Method for identifying structure of threshed tobacco strip in threshing and redrying process

Abstract

The invention relates to a method for identifying a tobacco strip structure after threshing in a threshing and redrying process. The method comprises the following steps: collecting picture data of the tobacco strip structure according to a mode of simulating a real scene; preprocessing the picture data, and dividing the picture data into a training set, a test set and a verification set; training, testing and verifying a preset model; and carrying out actual detection to obtain a prediction result of the flue-cured tobacco lamina structure on the real-time production line. The method has the beneficial effects that the traditional link of discretizing the tobacco lamina is avoided, and three links of discretizing the stacked tobacco lamina, identifying the area of a single tobacco lamina and finally calculating the tobacco lamina structure in the stack are combined into a link of directly predicting the tobacco lamina structure, so that the tobacco lamina structure in the stack can be directly predicted; according to the method, the structure of the stacked tobacco lamina is directly predicted from surface photos of the stacked tobacco lamina passing through an actual production line conveying belt through the trained improved EfficientNetv2-s model, and the method has instantaneity and certain representativeness, can quickly and timely reflect the field production state and is beneficial to guiding subsequent production.

Description

Technical Field

[0001] The invention relates to the technical field of cigarette processing and manufacturing, and in particular to a method for identifying the structure of threshed tobacco leaves in a threshing and re-baking process. Background Art

[0002] The structure of tobacco leaves refers to the proportion of large, medium, small and broken leaves after threshing. It is an important component of tobacco leaves quality index and an important indicator to measure the level of threshing and re-baking. At present, in the process of cigarette production, the main requirements for tobacco leaves structure are to control the large-leaf ratio, steadily increase the medium-leaf ratio, and reduce the small-leaf ratio.

[0003] Therefore, the essence of obtaining the structure of tobacco leaves in the tobacco industry lies in how to identify and calculate the proportion of large, medium, small and broken leaves in a pile of tobacco leaves.

[0004] Traditional industrial methods include direct measurement and indirect measurement. In the direct measurement method, during the actual production process, samples are taken from the tobacco pile on the production line manually. Generally, 3000±300g of tobacco is sampled, and then the area of ​​each tobacco piece is manually identified and calculated using the graph paper method, punch weighing method, etc. Finally, the tobacco piece structure is calculated according to the definition. The entire identification and calculation process is completely completed by manual operation, which is time-consuming and labor-intensive, but the calculation result is relatively accurate.

[0005] Compared with the direct measurement method, the indirect measurement method has some progress. In the actual production process, samples are taken from the tobacco leaves on the production line. Generally, 3000±300g of tobacco leaves are sampled and evenly dispersed on the surface of the conveyor belt. The quality control vibrating screen is started to screen out large, medium, small and broken leaves respectively. This is the identification process. Then, the tobacco leaves collected by each layer of the screen are used to obtain the tobacco leaves area using a leaf area meter. The tobacco leaves structure is calculated according to the definition. This is the calculation process. The advantages of the indirect measurement method are fast and highly automated, but there are the following disadvantages: less tobacco leaves can be processed each time, and the tobacco leaves need to be in a separate state. It takes a long time to calculate the area of ​​tobacco leaves, which is inconsistent with the state of a large number of overlapping and rapid passage of tobacco leaves on the production line in a short period of time. The measured results cannot timely and accurately reflect the current production status and guide subsequent production. The leaf area meter is expensive and has high maintenance costs.

[0006] In summary, both direct and indirect measurement methods are highly accurate, but the disadvantages are that they are time-consuming and labor-intensive, require the production line to be suspended, and are only suitable for small batches of leaf area measurement, not for large-scale industrial production needs, because it takes 30-40 minutes to get the results, and during this time, the weight of the tobacco leaves passing through the conveyor belt of the production line can reach more than 5,000 kg, far exceeding the weight of the sample. Therefore, the tobacco leaf structure obtained by the traditional direct and indirect measurement methods has serious lag and lack of representativeness, and cannot quickly and timely reflect the on-site production status and guide subsequent production.

[0007] With the development of machine vision, image processing technology has gradually become a hot topic in the field of blade area calculation research. The use of image processing to obtain blade area has the advantages of fast speed, low equipment cost, and low labor cost. However, the current machine vision method is: in the actual production process, it is still necessary to manually take samples from the tobacco sheet pile on the production line, and then separate the tobacco sheets one by one through the differential belt, and then use machine vision technology to quickly identify the area of ​​each tobacco sheet and complete the task of calculating the tobacco sheet structure according to the definition.

