Intelligent identification method and system for maturity level of cigar wrapper tobacco leaves

Abstract

Disclosed in the present invention are an intelligent identification method and system for the maturity level of cigar wrapper tobacco leaves. The method comprises the following steps: acquiring cigar tobacco leaf images, and performing classification processing thereon; by adjusting the number and size of convolutional blocks and residual blocks, improving a ResNet_18 model so as to construct a residual identification model, and using the residual identification model as an identification model for the maturity level of cigar tobacco leaves; determining a mini-batch size and an initial learn rate of the residual identification model; on the basis of the cigar tobacco leaf images that have been subjected to classification processing, training the residual identification model; and on the basis of the trained residual identification model, identifying the maturity level of cigar tobacco leaves. The intelligent identification method for the maturity level of cigar wrapper tobacco leaves provided in the present invention can achieve the rapid and accurate identification of the maturity level of wrapper tobacco leaves, and can overcome the defect of manual identification being highly subjective.

Description

A method and system for intelligently identifying the maturity of cigar wrapper tobacco leaves Technical Field

[0001] The present invention belongs to the technical field of tobacco leaf planting, and in particular relates to a method and system for intelligently identifying the maturity of cigar wrapper tobacco leaves. Background Art

[0002] Maturity is the state of maturity exhibited by tobacco leaves during their field growth and development, and is a key factor in determining tobacco leaf quality. Currently, cigar tobacco maturity is often assessed visually based on the number of days since transplanting and leaf color. Compared to flue-cured tobacco, cigar tobacco leaves exhibit less color variation across different parts, and visual assessment is susceptible to environmental influences, leading to significant subjectivity and errors.

[0003] With the rapid development of machine vision technology in recent years, combined with image processing and machine learning, machine vision has provided a scientific method for determining tobacco leaf maturity. While extensive research has been conducted on using images to identify the maturity of flue-cured tobacco, little research has been conducted on cigar tobacco leaves. Cigar and flue-cured tobacco leaves differ significantly in terms of variety and cultivation methods.

[0004] In view of this, it is urgent to propose an intelligent identification method for the maturity of cigar tobacco leaves. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention proposes a method and system for intelligently identifying the maturity of cigar wrapper tobacco leaves to solve the problems existing in the above-mentioned prior art.

[0006] To achieve the above object, the present invention provides a method for intelligently identifying the maturity of cigar wrapper tobacco leaves, comprising the following steps:

[0007] Obtain cigar leaf images and perform classification processing;

[0008] The ResNet_18 model was improved by adjusting the number and size of convolutional blocks and residual blocks to construct a residual recognition model, which was used as a recognition model for cigar tobacco leaf maturity.

[0009] Determining a mini-batch size and an initial learning rate for the residual recognition model;

[0010] Training the residual recognition model based on the classified cigar leaf images;

[0011] The maturity of cigar tobacco leaves is identified based on the trained residual recognition model.

[0012] Compared with the prior art, the present invention has the following advantages and technical effects:

[0013] The present invention uses an improved ResNet_18 model to automatically extract maturity recognition features from images of fresh wrapper tobacco leaves in the field, overcoming the shortcomings of manual feature extraction. It further builds a recognition system to achieve model visualization effects, and can realize rapid and accurate recognition of the maturity of wrapper tobacco leaves. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0015] FIG1 is a schematic diagram of the recognition accuracy of the AlexNet classic structure, optimized structure 1, and optimized structure 2 for test set samples according to an embodiment of the present invention;

[0016] FIG2 is a schematic diagram of the discrimination accuracy of AlexNet with different MiniBatchSizes on a test set according to an embodiment of the present invention;

[0017] FIG3 is a schematic diagram of the discrimination accuracy of AlexNet with different InitialLearnRates on a test set according to an embodiment of the present invention;

[0018] FIG4 is a confusion matrix diagram of the improved AlexNet test set according to an embodiment of the present invention;

[0019] FIG5 is a schematic diagram of the accuracy of the training process of ResNet_18 with different structures according to an embodiment of the present invention;

[0020] FIG6 is a schematic diagram of the training accuracy of ResNet-18 with different MiniBatchSizes according to an embodiment of the present invention;

