The detection of flavor substance content in tipping paper usually relies on destructive sampling methods. This study aimed to explore a non-destructive detection method for flavor substance content in tipping paper. Hyperspectral images of tipping paper with different flavor substance contents were collected using a hyperspectral imaging system, and an estimation model for flavor substance content in tipping paper was constructed. Considering the characteristics of hyperspectral images of tipping paper, a Reconstructed Regional Spectral Mean (RRSM) method was proposed to extract hyperspectral data. The effects of preprocessing methods such as first derivative (D1), Savitzky-Golay (SG) smoothing filter, and standard normal variate (SNV) transformation, as well as characteristic band selection methods including principal component analysis (PCA), genetic algorithm (GA), and competitive adaptive reweighted sampling (CARS), on the estimation model were investigated. By comparing the performance indicators of the XGBoost model under different combinations of data preprocessing and characteristic band selection methods, the D1-GA-XGBoost model was determined as the optimal estimation model for flavor substance content in tipping paper. For the validation set, the coefficient of determination (R2), root mean square error (RMSE), and mean absolute error (MAE) were 0.9122±0.0290, 0.1645±0.0259 and 0.0788±0.0106, indicating good generalization ability. The XGBoost model based on hyperspectral imaging can achieve accurate detection of flavor substance content in tipping paper and provide technical support for the future realization of online non-destructive detection.

Quantitative testing of spice content in tobacco leaves is crucial for maintaining cigarette quality during manufacturing. While hyperspectral imaging (HSI) offers a non-destructive testing (NDT) alternative to traditional contact-based offline measurements, its redundant high-dimensional data with significant nonlinearity poses a major challenge for accurate spice content testing. To address this, we developed a novel dimensionality reduction algorithm neighborhood structure collaborative preservation embedding (NSCPE) for improving the HSI nondestructive testing system. Lying in the theoretical fusion of manifold learning and collaborative representation, NSCPE constructs a neighborhood structure-aware collaborative representation (NSaCR) model to adaptively capture the precise local correlations of samples within neighborhoods of varying density. And this local structural information is preserved through a dual collaborative graph embedding framework, which maintains both point-to-point and set-to-set neighborhood structures in embedding simultaneously. This adaptive mining and comprehensive preservation mechanism for local structures mitigates data nonlinearity and extracts intrinsic features highly relevant to spice content variations. In experiments, NSCPE achieved superior performance, with the optimal results of R2 = 0.958 and RMSE = 0.349 on LanzhouTob and R2 = 0.861 and RMSE = 0.637 on ShandongTob data sets, outperforming several state-of-the-art methods and showing its potential in real-time quality monitoring in cigarette industry.

To accurately and rapidly detect high-moisture regions in cut tobacco, a detection method for tobacco cut wet clumps based on hyperspectral imaging and characteristic band difference index is proposed. First, the data collected by the hyperspectral detection platform are preprocessed using Moving Average (MA), Standard Normal Variate (SNV), and Min-Max Scaling (MMS). Based on the absorption peak of water molecules near 1230 nm in the short-wave infrared band and the absorption characteristics of the reference band at 1400 nm, a Standardized Water-Wavelength Difference Index (SWIRDI) is constructed to enhance the contrast between high-moisture regions and normalmoisture regions. Then, potential wet clumps are extracted by threshold segmentation, and isolated noise points are removed by morphological filtering. For irregular and scattered detection regions, a bounding box merging algorithm is used to integrate spatially adjacent or overlapping regions, and area-based screening is performed to achieve accurate and complete localization of wet clumps. The results show that the technical approach based on multiple preprocessing, SWIRDI construction, morphological processing, and bounding box merging can efficiently achieve accurate detection of wet clumps from tobacco materials, particularly demonstrating significant advantages in handling irregularly shaped and scattered wet clumps, providing strong technical support for online monitoring and quality control in the tobacco drying process.