[0008] It can be seen that the current machine vision solution is only a cheaper alternative to the function of the traditional leaf area meter and is faster and more accurate than the leaf area meter. However, it still requires manual sampling from the tobacco leaf pile on the production line and then separating the tobacco leaves one by one. These steps affect production continuity and require manual participation or other machine assistance. It is not possible to directly identify and calculate the tobacco leaf structure from the tobacco leaf pile on the fast-passing production line conveyor belt. In other words, there are always serious lags and insufficient representativeness, and it is impossible to quickly and timely respond to the on-site production status and guide subsequent production.

[0009] In summary, whether it is direct measurement, indirect measurement or current machine vision method, the process of obtaining the structure of tobacco sheets is essentially divided into three steps:

[0010] Step 1: Discretize the stacked cigarette pieces and separate them into individual cigarette pieces;

[0011] Step 2: Identify the area of ​​a single cigarette;

[0012] Step 3: Calculate the smoke sheet structure according to the definition.

[0013] Obviously, step one is the fundamental reason why the current method always has serious lags, lacks representativeness, and is unable to quickly and timely reflect the on-site production status and guide subsequent production. It is also an unavoidable link in the current method. Summary of the invention

[0014] In order to overcome the problems existing in the background technology: whether it is the direct measurement method, the indirect measurement method or the current machine vision method, manual sampling is required and the cigarette sheets in the stack need to be discretized to separate individual cigarette sheets, and then the area of ​​the single cigarette sheet is identified, so as to finally calculate the structure of the cigarette sheets in the stack. The discretization of the cigarette sheets causes the current method to have serious lag and lack of representativeness, and it is unable to quickly and timely reflect the on-site production status and guide subsequent production.

[0015] The present invention proposes a method for identifying the structure of tobacco leaves after threshing during the threshing and re-baking process, comprising: setting up an image acquisition and analysis device on a production line, and also comprising:

[0016] Collect image data of the cigarette structure in a way that simulates a real scene;

[0017] The tobacco strip structure is a structural type formed by the specific mass proportions of broken leaves, small leaves, medium leaves and large leaves in the stacked tobacco strips after the threshing and re-baking process;

[0018] Preprocessing the image data into a training set, a test set, and a validation set;

[0019] Use the training set to train the preset model, use the test set to test the preset model, and use the validation set to verify the preset model;

[0020] The preset model after the verification is completed is used to perform actual testing to obtain the prediction result of the tobacco sheet structure on the real-time production line.

[0021] Preferably, the image acquisition and analysis device includes an area array camera, a light source system, structural parts, a data processing device, an acquisition card, a data transmission line, and a high-performance processor.

[0022] Preferably, the collecting of image data of the cigarette sheet structure in a manner simulating a real scene comprises:

[0023] Take samples from the production line from the stack of tobacco leaves after threshing and redrying;

[0024] Separating broken leaves, small leaves, middle leaves and large leaves from the stack of sampled tobacco leaves;

[0025] Manually mixing and placing the separated broken leaves, small leaves, medium leaves and large leaves into N new piles according to a specific tobacco leaf structure, wherein the size of the new piles is the same as the size of the piles photographed on the actual production line;

[0026] The new stack includes all types of sheet tobacco structures;

[0027] Each new stack is repeated M times and the stack is shuffled using a toggle tumbling device. Each time the stack is shuffled, an area array camera is used to photograph the stack to form a data set of N*M images.

[0028] Preferably, the preprocessing of the image data into a training set, a test set and a validation set comprises:

[0029] Processing the image data using median filtering;

[0030] Enhance and expand the image data after the median filtering process by using flipping, rotation and translation scaling;

[0031] The enhanced and expanded image dataset is annotated with respect to the cigarette structure type.

[0032] The divided image data set is divided into a training set, a test set and a validation set in proportion.