[0021] FIG7 is a schematic diagram of the training accuracy of ResNet-18 at different InitialLearnRates according to an embodiment of the present invention;

[0022] FIG8 is a confusion matrix diagram of the improved ResNet_18 test set according to an embodiment of the present invention. DETAILED DESCRIPTION

[0023] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0024] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0025] Example 1

[0026] This embodiment provides a method for intelligently identifying the maturity of cigar wrapper tobacco leaves, comprising the following steps:

[0027] Materials and Methods

[0028] Test site and materials:

[0029] The experiment was conducted in Sujiaqiao Village, Shifang City, Deyang City, Sichuan Province (31.20°N, 104.08°W) and Yantan Village, Fengcheng Town, Dazhou City, Sichuan Province (31.54°N, 108.01°W). The experimental varieties were Dexue No. 1 and Chuanxue No. 1, and the tobacco leaf image acquisition device was a low-cost industrial camera sensor, model MVL-MF0828M-8MP.

[0030] Experimental Design:

[0031] The test selected mid-stage tobacco leaves (leaf positions 9-12 from the bottom) and set three maturity levels: unripe, mature, and overripe. The plant spacing in Shifang was 0.38m and the row spacing was 1.1m, while the plant spacing in Dazhou was 0.4m and the row spacing was 1.1m. Field measures such as land preparation, ridging, fertilization, and transplanting were uniformly managed in accordance with Shifang and Dazhou wrapper cultivation techniques. Nitrogen application in Shifang was 142.5kg / hm2. 2 , N:P2O5:K2O is 1:1.2:2.4, and the nitrogen application rate in Dazhou planting area is 202.5kg / hm 2 , N:P2O5:K2O is 1:1.1:2.

[0032] Tobacco leaf image acquisition:

[0033] After each tobacco leaf is harvested, it is immediately photographed. The top, bottom, left, right, and back of the shooting environment are all covered with black panels, with one-third of the front panel left open for easy access. Fill lights are installed on the top, left, and right sides, and tobacco leaves are placed at the bottom. The camera is connected to an external computer.

[0034] Discriminant model selection

[0035] Two classic deep convolutional neural networks, ResNet_18 and AlexNet, with 71 and 25 layers, respectively, were selected as the classification models in this example. The model building framework was Matlab's Deep Learning Toolbox. The architecture, optimizer, and hyperparameters of the two models were adjusted and compared to explore the classification performance of models with different numbers of layers for wrapper tobacco leaves of varying maturity. The models used required image pixel sizes of 224×224×3 and 227×227×3, respectively, compared to the original image pixel size of 4864×2498×3. Adjustment of the original image pixel size was necessary to accommodate the model requirements.

[0036] Optimizer selection:

[0037] The optimizer adjusts the model's loss function, helping it gradually approach the optimal solution and improve classification performance. Common optimizers include Stochastic Gradient Decent Momentum (Sgdm), Root Mean Square Prop (Rmsprop), and Adaptive Moment Estimation (Adam). In this experiment, both models used Sgdm as the optimizer.

[0038] Hyperparameter Study:

[0039] For deep convolutional neural networks, the initial learning rate (InitialLearnRate) and mini-batch size (MiniBatchSize) are the two hyperparameters that most influence model convergence and loss reduction. The initial learning rate, a hyperparameter that guides network weights by adjusting the gradient of the network's loss function, represents the rate of information accumulation over training time. When the learning rate is optimal, the model's effective capacity is maximized. A larger mini-batch size increases the output error and is more susceptible to abnormal data. A smaller mini-batch size slows down network convergence and increases network complexity. The mini-batch size is the number of samples selected for network training at a time. With the default training epoch of 1, if there are 10,000 samples and the mini-batch size is 20, the number of samples trained at a time is 500. Generally, a larger mini-batch size increases the network's computational power for matrix multiplication and improves training results, but also increases memory and GPU consumption. A too low mini-batch size can introduce more noise during error calculation. This example tests different initial learning rates and mini-batch sizes to find the optimal combination.

[0040] Model evaluation method:

[0041] The AlexNet and ResNet_18 training process and generalization ability test are evaluated using the accuracy indicator. The specific calculation formula is:

[0042] Where Pr is the number of correct classifications, and Tr is the total number of samples.

