| The quality of mulberry leaves is directly related to the high quality development of sericulture economy,so it is very important to identify the diseases of mulberry leaves quickly and accurately.With the development and maturity of deep learning technology,its application in plant disease recognition is being explored more and more intensively.In this thesis,compared with several classical object detection networks,YOLOv5 s network with good performance was selected and improved to obtain better detection for mulberry leaf disease.A new R2-YOLOv5 s network was designed by improving the backbone,feature fusion structure,detection head and post-processing mechanism of YOLOv5 s network.A new feature extraction module of C3R2 was constructed using Res2 x,which is a multi-scale residual structure.CBAM attention mechanism is added after each C3R2 module.Convolution pooling combined module MPC was used for downsampling.The CSS is used in the feature fusion structure to enhance the information transfer between the shallow layer and the deep layer.A decoupling head is introduced into the detection head to separate the classification and localization.The DIo U-NMS mechanism is used to replace the traditional NMS.To evaluate the contribution of the above improvements to network performance,a dataset of 6000 images of mulberry leaf disease was prepared.There are 3000 original images,and the rest are processed by image enhancement technology.The diseases involved are tan disease,powdery mildew and sooty blotch.Image enhancement techniques such as rotation,mirror,brightness adjustment and Mosica are adopted.Each image in the data set was accurately labeled with the Labelimg tool under the guidance of sericulture experts.The test experiments with this data set show that the designed R2-YOLOv5 s network has better comprehensive performance than the origin YOLOv5 s network.The R2-YOLOv5 s network has increased the m AP by 3.8%,the precision by2.1%,and the recall by 4.6%.In addition,compared with other networks,R2-YOLOv5 s network has the highest detection accuracy,and fast detection speed,up to 120 FPS.In addition,according to the practical demand of mulberry leaf disease identification,the R2-YOLOv5 s network was compressed by knowledge distillation method and deployed on the mobile terminal.The number of parameters and computation of the obtained R2-YOLOv5 n network is only 28% of that of R2-YOLOv5 s network,the m AP can be up to 91.3%,and the detection speed is about194 FPS.By using NCNN framework,a APP is deployed in mobile terminal,and used to realize Mobile real-time detection of mulberry leaf disease. |
