Research On Weed Detection In Farmland Based On Computer Vision

Weeds in agricultural fields have great influence on crop growth and yield,so timely and effective weed control is an important measure to ensure national grain reserves.However,most of the current weed recognition technologies have shortcomings in practical applications,such as traditional target recognition algorithms cannot achieve large-scale sample training,poor generalization ability and low recognition accuracy.In this paper,we take sugar beet and associated weeds in complex scenes of farmland as the research object,and use deep learning algorithms to train the acquired dataset,aiming to improve the model recognition accuracy and accelerate the model convergence speed,so as to achieve the accurate removal of weeds in the field and the accurate fertilization of crops.To accomplish the above study,the following works were done:(1)Introduce the basic structure of convolutional neural networks and discuss the current mainstream target detection algorithms,including One-stage based algorithms: such as SSD and Yolo series,and Two-stage based algorithms: such as Fast R＿CNN and Faster R＿CNN.Given that the target detection in complex scenes like farmland has high requirements for recognition accuracy.Therefore,this paper chooses to use Faster R＿CNN as the network model for weed recognition in farmland.(2)A new object detection algorithm model based on Faster R-CNN is constructed.First,a dataset that meets the experimental requirements is generated using publicly available datasets and employing data augmentation techniques(e.g.,Augmentor library and adversarial generative networks)and processed using annotation tools.Second,transfer learning is used to verify that the transfer learning technique can accelerate the convergence of the model.Then,the backbone feature extraction network is optimized to ensure that the feature information is extracted more adequately.Finally,different attention mechanisms are introduced to achieve adaptive attention to the detection object.Experimental results show that the improved algorithm in this paper not only shortens the training time and accelerates the convergence speed of the model,but also effectively improves the detection accuracy of the model.(3)A GUI tuning interface based on Faster R＿CNN is built,which also incorporates a speech system to guide the user through the detection process;for the same image data,the hyperparameters are adjusted according to the idea of the control variable method to optimize the detection accuracy of the model;in addition,multiple detection images are used to verify the effectiveness and convenience of the tuning interface,and all detection results are visualized.Through the in-depth study of this paper,the improved Faster R＿CNN algorithm can realize weed recognition in complex scenes,and the algorithm not only accelerates the model convergence but also effectively improves the recognition accuracy with the help of migration learning,which provides some theoretical support for the weed recognition technology in farmland.