[0033] Preferably, the preset model includes:

[0034] The preset model is the improved EfficientNetv2-s model;

[0035] The improved EfficientNetv2-s model includes adding TEM module and CBAM module to the traditional EfficientNetv2-s model

[0036] The beneficial technical effect of the present invention is that it completely avoids the link of "discretizing the cigarette sheets" which is unavoidable in traditional direct measurement, indirect measurement or current machine vision methods, and instead merges the three links of "discretizing the cigarette sheets in the stack, identifying the area of ​​a single cigarette sheet, and finally calculating the structure of the cigarette sheets in the stack" required for traditional calculation of the cigarette sheet structure into one link of "directly predicting the structure of the cigarette sheets", and uses the improved EfficientNetv2-s model to instantly predict the structure of the cigarette sheets in the stack from the surface photos of the cigarette sheet stack passing through the conveyor belt of the actual production line, and the improved EfficientNetv2-s model is more suitable for the actual production situation involved in the present invention than the traditional model, and has a better overall prediction effect. Therefore, the technical solution of the present invention is immediate and representative, can quickly and timely reflect the on-site production status, and is conducive to guiding subsequent production. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0038] Figure 1 Yes; the images collected under different tobacco sheet structures of the present invention, wherein the upper left corner image is tobacco sheet structure 9, the upper right corner image is tobacco sheet structure 3, the lower left corner image is tobacco sheet structure 6, and the lower right corner image is tobacco sheet structure 11;

[0039] Figure 2Yes; The structure diagram of the MBConv module and the Fused-MBConv module in the traditional EfficientNetv2-s network model;

[0040] Figure 3 It is: the structural diagram of the TEM module and the CBAM module in the improved EfficientNetv2-s network model of the present invention;

[0041] Figure 4 is: a flow chart of the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] Additional aspects and advantages of the present invention will be given in part in the following description, which will become obvious from the following description, or may be learned through practice of the present invention.

[0044] like Figure 1-4 As shown, in order to solve the problems existing in the background technology: whether it is the direct measurement method, the indirect measurement method or the current machine vision method, manual sampling is required and the cigarette sheets in the stack need to be discretized to separate individual cigarette sheets, and then the area of ​​the single cigarette sheet is identified, so as to finally calculate the structure of the cigarette sheets in the stack. The step of discretizing the cigarette sheets causes the current method to have serious lags and insufficient representativeness, and it is unable to quickly and timely reflect the on-site production status and guide subsequent production.

[0045] A method for identifying the structure of tobacco leaves after threshing in a threshing and re-roasting process, comprising: setting up an image acquisition and analysis device on a production line; collecting image data of the tobacco leaf structure in a manner simulating a real scene; the tobacco leaf structure is a structure type formed by the specific mass proportions of broken leaves, small leaves, medium leaves and large leaves in the stack of tobacco leaves after threshing and re-roasting; pre-processing the image data to divide it into a training set, a test set and a verification set; training a preset model with the training set, testing the preset model with the test set, and verifying the preset model with the verification set; and performing actual detection using the preset model after the verification to obtain a prediction result of the tobacco leaf structure on the real-time production line.

[0046] Specifically, the image acquisition and analysis device includes an array camera, a light source system, a structural component, a data processing device, an acquisition card, a data transmission line, and a high-performance processor. These devices ensure that the production line can be imaged in real time, and the acquired images can be processed in real time by the data processing device, and the processed results can be immediately fed back to the operator.

[0047] Specifically, from the perspective of production practice, although it is possible to directly collect images of the stacks of tobacco leaves actually produced on the production line, it is difficult to accurately sample the stacks within the image range for analysis and annotation of the tobacco leaf structure types, and these stacks may not necessarily contain all the current tobacco leaf structure types. This leads to a lack of sufficient and comprehensive data sets for subsequent model training and testing.

[0048] Therefore, the present invention adopts a method of collecting image data of the tobacco sheet structure in a manner of simulating a real scene, specifically including:

[0049] Sampling the tobacco leaves stacked after the threshing and re-baking process on the production line. This sampling method ensures that the tobacco leaves in the image data are from the actual tobacco leaves stacked on the real production line;

[0050] Then, a vibration screening method is adopted to separate broken leaves, small leaves, medium leaves and large leaves from the stack of sampled tobacco leaves, wherein the vibration screening method comprises: placing tobacco leaves on screens with different apertures, vibrating forward, and letting small-sized leaves fall through for further screening, and screening out tobacco leaves of different sizes in each layer;

[0051] Then, the separated broken leaves, small leaves, medium leaves and large leaves are manually mixed and placed into N new stacks according to a specific tobacco leaf structure. The size of the new stack is the same as the size of the stack photographed on the actual production line, which ensures that the size of the photo of the stack taken by the area array camera is consistent with the actual production line, ensuring the simulation of the real scene;