[0043] After the architecture and hyperparameters are determined, the model is evaluated using a confusion matrix. The confusion matrix, also known as the error matrix, is a common method for evaluating the accuracy of image classification models. The confusion matrix calculates four types of metrics: a true positive (TP) when both the true and predicted values ​​are positive; a false negative (FN) when the true value is positive and the predicted value is negative; a true negative (TN) when both the true and predicted values ​​are negative; and a true negative (TN) when the true value is negative and the predicted value is true.

[0044] Based on the confusion matrix, we can calculate precision, recall, and the F1 score. Precision is the proportion of correct predictions among all results for which the model's predictions were positive. Recall is the proportion of correct predictions among all results for which the true values ​​were positive. The F1 score combines the results of precision and recall to evaluate the overall model's predictions. Its value ranges from 0 to 1, with closer to 1 indicating better model predictions. The calculation formulas are as follows.

[0045] Where Pr is the precision rate and Re is the recall rate.

[0046] Data processing method:

[0047] Matlab was used to build the model and system. The Matlab software and hardware experimental environment and version information are shown in Table 1.

[0048] Table 1

[0049] Results and Analysis

[0050] Tobacco leaf image acquisition results:

[0051] A total of 3043 images were collected for the two varieties in the 2022 experiment, as detailed in Table 2 .

[0052] Table 2

[0053] AlexNet recognition model establishment:

[0054] Convolutional layer design:

[0055] AlexNet's input is the captured RGB color image, and its output is the name of the folder containing images of tobacco leaves of different maturity levels. Specifically, before training, each category of images is placed in a separate folder, named according to the maturity category: "Unripe," "Moderately Ripe," and "Overripe." The classic AlexNet architecture was built in Matlab's Deep Learning Toolbox app. By adjusting the number and size of convolutional blocks, optimized AlexNet structures 1 and 2 were constructed, resulting in the corresponding first and second convolutional recognition models. The results are shown in Table 3. The images of the two tobacco varieties were divided according to a 4:1 ratio, with 80% of the samples randomly assigned to training and 20% to testing for model training and testing.

[0056] Table 3

[0057] Figure 1 shows the discrimination accuracy of the three AlexNet architectures for the test set samples. It can be seen that AlexNet optimized architecture 2 has the highest discrimination accuracy (93.6%). Looking at the entire model training process, the classic architecture experienced smaller oscillations and a smoother training process. Optimized architectures 1 and 2 experienced larger accuracy fluctuations when the iteration cycle was 100 to 200, but ultimately achieved improved accuracy. Compared to training with a single variety, the combined image set training iteration cycle doubled, resulting in longer training time. In summary, AlexNet optimized architecture 2 achieved superior discrimination accuracy compared to optimized architecture 1 and the classic architecture. Therefore, the corresponding second convolutional recognition model was selected as the final convolutional recognition model.

[0058] Hyperparameter Study

[0059] (1) Different mini-batch sizes:

[0060] Comparative experiments were conducted using four mini-batch sizes: 16, 32, 64, and 128. The image sample partitioning rules were the same as above. The results of the test set after training are shown in Figure 2. A mini-batch size of 32 achieved the best discrimination performance, with an accuracy of 93.5%. For AlexNet, a mini-batch size of 32 achieved the best discrimination accuracy.

[0061] (2) Different initial learning rates:

[0062] We explored the discriminative performance of AlexNet by setting three different learning rates: 0.001, 0.0001, and 0.00001. The image sample division rules were the same as above. The test set results after training are shown in Figure 3. As can be seen, the highest discrimination accuracy was achieved at 0.001, reaching 92.7%. Therefore, the optimal learning rate for AlexNet was determined to be 0.001.

[0063] Establishment of the improved AlexNet model:

[0064] Based on the above experimental results, AlexNet optimized structure 2 outperforms optimized structure 1 and the classic structure in recognition performance. The accuracy when the mini-batch size is set to 32 is significantly higher than those of 16, 64, and 128. When the initial learning rate is set to 0.001, the model stability and accuracy are higher than 0.0001 and 0.00001, respectively. An improved AlexNet model (i.e., a convolutional recognition model) was established using optimized structure 2, a mini-batch size of 32, and an initial learning rate of 0.001. The results are shown in Figure 4. As can be seen, only 4 samples in the unripe category were misidentified, only 9 in the ripe category, and the overripe category had a higher number of misidentifications, 12.

