Eupatorium Adenophorum Detection Based On Deep Learning

China is constantly invaded by alien species,which can cause damage to ecological diversity and affect economic development in agriculture and forestry.Among the invasive plants suffered by alien species in China,Eupatorium adenophorum is one of the most typical examples.At present,the country has adopted various methods to control it.Detection is a key step in the prevention and control of Eupatorium adenophorum.How to correctly identify and detect Eupatorium adenophorum is an urgent problem to be solved.This thesis proposes a deep learning-based target detection method for classification and localization of Eupatorium adenophorum against a dataset of Eupatorium adenophorum under complex backgrounds.The work done includes the following aspects:(1)The object detection algorithm based on regional nomination are introduced in details and their performance are compared,the better Faster R-CNN algorithm is selected from these to realize the detection of Eupatorium adenophorum.(2)YOLOv3,as the best performance algorithm in regression-based object detection algorithms,has achieved good results in both speed and accuracy.Aiming at the single-type target detection problem of Eupatorium adenophorum,the thesis is based on the YOLOv3 model.First,the anchor box parameters are modified to improve the adaptability of the model to Eupatorium adenophorum.Secondly,modify the classifier and change the output to an output tensor of 18 dimensions to complete the classification and positioning tasks of Eupatorium adenophorum.(3)In order to improve the detection accuracy,the thesis makes the following improvements to the problems of YOLOv3 transfer learning method: First,the depth method requires massive amounts of data for training and testing to enhance the overall performance of the model.Therefore,the thesis has performed data flipping and shading adjustments on the Eupatorium adenophorum data,these adjustments don not loss of image information and can enhance the generalization ability of the network.Second,the loss function in YOLOv3 network includes coordinate box loss,confidence loss and classification loss.Each type of loss contributes differently to the network.Therefore,the thesis has adjusted the weights of various types of losses in YOLOv3,and continuously updated the weights to obtain the best value for the loss.Third,in view of the problem that the height and width loss in the coordinate box loss cannot well express the large target frame error and the small target frame error,the thesis proposes an area loss to make up for the shortcomings of width and height loss.Finally,in order to further improve the network performance,the data augmentation method is used in the Eupatorium adenophorum detection method based on YOLOv3 model migration while updating the weight of the loss function in the network to improve the network performance.The experimental results show that among all the detection methods proposed and improved in the thesis,the detection method of Eupatorium adenophorum with the updated loss function weight value has the highest accuracy,and its average precision reaches 67.22%.At the same time,the AP of the method is about 17% higher than that before the improved.The method can well meet the accuracy requirements for the detection task of Eupatorium adenophorum in prevention and control work.The data enhancement method and the loss function weight update method are implemented at the same time,which makes the detection AP value increase by about 19% compared with that before the improvement,further improves the overall network performance,and can better perform the detection task of Eupatorium adenophorum.