Study On Tobacco Leaf Disease Detection Based On Deep Learning

Green control of tobacco diseases is the premise of improving tobacco quality,and accurate identification of diseases is the basis of green control of tobacco.Traditional disease detection methods are time-consuming and laborious,with low accuracy and poor applicability.The detection method based on deep learning does not need to set artificial features,and has higher detection accuracy.However,there are few studies on tobacco diseases using deep learning methods.This paper constructs a tobacco disease data set for research,and uses deep learning methods to study tobacco leaf disease detection.The research content of this paper includes the following three parts :(1)An improved disease detection model based on YOLOv5 network is proposed.In order to solve the problem of weak feature extraction ability of the original model for small target lesions,three improved methods are proposed.The first method is to add a multi-scale enhancement module.Four kinds of dilated convolutions with different expansion rates are added to the Neck layer,and the receptive field of the model is increased without losing the resolution.Then,the prediction accuracy is improved by fusing the feature maps of different receptive fields of four branches.The second is to add a small target detection layer,by adding a large feature map detector to reduce the missed detection rate of small targets;the third is to add a Transformer structure containing a multi-head attention mechanism to obtain feature information from multiple dimensions,and homogenize the possible deviations caused by a single self-attention layer,which improves the performance of the model.Based on three improvement ideas,four improved models are proposed,which are YOLOv5-ME with multiscale enhancement module,YOLOv5-LT with small target detection layer,YOLOv5-ME-LT with small target detection layer and multi-scale enhancement module and YOLOv5-T with Transformer structure.The m AP of the four improved models is better than that of the original model,and the m AP of YOLOv5-ME-LT reaches 91 %,which is much higher than 78 % of the original model.(2)Construct a generative adversarial network based on improved Sin GAN.For the small sample data set of tobacco diseases,the checkerboard effect is mainly alleviated by selecting the upsampling of the bilinear interpolation algorithm.By adding the residual network and the channel attention module,the network depth is increased,the learning ability of the network features is improved,and the quality of the generated images is improved.The experimental results show that the m AP of the improved Sin GAN network enhanced data set after training is96.4 %,which is higher than 85.56 % of the traditional data enhancement and 70.21 % of the original data set.(3)Research and development of tobacco farmer assistant APP.In order to facilitate the management of tobacco fields by tobacco farmers,a small assistant APP for tobacco farmers was developed,which has the functions of tobacco disease encyclopedia and tobacco disease detection.The tobacco leaf disease detection function module uses the improved YOLOv5-MELT model to realize the visualization of tobacco disease detection.At the same time,the system has low equipment requirements and is easier to deploy to the embedded platform.