. PSAMLCA module: This is an improved mechanism introduced into the object detection model. The PSAMLCA module integrates the Hybrid Local Channel Attention (MLCA) mechanism.

[0042] like Figure 1 As shown, this example provides a method for establishing a database of tobacco shreds with uniformity grades through manual weighing and image acquisition, mainly including the following steps:

[0043] In step 1, the specific steps for constructing the benchmark sample and classifying uniformity are as follows:

[0044] S110: Based on the process requirements of the target cigarette product, three representative tobacco blending schemes were selected, and the specific component ratios are shown in Table 1. The values ​​in the table represent the percentage of each component relative to the mass of the tobacco leaves. The tobacco leaves are defined as the baseline component at 100%, and the remaining components, including stems, airflow tobacco, reconstituted tobacco, and expanded tobacco, are dynamically adjusted according to this baseline.

[0045] Table 1. Tobacco blending scheme (based on leaf quality)

[0046]

[0047] Note: The percentage values ​​in the table represent percentages relative to the baseline tobacco leaf mass. For example, in Formula 1, if the tobacco leaf mass is 100 grams, then the stem mass is 18 grams, the airflow tobacco mass is 6 grams, the reconstituted tobacco mass is 10 grams, the expanded tobacco mass is 6 grams, and the total blend mass is 140 grams. Formula 3 does not include airflow tobacco, reflecting the proportion adjustments under specific process conditions.

[0048] S120: Based on the baseline proportions of Formula 1 (100% leaf shreds, 18% stem shreds, 6% airflow shreds, 10% reconstituted tobacco shreds, and 6% expanded tobacco shreds), 24 uniformity grades are generated by systematically introducing errors in single or compound components. Each grade corresponds to a combination of proportion errors, with error ranges covering multiple gradients of ±1%, ±3%, and ±5%, as shown in Table 2.

[0049] Table 2 Uniformity rating scheme for Formula 1 (based on leaf fiber quality)

[0050]

[0051] Note: The percentage values ​​in the table represent percentages relative to the baseline leaf filament quality. For example, 101% leaf filament quality means that the quality of the leaf filament has been improved to 101% of the original baseline value, and 17.0% stem filament quality means that the quality of the stem filament has been reduced to 17% of the original baseline value.

[0052] S130: Formulas 2 and 3 establish uniformity grade schemes in the same way as Formula 1.

[0053] In step 2, the specific steps for constructing and preprocessing the tobacco shred database based on uniformity grades are as follows:

[0054] S210: Based on the preset uniformity grade division rules, select one uniformity grade, and weigh the leaf shreds, stem shreds, airflow shreds, reconstituted tobacco shreds and expanded tobacco shreds according to the mass ratio parameters of each grade, ensuring that the error of a single type during weighing is less than 0.5% and the error of the overall weight is less than 0.3%, and generate a certain number of samples for each grade independently.

[0055] S220: Image acquisition schematic diagram as shown below Figure 2 As shown, after each sample is thoroughly mixed, images are acquired from a top-down angle to ensure the imaging plane is parallel to the sample surface and the lighting conditions are uniform and stable. Three duplicate images are acquired for each sample to generate a feature database covering all levels. The original image resolution is uniformly set to 4000×3000 pixels, and the images are stored in a lossless format to preserve detail information. By fixing the shooting angle and lighting conditions, the interference of external variables on image features is eliminated.

[0056] S230: Repeat steps S210-S220 to obtain multiple sets of tobacco images with different uniformity levels. Use the LabelImg annotation tool to perform YOLO format data annotation, labeling the data as the uniformity level corresponding to the tobacco, and obtain the annotated tobacco uniformity image dataset.

[0057] S240: Adjust all image data to a uniform size (2048×1536). When adjusting the size, the aspect ratio of the images is maintained to avoid image distortion. Carefully check the dataset and remove images with errors or low quality, such as incorrectly labeled or severely blurry images.

[0058] S250: Data augmentation is performed on the removed data. Data augmentation mainly includes three methods: random rotation, random translation, random cropping, adding Gaussian noise, and changing pixel values. A comparison diagram of the augmented image and the original image is shown below. Figure 3 As shown.

[0059] S260: Organize all data and labels to jointly establish a database of tobacco shreds with uniformity grades.

[0060] YOLOv10 is an advanced object detection algorithm that can achieve low-latency, high-precision object detection. This invention makes innovative improvements based on the YOLOv10 model, such as... Figure 4 As shown, by introducing two core modules, RepGFPN (Reparameterized Generalized Feature Pyramid Network) and PSAMLCA (Position-Sensitive Aggregated Hybrid Local Channel Attention), an efficient network architecture adapted to the detection of tobacco blend uniformity is constructed. The improved model significantly enhances the perception accuracy and classification robustness of complex tobacco distribution features while maintaining the original real-time detection capability.

[0061] In step 3, the specific implementation method is as follows:

[0062] S310: The original YOLOv10 PANet structure is replaced with the RepGFPN module. This module is based on the Generalized Feature Pyramid Network (GFPN) and achieves multi-scale feature interaction through an improved Queen-Fusion cross-layer connection mechanism. In the feature fusion stage, a reparameterization technique is used to merge the multi-branch convolutional structure (including 3×3 convolutions, 1×1 convolutions, and skip connections) from the training stage into a single-path structure for the inference stage, preserving feature diversity while reducing computational latency. For fine-grained features of tobacco images, the channel dimensions of RepGFPN are dynamically allocated: the 1 / 8 resolution feature map uses 96 channels to capture local texture, and the 1 / 16 and 1 / 32 feature maps are expanded to 192 and 384 channels respectively to enhance global distribution modeling capabilities.

