Research On Identification Of Maize Seedlings And Weeds Based On Deep Learning

In recent years,computer technology has developed at a high speed.Computer vision technology has developed rapidly because of its close connection with the industry,and its results have greatly facilitated people’s lives.One important research field of computer vision is object detection,and it has achieved good results in many application fields,such as face detection,pedestrian detection,vehicle detection and so on.With the rise of deep learning technology,its powerful ability of feature extraction and fitting capabilities make it the mainstream of target detection.In agricultural production field,corn is one of important crops in China,and its growth will be seriously affected by weeds.However,existing methods such as mechanical weeding and chemical weeding have various shortcomings,such as pollution to the environment,increasing labor costs,etc.So the research of efficient intelligent weeding technology is very practical.Based on the above-mentioned problems,this paper aims to achieve rapid and accurate identification of corn seedlings and weeds,and uses techniques in computer vision and deep learning to analyze and study methods for automatically identifying corn seedlings and weeds.In this paper,based on the identification of corn seedlings and weeds in agricultural production,an efficient image object detection model was proposed,which can do end-to-end training.And a large number of experiments were carried out on the corn seedlings and weed datasets which was created by ourselves.Finally we proposed an effective method to solve the problem of weed identification of corn seedlings.At present,there are two main types of models for solving object detection problems based on the method of deep learning and convolutional neural network.One is a twostage model based on region proposal networks represented by Faster-RCNN,which has advantages in detection accuracy.The other is a single-stage detection model that can be end-to-end training represented by SSD,which has advantages in speed.Based on the SSD model,this paper proposes an end-to-end image object detection model with higher accuracy.The main contributions : First,the original SSD model is a light-weight model.We have added dilated convolution,Inception and DenseNet structure to enhance the feature extraction capabilities of the network.Second,learning from the characteristics of the Two-stage method of Faster-RCNN,a Refinedet structure is used in the model,which is a method that imitates the two-regression of bounding box,which can effectively improve the detection accuracy degree.On the VOC2012 dataset,the model we finally proposed which is named WeedNet showed a good detection result,which is better than several classic models in detection accuracy.For the specific topic of identification of corn seedlings and weeds,because there is no published dataset of corn seedling and weed,the first step is the production of datasets.We collected a large number of high-resolution images of corn seedlings and weeds which planted in greenhouse.The images are taking photos at conditions with different lighting,angles and other situations.Then we used the annotation tool called LabelImage to manually annotate the image and create a dataset which is in the VOC2012 format.With the transfer learning method and the characteristics of the weed dataset,we continued to train the model on the weed dataset based on the models trained on the VOC2012 dataset.We compare the experimental results with other deep learning detection models and traditional object detection methods.The results show that our models have certain advantages in detection accuracy and speed.