Research On PSO-BP Weed Image Recognition Method Applied To Weeding Robot

Weeding robot is a weed recognition device based on machine vision.Through the image analysis of the camera,weeds and crops can be distinguished accurately and weeding efficiently.Compared with the methods of artificial weeding and mechanical pesticide spraying,the weeding robot has the advantages of low cost,high efficiency and small pollution,it is also the most valuable weeding methods at home and abroad.The farmland weeding robot is an intelligent autonomous decision system that can be divided into three parts: autonomous path navigation,visual image recognition,precise weeding action.Among them,visual image recognition is used to accurately distinguish weeds from crops and provide decision information for weeding.It is a key technology for weeding robots.Focusing on the problem of visual grass recognition in corn,the identification and classification of maize and weeds were conducted by using the multi-feature fusion technology of color,shape and texture.The main work is as follows: In the light intensity is appropriate,the shooting angle of 45 degrees with the ground conditions,through the camera mobile terminal to capture a large number of maize in the seedling weeds pictures.First of all,the target image is segmented by using the method of super green graying and connecting component marking.Based on this,three aspects of the color,shape and texture of the seedling weeds were extracted and the characteristic parameter combinations were constructed.Finally we use pattern recognition algorithm and neural network algorithm to complete the seedling weeds recognition classification.First,based on the principle of minimizing the probability error,a Bayesian classifier based on multi-features of color,shape and texture is built.Secondly,according to the characteristics of large and nonlinear feature data,a three-layer BP neural network weed classification model with 15 inputs and 5 outputs was established.Thirdly,using the global search capabilities of Particle Swarm Optimization(PSO),the connection weight threshold of BP network is optimized,and the PSO-BP neural network weed classification model is constructed to improve the stability of the algorithm and predict the classification accuracy.In the MATLAB software environment,contrastive analysis of three recognition algorithms with the images of five kinds of seedling weeds including corn seedlings,broom grass,alfalfa,amaranth and bracken.The results show that the Bayesianclassifier has a simple set-up procedure,only the label and probability calculation,but it is difficult to solve the changes caused by feature reorganization,and the recognition effect in the actual image acquisition environment is not ideal.BP neural network classification model has the ability of self-learning for the change of characteristic data and predicts the classification accuracy is higher than that of Bayesian classifier,but the problem of poor stability still exists.This paper puts forward PSO-BP neural network weed classification model,effectively solves the dependence of BP network on initial value and predicts fluctuation,improves stability and recognition rate.The experimental results show that the PSO-BP algorithm is the most suitable for early weed classification in the three types of models.Finally,based on the PSO-BP neural network model,a seed-weeding recognition simulation system is designed,which integrates camera acquisition,image processing,feature extraction,and recognition classification,it validates the effectiveness and feasibility of the algorithm in practical applications.