[0065] The precision, recall, and F1 score of the convolutional recognition model are calculated and the results are shown in Table 4. It can be seen that the overall precision of the model is 96.1%, the recall is 94.4%, and the F1 score is 0.952.

[0066] Table 4

[0067] ResNet-18 recognition method:

[0068] Convolutional layer design:

[0069] The construction, folder naming, and sample partitioning rules for different ResNet_18 structures are the same as above. The residual blocks of the ResNet_18 classic structure, optimized structure 1, and optimized structure 2 are shown in Table 5. Optimized structure 1 is the first residual recognition model, and optimized structure 2 is the second residual recognition model.

[0070] Table 5

[0071] Figure 5 shows that different ResNet_18 architectures differ in their ability to discriminate between image sets. The accuracy differences during training for the classic architecture, optimized architecture 1, and optimized architecture 2 are relatively small, reaching 97.4%, 97.9%, and 98.2%, respectively. Therefore, the second residual recognition model corresponding to optimized architecture 2 was selected as the final residual recognition model.

[0072] Hyperparameter Study:

[0073] (1) Different MiniBatchSize

[0074] Following the above image partitioning rules, we conducted comparative experiments with different mini-batch sizes. The accuracy of the test set during training is shown in Figure 6. Accuracy varies, generally decreasing with increasing mini-batch size. Throughout the training process, when mini-batch sizes are 128 and 64, the oscillation amplitude between iterations 50 and 150 is relatively small, but the final accuracy decreases. The oscillation amplitude is larger when the mini-batch size is 16, followed by 32. Compared to a mini-batch size of 16, a mini-batch size of 32 exhibits smaller oscillations and higher accuracy. Therefore, a mini-batch size of 32 achieves optimal performance for ResNet-18.

[0075] (2) Different initial learning rates:

[0076] Based on the above division rules, we conducted comparative experiments with different initial learning rates. The accuracy of the test set training process is shown in Figure 7. The differences in accuracy are significant, with 0.001 achieving the best accuracy, followed by 0.0001, and finally the lowest accuracy. The oscillation amplitude during training was greatest when the initial learning rate was 0.00001, followed by 0.0001, and finally the lowest. Therefore, 0.001 was determined to be the optimal learning rate for ResNet_18.

[0077] Improved ResNet_18 model establishment:

[0078] Based on the above experimental results, the ResNet_18 optimized structure 2 outperforms optimized structure 1 and the classic structure in recognition performance. The accuracy when the mini-batch size is set to 32 is significantly higher than that of 16, 64, and 128. When the initial learning rate is set to 0.001, the model stability and accuracy are higher than 0.0001 and 0.00001, respectively. An improved ResNet_18 model (i.e., the residual recognition model) was established using optimized structure 2, a mini-batch size of 32, and an initial learning rate of 0.001. The results are shown in Figure 8. As can be seen, only 8 samples in the immature category were misidentified, only 2 in the mature category, and the overripe category had a higher number of misidentifications, 4.

[0079] The precision, recall, and F1 score of the residual recognition model were calculated and are shown in Table 6. The model achieved an overall precision of 97.8%, a recall of 97.7%, and an F1 score of 0.978. The model still achieved satisfactory results after 30 training cycles, demonstrating its strong stability.

[0080] Table 6

[0081] Model generalization ability test

[0082] Based on the above experimental results, the improved ResNet_18 model achieved higher precision, recall, and F1 scores than AlexNet, by 1.7%, 3.3%, and 0.026, respectively. The improved ResNet_18 model was tested for generalization. Generalization is a key component in determining whether the model is overfitting. The test tobacco leaves were cigar tobacco images taken in 2023. Fifty images of each variety and maturity level were randomly selected for testing. The test results are shown in Table 7. Table 7 shows that the model performed well in classifying the generalized test samples, with an average accuracy exceeding 79% for both varieties, and a 92% accuracy rate for identifying unripe Chuanxue No. 1.