At present, the current national standards for China's tobacco industry classify tobacco leaves into 42 grades based on 7 indicators. However, these indicators lack a unified quantitative standard, resulting in prominent problems of strong subjectivity and low efficiency in grading. To solve the issues of difficult image feature extraction caused by the curling and folding of redried tobacco leaves, and limited classification accuracy restricted by small inter-class differences during the sorting and grading process of redried tobacco leaves, this article proposes an intelligent tobacco leaf grading method that integrates multi-attention fine-grained features with color and texture features. Firstly, the RGB/HSI dual-space color statistical strategy, Local Binary Pattern (LBP), and Gray-Level Co-occurrence Matrix (GLCM) are used to extract color and texture information, respectively. Meanwhile, a fine-grained classification network model based on multi-attention sampling is designed, which utilizes the attention mechanism to capture global features and local texture details, making up for the defect that traditional convolutional networks are insufficient in distinguishing subtle features. Finally, the Particle Swarm Optimization (PSO) algorithm is introduced to optimize feature weights, and a multi-model integration framework is constructed by fusing color and texture features. In addition, an image dataset of tobacco leaves at various grades is established. Experimental results show that the fine-grained classification network fusion model designed in this paper achieves recognition accuracies of 84.9% and 79.6% for binary classification and five-class classification, respectively, both of which are superior to those of the ResNet and InceptionV3 models. The intelligent tobacco leaf sorting and grading system developed based on this method realizes fully automatic grading of tobacco leaves and effectively improves production efficiency.

Tobacco leaf disease hole detection is a critical step in ensuring the quality and stability of tobacco leaves within the tobacco industry. However, detecting disease holes during tobacco processing is challenged by interference from visually similar non-disease holes and insufficient feature extraction of small disease hole objects, resulting in detection accuracy that fails to meet the quality control standards required for refined tobacco processing. To address these challenges, this paper proposes a multi-modal Transformer-based detection method for tobacco leaf disease holes, leveraging spectral-spatial fusion. First, a dual-branch feature extraction network combining CNN and Transformer architectures is constructed. The CNN branch accurately captures the texture and edge details of tobacco leaf disease holes in visible light images through a cascaded structure of efficient feature extraction modules, while the Transformer branch explores global feature dependencies in hyperspectral images, enabling deep interactive fusion of multi-scale features. Second, a feature enhancement and fusion network is designed to optimize the distribution heterogeneity and scale differences of cross-modal features, thereby improving the model's semantic understanding in complex scenarios. Additionally, an adaptive deep detection network is developed to efficiently aggregate multi-level contextual information and accurately localize disease holes of varying sizes. Furthermore, a multi-scale adaptive combined loss function is proposed, incorporating a differentiated weighting mechanism and a progressive training strategy to balance the optimization of multi-scale objects and enhance detection performance for small disease hole objects. Experimental results demonstrate that the proposed method significantly improves detection accuracy and robustness, achieving an AP50 of 52.3%, a precision of 62.40%, and an F1-score of 56.43%. Notably, for diseased holes smaller than 2 mm in tobacco leaves, the AP50 reaches 31.16%. Based on the algorithm designed in this work, an online visual inspection system was developed, and test results demonstrate that it maintains stable detection accuracy even under complex scenarios. The method accurately localizes and recognizes tobacco leaf disease holes, providing an efficient solution for disease hole detection in tobacco processing and holding important practical significance for ensuring the quality and stability of cigarette products.

To explore a method for determining leaf drying intensity based on hyperspectral technology and machine learning classification algorithms, this study proposes a Dual Gradient-Successive Projections Algorithm (DG-SPA) for feature band selection. First, local gradients are calculated using second-order difference to capture subtle high-frequency spectral fluctuations. Regional gradients are then calculated using the first derivative after Gaussian weighted smoothing via a dynamic window, extracting overall trend features in the mid-to-low frequencies. Second, single-band dual-gradient feature importance scores are obtained through mean absolute value normalization. Grid search is used to optimize local and regional gradient weights, and a partial least squares (PLS) model combined with an early stopping mechanism determines the optimal initial screening ratio, completing the initial feature band screening. Finally, the SPA algorithm iteratively projects and filters the initially screened features, using the mean square error of cross-validation as an indicator to retain the optimal feature subset, achieving redundant information removal and focusing on key discriminative information. Experimental results show that the proposed model can effectively distinguish leaf filaments with different drying intensities. On the validation set, its accuracy, precision, recall, and Fl score all reach 0.95, significantly outperforming other advanced feature selection methods. This paper achieves stable and efficient indirect discrimination of leaf filament drying intensity by organically combining hyperspectral technology with a dual-gradient-continuous projection algorithm, providing technical support for the digital characterization of leaf filament drying intensity and precise control of drying processes.

