Research On Image Recognition Method Of Farmland Weeds Based On Deep Learning

Weeds in farmland can severely damage the growth of crops and cause great losses to agricultural production.At this stage,large-scale spraying of chemical herbicides is still the most important weeding method.Although it is fast and efficient,this method is likely to cause soil environmental pollution and food security problems,and is not in line with the agricultural green development.In order to use herbicides efficiently and rationally,and to avoid undifferentiated large-scale spraying,it is necessary to implement precise spraying,with precise identification of weeds in farmland as basis.In this paper,in order to solve the problems of low accuracy and slow speed of weed identification in the farmland environment,we choose the associated weeds of corn and rice as the research object.Based on the deep separable convolutional network model,and combined with the migration learning strategy,this paper studied the weed identification method deeply,which provides a theoretical basis and reference for the follow-up farmland weed control.The specific research work is as follows:(1)Image data of farmland weeds are collected and processed.We collect,classify and label common farmland weed data,and expand the data through data enhancement methods to increase the diversity of the data and to improve the robustness of the subsequent model;Based on the characteristics of the data set of weeds and corn seedlings,the data set is preprocessed.As the weeds in farmland are green,the normalized super green algorithm and Otsu threshold segmentation algorithm are selected to remove the background of the image,and only the main body of weeds is retained to improve the performance of subsequent models.(2)A farmland weed recognition method based on improved deep separable convolutional network is proposed.In this paper,the deep separable convolutional network Xception is improved by using the ELU activation function and the global maximum pooling layer,which can improve the convergence speed of the model and enhance the expression of the effective characteristics of weeds.(3)The network model based on the migration learning strategy is improved.A priori knowledge is used to train the weed recognition model through transfer learning strategy.Compared with the training from scratch,the training time of the model can be shortened,the convergence of the training model accelerated,and the recognition accuracy is improved by1.93%.(4)Development of farmland weed recognition system.This paper uses the interface development tool PyQt5 to design and develop an intelligent identification system for farmland weeds.The system is convenient for users to operate,and can realize the recognition of farmland weed images in a few simple steps,which is proved simple and efficient.In the field test,the accuracy can reach 95.9%,which has good practical application value.