Apple Leaf Disease Recognition Model Based On Lightweight Convolutional Neural Network

Apples are widely planted in China,and apple leaf diseases affect fruit growth and development,not only affecting fruit quality and yield,but also seriously threatening fruit farmers’ economic interests.Accurate identification and scientific control of apple leaf diseases are important to reduce fruit farmers’ economic losses and promote fruit industry development.In recent years,deep learning has been widely researched in crop disease identification,and crop disease identification based on convolutional neural network models has a higher accuracy rate.However,convolutional neural network models usually have complex structures and many parameters,which will occupy a large amount of system resources and are not conducive to the promotion of applications.In this thesis,we take apple leaf disease as the experimental object,and conduct the following research work around the light-weighting method of disease recognition model and disease degree grading:(1)Study the apple leaf disease recognition model based on convolutional neural network.Firstly,the lighter versions of Yolov5 model,Yolov5 s and Yolov5 n,were trained and tested on the constructed apple leaf disease image dataset.Yolov5 n is selected as the experimental benchmark model based on the training and testing results.(2)To study the lightweighting method of convolutional neural network and use it for apple leaf disease identification.In order to reduce the accuracy loss caused by lightweighting,the CBS module and C3 module in Yolov5 n are improved by lightweighting,and the lightweight models Ghost-Yolov5n1 and Ghost-Yolov5n2 are constructed.The experimental results show that compared with Ghost-Yolov5 n,the accuracy loss of GhostYolov5n1 and Ghost-Yolov5n2 is 1.8% and 1.5%,respectively,and the number of parameters is reduced by 17% and 30%,respectively.It is consistent with the research objectives of a more lightweight and efficient model.(3)To study the grading method of apple leaf disease degree.Three apple leaf diseases,including brown spot,gray spot and rust,were investigated,and a preliminary manual estimation was made for each disease level.The disease degree was graded by calculating the ratio of the disease spot detection frame area to the leaf detection frame area as the disease index k1.The disease index k1 was then optimized based on the results of the grading experiments,and a disease index k2 based on the area and number of disease spots was proposed.experimental results showed that the grading accuracy of brown spot,gray spot,and rust using disease index k2 reached 81.2%,80.0%,and 97.0%,respectively.Compared with the disease index k1,the accuracy increased by 3.9%,13.3%,and 3.0%,respectively.(4)Study on the deployment of apple leaf disease recognition models on cell phones.The trained model was deployed to the cell phone terminal,and the rapid recognition of apple leaf diseases was successfully achieved.In summary,this thesis provides an effective technical support for the accurate control of apple leaf diseases by constructing an apple leaf disease recognition model and lightweight improvement,and realizing disease degree grading based on disease spot area and number.