Deep Learning Based Tomato Leaf Disease Identification And Research

Tomato is one of the common vegetables in China.Due to the different environmental and growing conditions,tomato plants are susceptible to a variety of pests and diseases during their growth.In order to prevent and control tomato pests and diseases effectively,pest analysis is required.Due to the diversity and complexity of pests and diseases,traditional methods of monitoring and identification through manual observation and statistics are no longer sufficient to meet the needs of modern large-scale agricultural production for pest and disease prevention.Traditional methods of manual detection and identification are not only heavy and inefficient,but are also contingent,unable to achieve real-time scientific detection and do not meet the current long-term development needs of modern agriculture.Traditional machine learning algorithms such as decision trees,random forests and support vector machines are not very accurate for disease identification,so the use of deep learning-based target detection algorithms and semantic segmentation techniques for tomato leaf pest identification is an important research direction in the field of modern smart agriculture.This thesis combines public datasets and tomato leaf disease datasets obtained from the tomato experimental base in Hong Yang Town,Wuhu City,Anhui Province as research objects,and proposes a tomato leaf disease and pest recognition method based on target detection and semantic segmentation to achieve fast and accurate recognition of tomato leaf disease and pest images,the main research work of this thesis is summarised as follows:(1)Firstly,tomato leaf disease data sets were collected and classified and preprocessed.(2)The Shuffle Net V2 lightweight network,YOLOv5-Net,was added to YOLOv5-5.0.The Focus layer in YOLOv5 was removed and the YOLOv5 headers were channel cropped.The experimental results show that the computational volume and weight file size of the YOLOv5-Net model are substantially reduced.(3)Visualisation of the seven trained models,the experimental results show that the recognition rate of Faster RCNN is 92%;YOLO V3 recognition rate is 98.14%;YOLOv4recognition rate is 99.2%;YOLOR recognition rate is 93.17%;YOLOv5 recognition rate is97.98%;YOLOv5-Net recognition rate is 97.67%;The recognition rate of YOLOv5-D is90.95%.(4)To address the problem of low recognition of small target diseases such as early blight and powdery mildew in complex contexts,the deep learning-based semantic segmentation networks U-Net and Seg Net were used,and the reasons for selecting these two networks as the main research components were analysed.By using these two different classification methods,the classification of tomato leaf diseases was achieved and the algorithm was used in classification trials with Res Net neural networks,and the results showed that the algorithm achieved 90.7% correct diagnosis of tomato leaf diseases.