Research On Weed Recognition Method Of Soybean Seedling Stage Based On Deep Learning

Soybean is a very important oilseed crop and one of the key crops for maintaining national food security.Soybeans have a very important role in the diet of China and the world.Despite the complexity of the environment in which soybeans grow,they are able to thrive in nature.However,due to the presence of weeds,they can compete with soybeans for resources,causing their yield and quality to suffer or even hinder their development.To improve soybean production and economic efficiency,it is important to detect and treat weeds in a timely manner.However,traditional manual weed control methods are time-consuming and inefficient,and improper use of herbicides can contaminate the soil and growing environment.Therefore,the use of efficient soybean and weed identification methods is necessary to achieve accurate weed identification and monitoring.This paper addresses the problem of low accuracy of weed identification in soybean fields under natural environment,and selects soybean seedlings and their associated weeds as the experimental research object to optimize the target detection model in the deep learning method to provide a scientific method for subsequent accurate identification and monitoring of weed damage.The research content of this paper is as follows:(1)Weed image dataset and sample library of soybean seedling stage in the field were created.In this experimental study,two acquisition methods,cell phone and drone,were used to obtain soybean weed data at two scales,cell phone and drone,at the seedling stage in the field.The image annotation tool Label Img was used to label the labels,and the data enhancement methods of color transformation and geometric transformation were used.The purpose was used to expand the constructed dataset to increase the diversity of weed samples.Two kinds of cell phone scale soybean weed datasets including soybean,grass weeds,and broadleaf weeds,and low altitude drone scale weed datasets are constructed for experimental studies.(2)A weed identification method based on the optimized Faster R-CNN algorithm for soybean seedling stage is proposed.The training and identification were performed using data captured by cell phone devices.Firstly,by comparing the classification recognition effects of Res Net50,VGG16 and VGG19,VGG19 was determined as the backbone feature extraction network with the best model performance;secondly,the convolutional attention mechanism was embedded in three different positions after the pooling layer in the latter half of VGG19,and the best accuracy of the embedded model was determined after Block4 and Block5 to form the VGG19-CBAM structure,and finally,comparison experiments are done with SSD and Yolov4.Using the trained Faster R-CNN algorithm to recognize soybean and weeds in the field under natural environment,the experimental results show that the average recognition speed of single image of Faster R-CNN method with VGG19-CBAM structure as the backbone extraction network structure is 336 ms,the average recognition accuracy is 99.16%,which is 5.61% higher than before optimization,higher than SSD algorithm 2.24%,and 1.24% higher than the Yolov4 algorithm,which can realize the recognition of soybean seedling weeds at the cell phone scale under natural environment and provide a scientific approach for weed identification and control.(3)For the low altitude weed data images acquired by UAV equipment,apply them to the target detection method Faster R-CNN,propose a soybean weed recognition method based on low altitude UAV images and Faster R-CNN,experiments firstly compare four kinds of backbone feature extraction network structures,Res Net50,Res Net101,VGG16,VGG19,determine the most VGG16 structure is optimal,analyze the influence of different feature extraction structures on the model accuracy,secondly optimize the anchor frame for the UAV weed data own characteristics,and finally compare the experimentally optimized model with the target detection algorithm SSD and YOLOv3 for comparison test.The results show that by optimizing the anchor frame scale,the model of Faster R-CNN with VGG16 structure as the backbone feature extraction network has the best results,and the average recognition rate reaches 88.69%,which is 6.31% and 5.79% higher than SSD and YOLOv3 in recognition accuracy,respectively,and can effectively detect soybean seedling weeds at the low-altitude UAV scale,improve the detection of the model speed and enhanced the generalization ability of the model.