[0052] In order to ensure that each type of tobacco leaf structure has sufficient data sets for training and testing, the new stack includes all tobacco leaf structure types. Specifically, the present invention involves 14 tobacco leaf structure types, that is, N=14, as shown in Table 1;

[0053] Table 1 Ratios of large, medium, small and broken leaf rates under 14 types of tobacco leaf structures

[0054]

[0055]

[0056] Depend on Figure 1As shown, the upper left image is the tobacco structure 9, the upper right image is the tobacco structure 3, the lower left image is the tobacco structure 6, and the lower right image is the tobacco structure 11. It can be seen that different tobacco structure types contain different large, medium, small, and broken leaf rate ratios, and also show very different image textures, shapes, tobacco density, etc. at the same shooting size;

[0057] After all types of stacks are laid out, each new stack is repeated M times and the stack is shuffled using a toggle tumbling device. Each time the stack is shuffled, an area array camera is used to photograph the stack to form a data set of N*M images.

[0058] Specifically, after completing the image data collection, the image data is preprocessed and divided into a training set, a test set, and a validation set, including:

[0059] Depend on Figure 1 As shown in the figure, in the actual collected cigarette dataset, the cigarette almost occupies the entire image, and only a small amount of background is exposed in the image, so the background segmentation of the collected image is not performed. Before the training starts, the data is preprocessed using median filtering to filter out the salt and pepper noise while retaining most of the edge information of the image. Then, data enhancement is performed using flipping, rotation, and translation scaling to expand the dataset, so that the expanded dataset can also be used, and the work of on-site sampling and collection is reduced, which improves efficiency.

[0060] In this embodiment, the cigarette sheet image dataset has a total of 13,450 images, among which the number of images in the training set, test set and validation set is 8067:2656:2818, that is, the dataset is roughly divided according to 6:2:2, and the training set, test set and validation set contain images of the 14 types of cigarette sheet structures.

[0061] The model used in the present invention is an improved EfficientNetv2-s network model. The traditional EfficientNetv2-s network structure is shown in Table 2:

[0062] Table 2 Network structure of EfficientNetv2-s

[0063]

[0064] The EfficientNetv2-s network model is composed of various stages connected in numerical order, and the number is the number of corresponding modules. For example, the first two rows in Table 2 show that the first few layers of the EfficientNetv2-s network consist of a Conv3×3 module followed by 2

[0065] The main modules of EfficientNetv2-s are MBConv (Mobile Inverted Bottleneck Convolution) and Fused-MBConv. MBConv is a module in MobileNetV3, but EfficientNetv2-s adds SE (Squeeze and Excitation) module to it, and the activation function is changed from ReLu to Swish activation function proposed by Google. The structure diagram of MBConv module in EfficientNetv2-s is as follows: Figure 2 Shown on the left.

[0066] Fused-MBConv is an improved MBConv module, which is mainly used for server acceleration. It replaces the Expansion Conv1×1 (written as Conv1×1 in the structure diagram) and Deepwise Conv3×3 in MBConv with a common Conv3×3. Experiments on EfficientNetB4 show that the use of Fused-MBConv can significantly improve the training speed while only adding a few network parameters and FLOPs overhead, while ensuring training accuracy. The structure diagram of Fused-MBConv in EfficientNetv2-s is shown in the figure below. Figure 2 Shown right.

[0067] The improved EfficientNetv2-s model includes adding a TEM module and a CBAM module to the traditional EfficientNetv2-s model.

[0068] More specifically, since the structure of tobacco leaves is determined by the rate of large, medium, small, and broken leaves, different structures of tobacco leaves will result in different areas occupied by large, medium, small, and broken leaves on the tobacco leaf image, and the distribution of tobacco leaves of different sizes on the image constitutes the texture of the image. The TEM (Texture Enhanced Module) module is a texture enhancement module. Its main idea is that in the human visual system, a group of different-sized group receptive fields is conducive to enhancing the area where the retina is sensitive to small spatial changes, thereby learning more spatial and scale-related information. Considering that tobacco leaves of different sizes and their overlapping and adjacent positions constitute the scale and spatial information of the image, and broken leaves account for a small proportion of tobacco leaves and are very small in scale, they can be ignored compared with other leaves. The receptive field branches of TEM are exactly three, which makes it easier to extract the features of large leaves, medium leaves, and small leaves of different sizes, which is consistent with the actual situation. Therefore, TEM is used to extract more discriminating features from tobacco leaf images. The structure diagram of the TEM module is shown below. Figure 3 As shown above.

