Research On Recognition Method Of Tobacco Leaf Maturity And Flue-cured Stage Based On Machine Vision

Strict classification or identification work is required during the tobacco leaf fluecured process to standardize and match appropriate flue-cured techniques,reduce fluecured losses,improve tobacco leaf flue-cured quality,and achieve the goal of benefiting the people and increasing income.The recognition of tobacco leaf maturity and flue-cured stage is crucial for tobacco leaf flue-curing.The moisture content and other indicators of tobacco leaves with different maturity levels are different,resulting in significant differences in their flue-curing resistance.Generally speaking,as the maturity of tobacco leaves increases,the flue-curing resistance will decrease.If tobacco leaves with different maturity levels are mixed for flue-curing,the flue-cured process is difficult to balance the flue-curing characteristics of tobacco leaves with different maturity levels,which can easily cause unnecessary flue-cured losses.Manual classification is difficult to form a unified standard,consumes a lot of manpower,but the classification purity is not enough.Therefore,machine vision technology can be used to identify and classify the maturity of tobacco leaves,reducing the difficulty of matching subsequent flue-cured processes,thereby improving the overall flue-cured quality of tobacco leaves and reducing fluecured losses.However,the current classification of tobacco leaf maturity research is limited to identifying whether the harvested tobacco leaves are mature or not,and there is no detailed classification of maturity level.After harvesting,it is often necessary to reclassify them before hanging them for flue-curing,And the research is mainly conducted in a laboratory environment,and the generalization of the recognition model in the production site is difficult to guarantee.During the tobacco leaf flue-cured process,the flue-cured master needs to continuously observe the flue-cured status of each room’s tobacco leaves and adjust the temperature and humidity based on the flue-cured status.This manual inspection and adjustment method is greatly influenced by the subjective influence of the flue-cured master,and is prone to visual fatigue.Therefore,using machine vision methods to identify the flue-cured stage of tobacco leaves during the fluecured process and assist the flue-cured master in judging the flue-cured process can greatly reduce the subjective influence of the flue-cured master,ensure the accuracy of the flue-cured process,and improve the flue-cured quality of tobacco leaves.However,due to the high dependence of the recognition process on equipment performance,it is still difficult to test the application on site.In response to the current problems,the main research content and conclusions of this article are as follows:(1)Collected tobacco leaf images of different maturity levels and different fluecured stages with relatively complete categories,and constructed datasets separately.This article divides the original data into training,validation,and testing sets in a 7:2:1 ratio,and enhances the data of the divided datasets.Gaussian noise was added to simulate the complex on-site environment of tobacco leaf images with different maturity levels in this article,and methods such as vertical flipping,horizontal mirroring,and clockwise and counterclockwise center rotation were used to expand the data.Finally,a tobacco leaf maturity image dataset was constructed.For the tobacco leaf flue-cured stage images used in this article,methods such as clockwise and counterclockwise center rotation and brightness change were used to expand the data,and finally a tobacco leaf flue-cured stage image dataset was constructed.(2)Based on the Mobile Net V2 model,an MSNet-FTLI machine vision recognition model for tobacco leaf maturity was constructed.Compared with mainstream advanced models,the MSNet-FTLI model achieved higher recognition accuracy with fewer parameters and faster recognition speed.In this paper,a multi-scale channel shuffle module is proposed structurally to realize the diversified gradient calculation area,and the FRe Lu activation function is used to complete the model training in cooperation with the SGDM optimizer.The accuracy of the MSNet-FTLI model on the test set has reached94.59%,the model parameters are only 1.34 M,and the detection time of 100 tobacco images on the computing server is only 3.85 s.Through gradient calculation of thermal maps and shallow and deep feature maps of the model,the main focus areas,feature activation within the feature maps,and the number of effective feature maps were analyzed.The results showed that the MSNet-FTLI model proposed in this paper has a higher attention to the overall tobacco leaf,and feature extraction fully considers the overall correlation of image features.The number of activated features and effective feature maps applied to image classification has significantly increased.The classification performance of the model for each maturity category was compared and analyzed through statistical confusion matrix,F1-score calculation,radar chart comparison and other methods.The results showed that the overall classification effect of MSNet-FTLI model was better,especially for the classification effect of upper over mature tobacco leaves.The model proposed in this article is based on the principle of gradient calculation,which directly improves the model structure to improve model performance,reduces complex image preprocessing processes,improves model robustness,and maintains high accuracy in tobacco maturity recognition while ensuring sufficient low hardware resource consumption.(3)A machine vision recognition model for Ghost Net-TFSI tobacco leaf flue-cured stage was constructed based on the Ghost Net model.On the premise that the recognition accuracy is close to that of large models,the Ghost Net-TFSI model has the characteristics of fewer model parameters and higher application of embedded devices in the flue-cured room.This article uses a more efficient ECA attention module to reduce model resource consumption,and cooperates with the SGDM optimizer for model training.The test accuracy on the flue-cured stage images collected in the flue-cured room reached 92.05%,with only 4.43 M model parameters.Through statistical confusion matrix,F1-score calculation,radar chart comparison and other methods,the classification performance of the model for each flue-cured stage was compared and analyzed.The results showed that the overall recognition performance of the improved Ghost Net-TFSI model improved significantly,and while ensuring the recognition accuracy,it reduced the complexity of branch structures such as attention mechanism,reduced the model resource occupation,and had better applicability in the field of tobacco flue-cured stage identification.(4)A tobacco maturity and tobacco baking stage recognition system was designed based on the Nvidia Jetson Nano embedded platform and Py Qt5 software development tool.Based on the application scenarios and functional requirements of the system,the hardware equipment and software functions of this system have been determined.We have built a hardware platform that includes Nvidia Jetson Nano development board,cameras,USB relays,and more.Based on the tobacco maturity recognition model and tobacco baking stage recognition model constructed in chapters 3 and 4,a Py Qt5 based tobacco recognition system operation graphical interface has been developed.And based on the intelligent tobacco baking related project,the system was tested.When detecting the maturity of tobacco leaves on the Nvidia Jetson Nano mobile detection platform,the speed can reach 0.17 seconds per sheet;When identifying the tobacco leaf baking stage,it can reach 0.87 seconds per sheet.The recognition speed basically meets the timeliness requirements of tobacco leaf recognition,and also provides theoretical basis and technical support for the subsequent design of tobacco leaf detection equipment..