[0083] Table 7

[0084] Maturity Identification Application (Application, APP) Construction

[0085] APP practicality analysis:

[0086] A practicality analysis involves analyzing the functionality, interface layout, and usability of the app being built, providing a basis for action. The app's primary purpose is to quickly and accurately identify the maturity of wrapper tobacco leaves, overcoming the subjective nature of manual identification. Furthermore, the app can also provide a platform for users interested in tobacco leaf maturity identification. Based on the app's intended functionality, the process can be divided into three main steps:

[0087] (1) Image upload:

[0088] In order to solve the problem of different image formats taken by different devices, the APP needs to be able to import images of all formats, such as JPG, BMP, PNG, JPEG, etc.

[0089] (2) Image maturity determination:

[0090] This function needs to determine the maturity of the uploaded image, and this determination is achieved through the built classification model. Therefore, this step requires the callback of the classification model to determine the image maturity.

[0091] (3)Discrimination results display:

[0092] In this embodiment, the classification results are divided into three maturity levels: unripe, mature, and overripe. Therefore, this step primarily displays the classification results on the main interface. Furthermore, based on the results of this study, this step should also provide corresponding recommendations based on the classification results. The rules are as follows: If the result is unripe, the recommendation is: the tobacco leaves are insufficiently mature and harvesting is delayed; if the result is mature, the recommendation is: the tobacco leaves are at an appropriate maturity and harvesting is recommended immediately; if the result is overripe, the tobacco leaves are relatively mature and harvesting is recommended immediately.

[0093] APP building method:

[0094] The APP building environment is Matlab's APP design tool.

[0095] (1) Image upload function implementation:

[0096] [file,path] = uigetfile({'*.jpg;*.png;*.jpeg','Image Files(*.jpg,*.png,*.jpeg)'}) allows you to select images in all image formats. image = imread(fullfile(path,file)) opens an image from the device image library. imshow(image,'Parent',app.UIAxes) displays the imported image on the app interface. app.I = image implements a callback to the entire app.

[0097] (2) Classification function implementation:

[0098] I2 = imresize(app.I,[224,224]) to compress the image size. da = load("trainedNetwork.mat",'net') to call back the deep network model. [preLabels,probs] = classify(net,I2) to call the model for classification.

[0099] (3) Classification results display:

[0100] app.EditField.Value = char(preLabels), app.TextArea.Value = aa(nn), respectively realize the function of displaying the classification results and corresponding suggestions on the interface.

[0101] APP running:

[0102] After completing the APP design using the Matlab design tool, package the APP and install it in Matlab to complete the entire APP construction. Open the created APP in the APP tab of Matlab and click to enter the system. The next step is to select an image for maturity judgment.

[0103] The present embodiment also provides an intelligent recognition system for the maturity of cigar wrapper tobacco leaves, including: an image acquisition module, used to acquire cigar tobacco leaf images and perform classification processing; a model construction module, connected to the image acquisition module, used to adjust the number and size of convolution blocks and residual blocks to improve the ResNet_18 model, construct a residual recognition model, and determine the mini-batch size and initial learning rate of the residual recognition model; a model training module, connected to the model construction module, used to train the residual recognition model based on the classified cigar tobacco leaf images; and an intelligent recognition module, connected to the model training module, used to identify the maturity of cigar tobacco leaves based on the trained residual recognition model.

[0104] The model construction module can be implemented to include: a model construction unit, which is used to obtain a first residual recognition model based on the ResNet_18 model structure by changing the size of the convolution layer and increasing the number of residual layers; and a second residual recognition model by changing the size of the convolution layer and reducing the number of residual layers; a model comparison unit, which is used to compare the accuracy of the cigar tobacco leaf recognition results of the unimproved ResNet_18 model, the first residual recognition model and the second residual recognition model, and use the second residual recognition model as the final residual recognition model.

[0105] In this embodiment, an AlexNet classification model, i.e., a residual recognition model, is established using RGB images taken by an industrial camera, and a comparative study is conducted on the structure and hyperparameters of the model. The results show that the recognition accuracy of the test set of the optimized structure 2 is higher than that of the classic structure and the optimized structure 1. The results of the hyperparameter study show that when MiniBatchSize and InitialLearnRate are 32 and 0.001, respectively, the model recognition accuracy is the highest. After the model is established, it is usually necessary to further establish a recognition applet or application software to realize the practical application of the model. In this embodiment, the constructed model is successfully designed as a recognition APP. In future applications, it is only necessary to upload the cigar tobacco leaf image to the computer to realize the recognition of new images.