To achieve precise localization and counting of abnormal-cut tobacco in complex cigarette manufacturing environments with interference and occlusion, this study proposes an enhanced YOLOv5s-based method using rotated bounding boxes for detection and counting. First, a C3-DEBlock module is developed in the backbone network by integrating the Efficient Multi-scale Attention (EMA) module, the Dynamic Snake Convolution (DSConv), and the C3 structure. This design adaptively adjusts the receptive field to enhance feature extraction capability. Second, a Context Anchor Attention–Bidirectional Feature Pyramid Network (CAA-BiFPN) is incorporated into the neck network. This structure not only reduces computational costs but also captures long-range contextual information, thereby strengthening multi-scale feature fusion. Finally, the Kullback–Leibler divergence (KLD) between the Gaussian distributions is adopted as the regression loss function, enabling dynamic adjustment of parameter gradients based on object characteristics for more accurate bounding box regression. Experimental results demonstrate that the proposed model outperforms mainstream detection models–Faster R-CNN, YOLOv4-tiny, and YOLOv5s–with improvements in mean average precision (mAP) of 14.91, 25.21, and 2.61

To address the issues of complex operation, long time consumption, low efficiency, and difficulty in online monitoring of traditional methods for detecting the proportion of flavoring in tobacco shreds, this study proposes a non-destructive detection method for the proportion of flavoring in tobacco shreds based on near-infrared hyperspectral imaging technology combined with machine learning. In the research, a hyperspectral dataset of tobacco shreds with various flavoring proportions was constructed, and the collected near-infrared hyperspectral data were preprocessed using improved Direct Orthogonal Signal Correction (DOSC) and other algorithms to remove noise interference and enhance signal features. On this basis, multiple models such as Partial Least Squares Regression (PLS), Support Vector Regression (SVR), and Random Forest Regression (RF) were constructed to predict and analyze tobacco shred samples with different flavoring proportions. The results show that the data preprocessed by DOSC have the highest prediction accuracy on the PLS model, especially when the flavoring proportion is 0.6% or higher, with a prediction accuracy of 91.8% and a root mean square error of 0.010. This method provides an effective means for rapid and non-destructive detection of the proportion of flavoring in tobacco shreds and has good industrial application prospects.

The complex industrial process is often characterized by strong multivariate coupling and nonlinear dynamic changes, which pose great challenges to modeling and prediction. Traditional deep learning methods are difficult to effectively capture spatiotemporal characteristics of industrial processes, resulting in poor prediction accuracy. To tackle this issue, we propose a novel end-to-end method named STA-TCN, which utilizes a temporal convolutional network (TCN) with both spatial and temporal attention mechanisms. The TCN uses causal and dilated convolutions to capture long temporal patterns in time series data. The spatial attention identifies the significance of different features, while the temporal attention focuses on crucial time steps. This design assigns adaptive weights to different features and emphasizes key moments to improve the accuracy of dynamic processes. We conduct experiments on two industrial datasets and show that the proposed STA-TCN method achieves significantly improved predictive performance compared to TCN for quality prediction of industrial processes. The results validate the effectiveness and robustness of the proposed method.

The tobacco industry attaches great importance to the development of slim cigarettes, and the content of stem sticks in slim cigarettes is extremely important to the quality of cigarettes. Therefore, in order to solve the problem of difficult detection of stem sticks in cut tobacco, a stem sticks detection algorithm in cut tobacco based on hyperspectral image technology combined with improved YOLOv8n is proposed. First, a principal component analysis method was used to process the hyperspectral image data to improve the differentiation between cut tobacco and stem sticks, and to construct the dataset. Second, the YOLOv8n algorithm was optimized to obtain the GMCM-YOLOv8n algorithm. Multiscale convolutional attention was introduced in the backbone network to capture detail information. Then, ghost convolution (GhostConv) was introduced to replace the regular convolution to simplify the network. M-BiFPN modules are proposed in neck networks as a way to improve the detection of small-sized stem sticks. The C2f module is also improved to obtain P-C2f with a view to reducing the model parameters and computational volume. Finally, the effectiveness of the GMCM-YOLOv8n algorithm is experimentally verified on self-constructed dataset. The results of the experiment showed that: the algorithm achieved a mean average precision of 93.9%, with parameters and floating point operations of 2.2 M and 6.2 G, respectively, and frames per second maintained at 73.5 fps. Compared with YOLOv8n, the proposed improved algorithm exhibited better comprehensive performance, which provided a valuable reference for realizing the task of quickly and accurately detecting the content of stem sticks in cut tobacco in practical production.

