Research On Automatic Classification Method Of Flue-cured Tobacco Based On Machine Vision

China has a great demand for cigarettes and is the largest cigarette producer and consumer all over the world.However,at present,tobacco leaf sorting in my country is completely by artificial operation,artificial separation with strong subjective consciousness,and for a long time for this repetitive operation would lower the separation efficiency and accuracy,combined with our country promulgated GB2635-92 standard for tobacco leaves of flue-cured tobacco separation property descriptions are fuzzy,many properties without quantitative division,In addition,due to the large turnover of personnel,flue-cured tobacco companies will conduct training for sorting personnel every year,which increases the cost.As the technology develops,many agricultural products have achieved automation in the field of machine vision,so the demand for automatic tobacco grading is more and more urgent.Most of the existing methods use color features or shape features to sort tobacco leaves,but tobacco leaves sorting is in-class sorting,which belongs to the image classification of delicacy.If only using machine vision to classify tobacco leaves,can’t guarantee accuracy,even not as high as manual sorting rate.Therefore,in order to solve the problems of high labor cost and low accuracy of tobacco leaf sorting,this thesis proposes a network model that combines traditional features and deep learning features to classify tobacco leaves.Thesis’ s main work is as follows:1.Build data sets.Since there is no data set related to tobacco grading in the existing data set,the data set needs to be constructed before deep learning training neural network.In general,the greater the amount of training data for any deep learning task,the better the results.Therefore,this paper uses flip and rotation operations to enhance the data quantity of the training set data.2.This thesis studies a maximum edge foreground extraction algorithm suitable for tobacco leaves,and extracted the foreground of the image to reduce the interference factors of subsequent image operation and improve the running speed.3.This thesis designs a new morphological feature extraction algorithm,combined with custom grayscale,to solve the shortcomings of inaccurate and low robustness in extracting morphological features of the previous algorithm;The color features of tobacco leaves are extracted by using HSI color space and color moment methods;A Gabor filter bank is selected to extract the texture features of tobacco leaf image,and the average and variance of each output image are obtained.4.In this thesis,convolutional neural network efficient Netv2-S was optimized and applied to the tobacco grading task.Support vector machine was used to determine the difficult samples,and the difficult samples were added to the training set.The transfer learning method was used to assign initial values to the model,which reduced the overfitting phenomenon of the model and improved the prediction accuracy.The traditional features obtained by machine vision technology,such as shape feature,color feature and texture feature,were integrated into the neural network model,and efficient Netv2-S inner layer was modified to improve the accuracy of the model sorting,making the accuracy up to 90.0%.