[0106] This embodiment collects images of cigar tobacco leaves of different maturity levels to identify maturity, and the results are ideal. The following conclusions can be drawn: the accuracy of ResNet_18 and AlexNet in the training process with optimized structure 2 is higher than that of the classic structure and optimized structure 1, and the accuracy is best when InitialLearnRate and MiniBatchSize are 0.001 and 32 respectively. The accuracy, recall rate, and F1 score of the improved ResNet_18 are higher than those of AlexNet, by 1.7%, 3.3%, and 0.026 respectively. The model generalization test results are ideal, with an average accuracy of more than 79% for both varieties, and the accuracy of identifying the unripe Chuanxue No. 1 is 92%. The application built can quickly and accurately identify the maturity of cigar tobacco leaves, providing a method for determining maturity.

[0107] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for intelligently identifying the maturity of cigar wrapper tobacco leaves, characterized in that: The following steps are involved: Obtain cigar leaf images and perform classification processing; The ResNet_18 model was improved by adjusting the number and size of convolutional blocks and residual blocks to construct a residual recognition model, which was used as a recognition model for cigar tobacco leaf maturity. Determining a mini-batch size and an initial learning rate for the residual recognition model; Training the residual recognition model based on the classified cigar leaf images; The maturity of cigar tobacco leaves is identified based on the trained residual recognition model.

2. The intelligent identification method for cigar wrapper tobacco maturity according to claim 1, characterized in that: The categories of the cigar tobacco leaf images include at least unripe, ripe and overripe.

3. The intelligent identification method for cigar wrapper tobacco maturity according to claim 1, characterized in that: The process of constructing the residual recognition model includes: based on the ResNet_18 model structure, by changing the size of the convolution layer and increasing the number of residual layers, obtaining a first residual recognition model; by changing the size of the convolution layer and reducing the number of residual layers, obtaining a second residual recognition model; by comparing the accuracy of the cigar tobacco leaf recognition results of the unimproved ResNet_18 model, the first residual recognition model and the second residual recognition model, the second residual recognition model is used as the final residual recognition model.

4. The intelligent identification method for cigar wrapper tobacco maturity according to claim 3, characterized in that: The mini-batch size of the residual recognition model is 32, and the initial learning rate is 0.

001.

5. The intelligent identification method for cigar wrapper tobacco maturity according to claim 1, characterized in that: The optimizer of the residual recognition model is Sgdm.

6. An intelligent recognition system for cigar wrapper tobacco maturity, characterized in that: include: An image acquisition module is used to acquire cigar tobacco leaf images and perform classification processing; The model building module is connected to the image acquisition module and is used to adjust the number and size of convolution blocks and residual blocks to improve the ResNet_18 model, build a residual recognition model, and The model's mini-batch size and initial learning rate are determined; a model training module, connected to the model building module, for training the residual recognition model based on the classified cigar leaf images; The intelligent recognition module is connected to the model training module and is used to identify the maturity of cigar tobacco leaves based on the trained residual recognition model.

7. The intelligent identification system for cigar wrapper tobacco maturity according to claim 6, characterized in that: The model building module includes: A model construction unit is used to obtain a first residual recognition model based on the ResNet_18 model structure by changing the size of the convolution layer and increasing the number of residual layers; and to obtain a second residual recognition model by changing the size of the convolution layer and reducing the number of residual layers; The model comparison unit is used to compare the accuracy of cigar leaf recognition results of the unimproved ResNet_18 model, the first residual recognition model and the second residual recognition model, and use the second residual recognition model as the final residual recognition model.

8. The intelligent identification system for cigar wrapper tobacco maturity according to claim 7, characterized in that: The network structure of the residual recognition model includes: convolutional layer 1, [7×7, 64, [2 2]]; residual block 2, group; residual block 3, Branch, [1×1, 128, [2 2]]; residual block 4, Branch, [1×1, 256, [2 2]].

Images