Three-dimensional convolutional networks (3DC-NNs) have proven to be powerful tools for hyperspectral image (HSI) classification. However, most existing 3DCNN networks optimize global spectral bands, neglecting the distinct reflective characteristics exhibited in different spectral ranges. To address this limitation, we propose a novel Interactive 3D Group Convolution Network (IGCNet) that incorporates a cross-group interaction strategy. This model explores unique information within different spectral ranges and leverages the high relevance between adjacent groups to enhance the complementarity of information. Additionally, to overcome the constraints of local receptive fields, a non-local fusion attention mechanisms is designed to account for the global correlations of feature map and emphasize important spatial regions during the fusion of group convolutional features, thereby improving the spatial-spectral representation capability of features. Experimental results on two data sets demonstrate the superior performance of IGCNet.

In the tobacco industry, impurity detection is an important prerequisite for ensuring the quality of tobacco. However, in the actual production process, the complex background environment and the variability of impurity shapes can affect the accuracy of impurity detection by tobacco robots, which leads to a decrease in product quality and an increase in health risks. To address this problem, we propose a new online detection method of tobacco impurities for tobacco robot. Firstly, a BCFormer attention mechanism module is designed to effectively mitigate the interference of irrelevant information in the image and improve the network's ability to identify regions of interest. Secondly, a Dual Feature Aggregation (DFA) module is designed and added to Neck to improve the accuracy of tobacco impurities detection by augmenting the fused feature maps with deep semantic and surface location data. Finally, to address the problem that the traditional loss function cannot accurately reflect the distance between two bounding boxes, this paper proposes an optimized loss function to more accurately assess the quality of the bounding boxes. To evaluate the effectiveness of the algorithm, this paper creates a dataset specifically designed to detect tobacco impurities. Experimental results show that the algorithm performs well in identifying tobacco impurities. Our algorithm improved the mAP value by about 3.01% compared to the traditional YOLOX method. The real-time processing efficiency of the model is as high as 41 frames per second, which makes it ideal for automated inspection of tobacco production lines and effectively solves the problem of tobacco impurity detection.

The threshing and redrying process is a crucial stage in cigarette production. During this stage, threshing equipment is used to separate the stems from the initially dried tobacco leaves, followed by the redrying of the leaves. Tobacco stems in the leaves affect various cigarette quality evaluation indices, making stem detection extremely important. In actual production, the methods used are mostly manual screening or winnowing equipment. Existing computer vision-based methods also require offline sampling and manual spreading for detection. The above methods are time-consuming and labor-intensive. Therefore, we propose a tobacco stem detection method based on an improved YOLO model, capable of directly detecting tobacco stems in the mixed materials on the production line without the need for offline sampling. This method, built on the YOLO framework, incorporates the SimAM attention module to make the network focus more on the tobacco stem regions. Additionally, it uses a texture enhancement module to extract concealed edge features of the stems, ultimately achieving a detection accuracy (AP50) of 83.6%. This method can effectively improve the detection efficiency of tobacco stems in the threshing and redrying process.

Accurate prediction of moisture content is significantly crucial for ensuring process stability and product quality in the cylinder drying process. However, the drying process exhibits complex spatio-temporal characteristics and strong interference, which make accurate prediction challenging for the deep learning approach. To address this issue, this article proposes a new spatio-temporal attention-based bidirectional long-short temporal memory network (STA-BiLSTM) model for accurate moisture content prediction. First, Maximum Relevance Minimum Redundancy (mRMR) is adopted to identify optimal features highly related to moisture content. Secondly, bidirectional long-short temporal memory (Bi-LSTM) network is utilized to extract temporal dependencies from the sequential data. Subsequently, spatio-temporal attention mechanisms are designed to adaptively focus on the most relevant features and timesteps, enhancing the model's generalization ability. Finally, due to the harsh industrial environment, eXtreme Gradient Boosting (XGBoost) is adapted to improve generalizability and robustness. Extensive experiments on a real industrial dataset of the drying process demonstrate that the proposed STA-BiLSTM approach significantly outperforms alternative approaches for predicting moisture content, validating its effectiveness and superiority.