[0063] S320: The PSA module at the backbone end is replaced with a PSAMLCA module to achieve joint optimization of local channel information and spatially sensitive features. Its core is Hybrid Local Channel Attention (MLCA), which divides the input feature map into a 5×5 local grid. Each grid independently performs channel compression and activation operations, generating a channel weight matrix reflecting the importance of the region. Simultaneously, position-sensitive convolutional layers in the horizontal and vertical directions capture long-range spatial dependencies, outputting a spatial attention map. Finally, a dynamic gating mechanism is used to fuse local channel weights and spatial attention, avoiding the dimensionality compression problem of traditional attention modules.

[0064] S330: Divide the data in the tobacco database of uniformity grades established in S2 into training set, validation set and test set in a ratio of 8:1:1, and ensure that the samples of each uniformity grade are evenly distributed in the subsets through stratified sampling.

[0065] S340: The improved YOLOv10-RepGFPN-PSAMLCA model is trained using the dataset obtained in S260; the model parameters are initialized by loading pre-trained weights, and iterative training is performed using the stochastic gradient descent algorithm combined with a dynamic learning rate adjustment mechanism.

[0066] S350: Based on the initial model trained by S340, evaluate its performance metrics on the validation set, including average precision (mAP@0.5), recall for each class, and inference speed (FPS). To address overfitting issues in the validation results (such as training set precision being significantly higher than validation set precision), an early stopping strategy is adopted (training is terminated if the loss does not improve after 10 consecutive validation rounds). Finally, the model weights with the highest mAP@0.5 on the validation set are saved.

[0067] like Figure 5 and 6 As shown, the recall and precision curves after the algorithm improvement are presented respectively. The recall curve shows that the overall recall rate for all categories is 0.82 at a confidence threshold of 0.5 and remains at 0.72 at a confidence threshold of 0.7, indicating that the model can still stably capture true targets in the high confidence interval, with a significant reduction in the false negative rate. The precision curve shows that the overall precision rate for all categories reaches 1.00 at a confidence level ≥ 0.992, and the precision rate for the tobacco leaf category is as high as 0.98 at a confidence level of 0.8, verifying the high reliability of the model in principal component identification. The overall curve trend is smooth, and the precision rate for all categories is > 0.95 when the confidence level is > 0.7, indicating that the model has both high precision and stability in the high confidence interval. The improved model's mAP@0.5 increased from 0.921 to 0.95, significantly optimizing the overall performance of tobacco blending uniformity detection, and can support efficient interception and accurate classification of minor ratio deviations in industrial quality inspection scenarios.

[0068] In step 4, the specific steps for model validation and performance evaluation are as follows:

[0069] S410: Input the labeled tobacco image into the pre-trained YOLOv10 model. The model outputs the prediction level and confidence score, and the output tobacco image is as follows. Figure 7 As shown;

[0070] S420: Manually verify whether the uniformity level with the highest confidence level is consistent with the uniformity level of the actual ratio. Manually re-verify the top 10% of the detection results (confidence level ≥ 0.9) and record the consistency ratio of the uniformity level judgment. If the manual-model consistency rate of 20 consecutive images is < 90%, the model retraining process is triggered (return to stage S2).

[0071] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the scope defined in the claims.

Claims

1. A method for image classification of tobacco blending uniformity, characterized in that, The following steps are used: S1. Determine the blending ratio of each tobacco shred in the set cigarette, then adjust the mass ratio of the set components to form several different tobacco shred blending uniformity grades, and establish tobacco shred uniformity grade standards. S2. Prepare corresponding uniformity level samples according to the matching parameters corresponding to each uniformity level, mix the samples evenly and collect images of the samples entering the ground to form a multi-level image database covering the entire uniformity level range. S3. Based on the YOLOv10 deep learning framework, the image data in the multi-level image database of step S2 is divided into training set, validation set and test set according to a set ratio. The model parameters are initialized with ImageNet pre-trained weights. The model is iteratively optimized by adjusting the dynamic learning rate and the stochastic gradient descent algorithm. Finally, the model weights with the best performance in the validation set are saved to achieve high-precision automated recognition of the uniformity of tobacco blending. S4. Based on the trained YOLOv10 model, the model's performance in recognizing the uniformity level of tobacco blending was verified using an independent test dataset.

2. The image classification method for tobacco blending uniformity according to claim 1, characterized in that, In step S1, the tobacco shreds are one or more of the following: leaf shreds, stem shreds, sheet shreds, air-flow shreds, reconstituted tobacco shreds, or expanded tobacco shreds.

3. The image classification method for tobacco blending uniformity according to claim 1, characterized in that, In step S2, the acquired images are preprocessed and labeled to establish a database of tobacco shreds with uniformity levels.

4. The image classification method for tobacco blending uniformity according to claim 1, characterized in that, Step S3 also includes introducing RepGFPN and PSAMLCA modules to improve the detection model.

5. The image classification method for tobacco blending uniformity according to claim 1, characterized in that, In step S3, the image data in the multi-level image database is divided into training set, validation set and test set in a ratio of 8:1:

1.

6. The image classification method for tobacco blending uniformity according to claim 1, characterized in that, In step S3, the localization loss, classification loss, and average accuracy of the validation set are monitored in real time during the training process. If the validation loss does not improve for 10 consecutive rounds, the early stopping mechanism is triggered.

7. The image classification method for tobacco blending uniformity according to claim 1, characterized in that, In step S4, the model outputs the prediction level and corresponding confidence level by inputting the preprocessed test set image.

Images