Research On Grape Leaf Disease Identification Method Based On Deep Learning

Grape is rich in nutritional value and has a wide planting area in China.Therefore,grape industry has become one of the main industries for farmers to increase their income.However,in the process of grape planting,it is easy to be affected by bacteria,environment and other factors,resulting in diseases,most of which are on grape leaves.Therefore,the use of computer vision technology to identify grape leaf diseases and provide users with timely and accurate information and control measures is of great significance to improve grape yield and quality.In this study,eight common diseases of grape leaves were identified by image enlargement,image segmentation and recognition.The main work of this paper is as follows:(1)Image preprocessing method of grape leaf disease.Through the analysis of the characteristics of grape leaf disease,aiming at the problem that the diseased samples are not evenly distributed,the model is prone to over fitting.The data set is expanded mainly by image rotation and mirror image,brightness adjustment and image enhancement based on antagonistic neural network.The image is zoomed by bilinear interpolation method.After the expansion,there are 4000 data sets.Labelme was used to label the region of grape leaves,which laid a foundation for the extraction of grape leaves.(2)Study on the extraction method of grape leaves.In order to improve the training efficiency of the network model and reduce the influence of complex background on grape disease identification,the paper first pre trains the Mask R-CNN network on the Image Net data set to get the model parameters,and then fine tunes the model on the data set of this study to realize the parameter migration of the model.Using Mask R-CNN to extract grape leaves,and then according to the proportion of leaf area in the filled image,decide whether to scale the leaves.When the ratio is less than 50%,bilinear interpolation is used again to restore the leaf area to its original size.The results showed that the AP of normal grape single leaf,normal grape multi leaf,disease grape single leaf and disease grape multi leaf were 0.9248,0.9129,0.9042 and 0.8924,respectively.(3)Study on the identification method of grape leaf diseases.In order to improve the recognition accuracy of grape leaf disease,the multi-scale convolution kernel combination is introduced to improve the response of Res Net bottom layer to different scale features,and the squeeze and exception networks(SENet)is added to improve the feature extraction ability of the network.The experiment shows that the proposed Multi-Scale Res Net's recognition accuracy is 90.83%.This paper develops and implements a small program to identify grape leaf disease,which is convenient to provide users with grape leaf disease knowledge,disease recognition and disease control information.