In cigarette production, tobacco strip structure directly affects the quality of tobacco in subsequent flue-cured tobacco processes, which in turn is related to the quality of cigarettes and the consumption of tobacco leaves.We propose an improved STDCNet algorithm for real-time segmentation of tobacco leaf images, which has excellent performance in our dataset. The IoU reaches 94.82%, which is 3.61% higher than the baseline algorithm STDCNet; FPS has reached 108.54, meeting the real-time requirements in industrial production.

为解决鲜烟叶成熟度、烘烤质量评价及青烟等级判定过程中存在人为因素对判定结果影响较大等问题,基于高光谱成像技术,通过分析鲜烟叶黄化过程和烤烟正常区域与泛青区域光谱特征差异,采用光谱吸收指数(spectral absorption index,SAI)对烟叶泛青特征进行量化表征.结果表明:①在400～780 nm波段,烟叶泛青区域光谱呈现"两谷一峰"特征,随着泛青程度的减弱,吸收谷深度逐渐减小,烘烤后烤烟正常区域在680 nm左右处的波谷基本消失;②烟叶泛青程度越高,其SAI值越大;③烤烟SAI阈值范围设置为1.19～1.28时,可准确区分正组、微带青组、青黄1级、青黄2级烤烟组别;④设定不同的SAI阈值,结合颜色映射,可实现烟叶不同泛青程度相应的泛青区域位置分布.该方法可为鲜烟叶成熟度评价、烘烤变黄期进程的衡量及烤烟等级的判定提供技术支持.

Due to the susceptibility of deep learning-based object detection algorithms to lighting and image quality in cigarette stem detection application scenarios, which leads to the false detection rate and high hardware occupancy. To address this issue, a CS-YOLOv5 cigarette stem detection model with excellent performance is proposed. First, a novel convolutional neural network (ConvNeXt) is used as the backbone network to improve the ability to extract abstract semantic features. Second, after using selective kernel networks (SKNets) with different kernel sizes in three effective special layers, the receptive field size can be adjusted. Finally, we create a tobacco stem dataset and use it to demonstrate that our improved model is excellent at detecting tobacco stems. The performance of the designed model was evaluated by conducting experiments on the tobacco stems dataset, and it showed excellent results.

为提高梗丝在细支卷烟中的应用比例,采用新型复切设备对梗丝进行复切,将制得的辊切梗丝以不同比例掺配至细支卷烟中,以不掺配辊切梗丝卷烟为对照,对比分析辊切梗丝对卷烟综合质量的影响.结果表明:在试验范围内,①随着辊切梗丝掺配比例增加,烟支总通风率增加,吸阻标准偏差、硬度标准偏差和密度降低;与对照样相比,掺配辊切梗丝烟支物理质量稳定性提高,吸阻标准偏差降低8%以上.②烟支静态燃烧速率、动态燃烧速率、燃烧锥平均特征温度随辊切梗丝掺配比例增加而增加,中心温度范围逐渐扩大;掺配辊切梗丝烟支动态吸阻标准偏差、一致性、落头倾向以及燃烧锥体积低于对照样,燃烧锥体积降幅达到23.4%以上,且当掺配比例为5%时,烟支燃烧锥形态、温度场分布更加均匀,卷烟燃烧过程稳定性更好.③烟支主流烟气各项指标随辊切梗丝掺配比例增加而降低,且均低于对照样,其中TPM量降低0.77 mg/支以上.④刺激性、杂气和余味随掺配比例增加呈现先降低后上升的趋势,当辊切梗丝掺配比例为5%时,烟气协调、蓬松,感官质量较好.适当提高辊切梗丝掺配比例,可以提高细支卷烟物理质量稳定性、燃烧状态和感官质量.

Tobacco strip structure is the size and sheet structure of tobacco leaves in the process of leaf beating, which is the key economic indicator of cigarette manufacturing. In this article, we turn the problem of getting the tobacco strip structure into a classification problem by dividing our photographed images in the complex industrial scenarios into 14 types of tobacco strip structures according to expert knowledge and then using deep learning to recognize them. We use Efficientnetv2 as the backbone and develop it based on the characteristics of our dataset. As for the proposed method, the overall image classification accuracy reaches 92.82%, which is 2.33% higher than the accuracy before our improvements. 14 types of tobacco strip structures after leaf beating can be identified quickly and efficiently by using our present model. The results of this study can provide a theoretical basis and technical support for the intelligent identification of tobacco strip structure after leaf beating.