Research On The Methods Of Grape Leaf Disease Identification Based On Deep Learning

In recent years,due to the particularity of grape cultivation methods and the influence of factors such as global climate change,the frequent and serious occurrence of grape diseases has become a major limiting factor for the safe and high-quality production of grapes.The production method that relies on manual disease identification,diagnosis and decision-making has low efficiency and high labor cost,so it is urgent to realize intelligent diagnosis of disease.As one of the cores of information technology,artificial intelligence technique provides important support for the realization of agricultural informatization and intelligence.With the rapid development of deep learning techniques,the research on the identification,detection,segmentation and counting of crop diseases and insect pests in the agricultural field has made certain progress,which is of great significance for achieving accurate control of diseases and insect pests,reducing economic losses and biological breeding.On the basis of combining domestic and foreign studies,based on different deep learning algorithms,this study has carried out disease identification and detection and other related technology research for a variety of common grape leaf diseases,in order to provide technical support for intelligent identification and detection of grape leaf diseases,the main research contents and related conclusions of this paper are as follows:(1)In order to improve the detection accuracy and efficiency of leaf spots,this study proposes a field grape leaf disease detection method based on improved Faster R-CNN.Res Net-50,which contains inverted residual structure,is used as the feature extraction network,and the feature map extracted by the backbone network is optimized by the attention mechanism.Due to the downsampling operation in backbone,the size of feature map is reduced,a large amount of detail information is lost,and due to the large difference in the scale of lesion in the data set,in order to improve the adaptability of the model to different scale features,a pyramid with strong semantic features at all scales is constructed,and the extracted high-level semantic information is integrated into the underlying detail information to enhance the scale sensitivity of the network.Aiming at the high degree of spots aggregation in some disease images,soft non-maximal suppression strategy and Ro I Align feature mapping method are used to reduce the missed detection rate and improve the bounding box regression accuracy.The experimental results show that the proposed method has 94.27% m AP for the detection of different diseases,which can provide a technical reference for intelligent detection of crop leaf diseases in natural scenarios.(2)The single-stage object detection network has the characteristics of high modularity and flexibility,but the adaptability of the network is not strong,and the problem of unbalanced training samples in the data set is very easy to produce network deviation,resulting in weak generalization ability,poor robustness,low level and quality of accuracy indicators of the model,which cannot be applied to larger-scale and more complex data sets in the natural environment.Aiming at above problems,a grape leaf spot detection model based on the optimization of the YOLOX object detection network is proposed,which optimizes the CSPDarknet and the normalized operation of channels to avoid the deviation of the model during learning process.In order to verify the effectiveness of the model,the effect of YOLOX-NAM with other corresponding object detection networks is compared.The experimental results show that the m AP of the YOLOX-NAM on different grape leaf spots is 92.31%,and the AP value for powdery mildew is increased by 8% compared with that before optimization,which greatly improves the adaptation of network to unbalanced samples without increasing the parameter amount.The m AP index obtained by the network diseases spot detection is of higher quality.(3)It is difficult to distinguish the texture,color and other information of disease spots on grape leaves in the field,and the leaf background is complex and irregular.Aiming at the problem that the shallow network cannot meet the extraction of disease spot features,a deep learning model CA-ENet based on convolutional neural network is proposed.The Coordinate Attention Block(CAB)module is integrated into the Efficient Net-B4 network architecture to realize the integration of feature channels and spatial information attention,and to strengthen the model’s ability to learn important features of the lesion area.The experimental results show that the accuracy of proposed model is 98.33%,and the average F1-score can reach 0.983,which can effectively identify different grape leaf diseases under complex conditions,and can also provide reference for the research on other crop diseases identification.(4)Aiming at the problem that the commonly used convolutional neural network has a large amount of parameters and is difficult to deploy,a lightweight model of local cross-channel interactive attention mechanism is proposed.Based on Shuffle Net lightweight network and efficient channel attention module,the model’s ability to extract fine-grained features of lesions is enhanced while reducing the amount of parameters.Lightweight models ECA-SNet with different scales are obtained by simplifying the network layer structure.The experimental results show that the model parameters of ECA-SNet 0.5× are only 1/4 of Shuffle Net-v2 1.0×,but the recognition accuracy of ECA-SNet is increased by3.3 percentage points,reaching 98.50%,which is significantly better than other commonly used lightweight models and lays a foundation for the terminal deployment of lightweight models.