[0069] CBAM (Convolution Block Attention Module) is a lightweight attention module, and its main submodules are the spatial attention module SAM (Spartial Attention Moudel) and the channel attention module CAM (Channel Attention Moudel). For the smoke scene, the stacking in the smoke image is relatively serious, and the image scene is relatively complex. Only using the traditional EfficientNetv2-s cannot learn some key features well. The attention module can make the network pay more attention to the main features, learn more key information, improve the network's discrimination accuracy, and save network parameters and computing power. Therefore, the CBAM module is added to improve the network classification performance. The CBAM module is as follows: Figure 3 As shown below.

[0070] Therefore, the network structure of the improved EfficientNetv2-s model proposed in the present invention is shown in Table 3:

[0071] Table 3. Improved network structure of EfficientNetv2-s

[0072]

[0073] In order to evaluate the improved EfficientNetv2-s model in the present invention, the present invention adopts accuracy as a relevant evaluation indicator.

[0074] The accuracy of image classification is defined as P. The calculation method of P is shown in Formula 1.

[0075]

[0076] Among them, for the accuracy rate P of the sheet structure type x, Sum is the total number of images of this sheet structure, and TP is the number of images of this sheet structure correctly classified as type x by the network. For the overall accuracy rate P, Sum is the total number of images of all sheet structure types, and TP is the total number of images correctly classified by the network.

[0077] Specifically, first build a Python-based deep learning environment. Hardware configuration: CPU is Intel(R) Core(TM) i9-10900K CPU@3.70GHz 3.70GHz, memory is 32G, and graphics card is NVIDIA GeForce RTX 3090. The deep learning environment is Python3.6 version, PyTorch version is 1.10.2, and CUDA version is 11.3. In this embodiment, the image is randomly cropped and scaled to 334*334 size and then sent to the improved EfficientNetv2-s model for training. The epoch is 30, the batch size is 32, and the initial learning rate lr and the cycle learning rate lrf are 0.01.

[0078] More specifically, the improved EfficientNetv2-s is used to classify the sheet cigarette structure on the test set, and the classification results are shown in Table 4. As can be seen from Table 4, for a single image, the improved EfficientNetv2-s network can achieve a classification accuracy of 100% on 3 types of sheet cigarette structure images, and a classification accuracy of 89% or more on 13 types of sheet cigarette structure images except sheet cigarette structure 9. The overall image classification accuracy can reach 92.82%, and the classification effect on 14 types of sheet cigarette structures is good, which can achieve fast, efficient and accurate sheet cigarette structure classification, and can be applied to actual production needs. The sheet cigarette structure can be output in real time during the sheet cigarette production process to provide guidance and reference for subsequent production processes.

[0079] Table 4 Performance of the improved EfficientNetv2-s on the test set

[0080]

[0081] More specifically, an ablation experiment is conducted on the improvements made in this paper to verify the effectiveness of the model improvement. The ablation experiment results are shown in Table 5.

[0082] Table 5 Related ablation experiments

[0083]

[0084] It can be seen from the second and third columns of Table 5 that although the addition of the CBAM module slightly reduces the overall classification accuracy of the network, the classification accuracy of the sheet cigarette structure 9, where the original network classification effect is poor, is significantly improved. The addition of the CBAM module improves the robustness of the network. It can be seen from the second and fourth columns of Table 5 that the addition of the TEM module, although the classification accuracy is slightly reduced under some sheet cigarette structures, effectively improves the classification accuracy on sheet cigarette structures 8 and 9, and also improves the overall network classification performance by 1.63%. The addition of this module is effective in improving the network classification performance. It can be seen from Table 5 that the addition of the CBAM module and the TEM module at the same time can achieve the optimal image classification of 6 sheet cigarette structures and the suboptimal image classification of 4 sheet cigarette structures under 14 sheet cigarette structures, and the overall image classification accuracy reaches the optimal. At the same time, for sheet cigarette structure 8, a 23.93% accuracy improvement can be achieved.

