| Flue-cured tobacco planting is a typical labor-intensive industry.For a long time,the labor cost generally accounts for more than half of the total planting cost.Especially in the northeast tobacco area,the main labor force of tobacco farmers’families is lost,and the tobacco production is facing the aging of agricultural workers and the non-agriculturalization of rural labor force.The current situation of difficult and expensive employment seriously restricts the development of the local tobacco industry.The research of this paper is based on solving the problem that the transplanting technology of flue-cured tobacco under the film in the northeast tobacco area is heavily dependent on manual completion,the work is heavy,and the degree of mechanization is very low,which seriously restricts the development of the tobacco industry.Combined with the integrated equipment of cutting seedlings and covering soil developed and designed in the previous research of the project,through the screening and testing of many target recognition and positioning methods,a target recognition and positioning algorithm of tobacco seedlings under the film based on machine vision is proposed,and the embedded plant distance control system based on vision is preliminarily built.The main research contents of this paper are as follows:(1)The characteristics of tobacco seedlings under film were systematically studied through field experiments.In this study,a standardized test tobacco field and a mobile information acquisition platform were built,and about 3000 tobacco seedling image samples under the film were collected.The sample images were preprocessed by cutting and adaptive global image equalization.Then,the tobacco seedling target under the film was extracted and segmented by HOG(histogram of oriented gradient)feature extraction method,LBP(texture)feature extraction method and LAB color space threshold segmentation method.The experimental results show that the irregular distribution of water droplets under the film will produce a microlens effect on the light passing through the film.The aging and wrinkle of the film itself,as well as the dust and water stains on the outer surface,will lead to a gradual decrease in light transmittance.In addition,the smooth outer surface of the film will also refract the light.The continuous growth of tobacco seedlings and the continuous change of light will also lead to the continuous change of its characteristics covered by the film.The existence of these multiple random factors makes it difficult for the hand-designed feature extraction algorithm and lidar/infrared vision method and sensor method to adapt to the actual working environment.(2)A transfer learning method based on Mobile NetV2 lightweight pre-trained neural network model is proposed to realize the target recognition and accurate positioning of tobacco seedlings under film in real field environment.Firstly,the composition of the sample data set was re-sorted and collected in combination with the characteristic distribution of tobacco seedlings under the film.The image samples of 6-15 days after transplanting were used to form a new data set according to the law of increasing in number,and the hyperparameters(learning rate,width factor and resolution factor,etc.)were optimized.The verification results of the model on the verification set and the test set show that the correct rate of the model is 97.60%and the F1 value is 99.20%.The reasoning time is short,the model size is appropriate,and it meets the design requirements.Subsequently,the model was deployed with embedded devices and field experiments were carried out.The experimental results showed that the accuracy of target recognition and location of tobacco seedlings under film was 90.29%,and the mean values of missed detection rate and false detection rate were 3.00%and 6.29%,respectively,which verified that the model had strong generalization in real field working environment.(3)An embedded plant spacing control system based on Mobile NetV2 lightweight pre-trained neural network transfer learning model was initially built,and the robustness and generalization of the control system based on this method were further verified in field experiments.Firstly,the corresponding relationship between the output results of the model and the location of the tobacco seedlings under the film in the real environment was analyzed,and the ’eye in hand’ control program was written.Then,the embedded control system of information acquisition-logic processing-executing mechanism was realized through the serial communication protocol.Finally,the field simulated seedling digging test was carried out by using the seedling digging and soil covering equipment equipped with the seedling digging mechanism(simplified as a mechanical arm).The test results showed that the correct operation rate of seedling digging based on the embedded plant spacing control system was 83%.The robustness of the system and the feasibility of using the target recognition and positioning method to replace the original ’walking wheel-transmission-electromagnetic valve’ plant spacing control system scheme are further verified.A series of experiments show that the method based on lightweight pre-training model transfer learning proposed in this paper can accurately identify the tobacco seedlings under the film and has excellent positioning accuracy.The proposed method has good robustness and generalization,and can adapt to the complex working environment in the field.(4)Taking the tobacco planting area of Ning ’an City,Mudanjiang City,Heilongjiang Province as an example,the theoretical analysis of the expected benefits generated in the actual production of the intelligent digging seedling and covering soil automation equipment using the research method of this paper is carried out.The analysis results show that the use of the equipment can reduce the cost of tobacco farmers in the process of digging seedlings and covering soil.More than half,and there are potential benefits to reduce planting risks. |
PDF Full Text Request