Research On Position Detection Of Tobacco Seedling Pit Based On Convolution Neural Network

Tobacco is an important economic plant in our country.The development of the tobacco industry increases the national financial revenue,promotes the development of related industries,is conducive to improving the income of tobacco farmers,and is an important economic industry for the benefit of the country and the people.From the beginning of tobacco seedling,the whole process of tobacco planting should be strictly controlled so as to harvest tobacco with high yield and good quality.According to the rhythm of tobacco growth and development,between seedbed stage and field stage,there is a seedling stage for tobacco to go through.Seedling stage is the key period for survival during transplanting,when the seedling has to adjust to complex field environment.The pit-transplanting technology of tobacco seedlings creates a good growth environment for tobacco plants in the field transplanting process,improves the root,s growing,effectively shortens the adjustment period,makes the tobacco seedlings grow vigorously,and is conducive to improving the yield and quality of flue-cured tobacco.The pit-transplanting technology of tobacco seedlings creates a good growth environment for tobacco plants in the field transplanting process,which makes tobacco seedlings grow in a better temperature and humidity environment,helps the growth and development of root system,enhances the vigor of root system and rapid growth of tobacco seedlings,and improves the yield and quality of flue-cured tobacco.At present,tobacco seedlings5 transplanting is manual work with low degree mechanization.Detection of the location of tobacco transplanting pit is the key to acheive mechanized seedlings5 throwing to the pit.This paper is focusing on detection of tobacco transplanting pit.In this paper,image data set of tobacco transplanting pit as the basic data of target detection algorithm model is made.Based on the characteristics of tobacco transplanting pit image data,the convolution neural network models for object detection is analyzed,and YOLO comes to the basic network structure for pit detection model.The theory of transfer learning,back propagation algorithm and batch normalization processing methodare applied to training the model.In the further process,the key configuration,IOU PredTruth and Con(object),as well as the input-image size were optimized by orthogonal tests,and the performance of the model was evaluated in the test set with negative samples.With calibrating the camera,the conversion between the pixel plane coordinate system,the image plane coordinate system,the camera coordinate system and the world coordinate system in the imaging process is acheived.The field test of the tobacco transplanting pit position detection model is carried out,and the prediction error is recorded.In this paper,with oversampling and data augmentation,the number of image samples is extended from 500 to 1320.Meanwhile,the number of strong-lighted the number of samples and weak-lighted samples are almost the same.Therefore,the image data set is insensitive to the intensity of illumination and it would support the model to fit the natural environment as its training set.The principle and function of different layers of convolutional neural network are analyzed and the related algorithms focusing on object detection are compared,while the basic position detecton model for tobacco transplanting pit is chosen.Based on the transfer learning theory,the weight function obtained by pre-training in large image database is used to initialize the model,and the image data set of tobacco transplanting pits in this paper is used as training data.When the model is training,the weights would be updated through back propagation and the batch normalization method is helpful to optimize the training model efficiently.By adjusting the configuration of trained model,it reaches the recall rate of 80%.Field experiments show that the average prediction error and the maximum prediction error are 29.3 mm and that could reach of tobacco transplanting pit s position.