[0085] To sum up, the technical solution of the present invention completely avoids the step of "discretizing the cigarette sheets" which is unavoidable in traditional direct measurement methods, indirect measurement methods and current machine vision methods, and instead combines the three steps of "discretizing the cigarette sheets in the stack, identifying the area of ​​a single cigarette sheet, and ultimately calculating the structure of the cigarette sheets in the stack" required for traditional calculation of the cigarette sheet structure into one step of "directly predicting the structure of the cigarette sheets".

[0086] In the link of "directly predicting the structure of tobacco sheets", the present invention chooses to add TEM modules and CBAM modules to the traditional EfficientNetv2-s model according to the characteristics of tobacco sheet stacking in actual production, wherein the TEM module enhances the texture characteristics of the three types of tobacco sheets, large, medium and small, in the tobacco sheet structure, and the CBAM module better handles the complex scene problems caused by the stacking of tobacco sheets in the actual production line. In the evaluation experiment, the improved EfficientNetv2-s model has a better overall effect than the traditional EfficientNetv2-s model. Therefore, the improved EfficientNetv2-s model proposed by the present invention is more suitable for the state of tobacco sheet stacking in actual production, and achieves better prediction results.

[0087] Finally, the structure of the stacked tobacco leaves is predicted directly from the surface photos of the stacked tobacco leaves passing through the conveyor belt of the actual production line through the trained improved EfficientNetv2-s model. The invention is instant and representative, can quickly and timely reflect the on-site production status, and is conducive to guiding subsequent production.

[0088] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art in the art to which the present invention belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless defined as herein.

[0089] Those skilled in the art can understand that the components in the device in the embodiment can be distributed in the device in the embodiment according to the description of the embodiment, or can be changed accordingly and located in one or more devices different from the embodiment. The components in the above embodiment can be combined into one component, or can be further divided into multiple sub-components.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate rather than limit the technical solutions of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the present invention can still be modified or replaced by equivalents. Any modification or partial replacement that does not depart from the spirit and scope of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A method for identifying the structure of tobacco leaves after threshing and re-roasting, comprising setting up an image acquisition and analysis device on a production line, characterized in that: Also includes: Collect image data of the cigarette structure in a way that simulates a real scene; The tobacco strip structure is a structural type formed by the specific mass proportions of broken leaves, small leaves, medium leaves and large leaves in the stacked tobacco strips after the threshing and re-baking process; Preprocessing the image data into a training set, a test set, and a validation set; Use the training set to train the preset model, use the test set to test the preset model, and use the validation set to verify the preset model; The preset model after the verification is completed is used to perform actual testing to obtain the prediction result of the tobacco sheet structure on the real-time production line.

2. The method for identifying tobacco leaf structure after threshing and re-roasting according to claim 1, characterized in that: The image acquisition and analysis device comprises an area array camera, a light source system, structural parts, a data processing device, an acquisition card, a data transmission line, and a high-performance processor.

3. The method for identifying tobacco leaf structure after threshing and re-roasting according to claim 1, characterized in that: The collecting of image data of the cigarette sheet structure in a manner simulating a real scene includes: Take samples from the production line from the stack of tobacco leaves after threshing and redrying; Separating broken leaves, small leaves, middle leaves and large leaves from the stack of sampled tobacco leaves; Manually mixing and placing the separated broken leaves, small leaves, medium leaves and large leaves into N new piles according to a specific tobacco leaf structure, wherein the size of the new piles is the same as the size of the piles photographed on the actual production line; The new stack includes all types of sheet tobacco structures; Each new stack is repeated M times and the stack is shuffled using a toggle tumbling device. Each time the stack is shuffled, an area array camera is used to photograph the stack to form a data set of N*M images.

4. The method for identifying tobacco leaf structure after threshing and re-roasting according to claim 1, characterized in that: The preprocessing of the image data into a training set, a test set and a validation set includes: Processing the image data using median filtering; Enhance and expand the image data after the median filtering process by using flipping, rotation and translation scaling; The enhanced and expanded image dataset is annotated with respect to the cigarette structure type. The divided image data set is divided into a training set, a test set and a validation set in proportion.

5. The method for identifying tobacco leaf structure after threshing and re-roasting according to claim 1, characterized in that: The preset models include: The preset model is the improved EfficientNetv2-s model; The improved EfficientNetv2-s model includes adding a TEM module and a CBAM module to the traditional EfficientNetv2-s model.

Images