Dynamic Monitoring Of Grape Leaf Disease Based On Deep Learning

In recent years,the wine making industry has developed steadily and the scale of the grape growing industry has gradually increased in China.In the process of grape growth,diseases have been become one of the important factors that restrict its high yield,high quality and high efficiency,which would cause the grapes metamorphism and the yield dropping sharply.Large-scale,high-density and clustered planting patterns of wine grapes and table grapes led to great challenges to the prevention and treatment of infectious diseases.The traditional detection and identification of disease were mostly identified by experienced experts in the field,which would need a long circle and was a waste of time and manpower and was low efficiency.The rapid and accurate identification of disease based on deep learning was of great significance for effective prevention of diseases.In this paper,we presented a dynamic disease monitoring method for wine grape,which inferred whether the disease had existed not only by the disease classifier but also by status changing observed over time from sequential images.The method of this paper can eliminate some false diseases,and the reliability of disease monitoring was improved,which can be applied to continuous online monitoring of grape diseases under natural conditions.In the aspect of image detection and tracking of time series blade,the grape leaves in the first frame of the video were detected by Faster R-CNN every day,then,tracked them in the following frames to find out the frontal snaps of leaves.In terms of tracking,the detected blade was tracked by improved Kalman filter to obtain the front image of the blade.In order to realize multi-blade tracking and solve the problem of tracking failure caused by occlusion,this paper proposed a new tracking method based on Kalman filter and Hungarian algorithm.This method combined motion measurement and depth appearance information to match the tracking target.The Markov distance was used for motion matching,and the minimum cosine distance was used for appearance matching.Secondly,the frontal image of the blade of different dates was matched by SIFT(scale-invariant feature transform),and a set of growth sequence images of the same blade was found.Then the disease recognition was performed by the deep learning technique in the sequence image.Finally,the presence of the diseases were confirmed by monitoring changes in the relative area of the lesion on the leaf sequence image or whether the number of lesions were increased.The experimental results showed that the average multi-target tracking accuracy of the tracking algorithm is 73.6%,and the multi-target tracking accuracy is 74.6%.The traditional tracking algorithms for model color features are 14.3% and 61.3%,respectively.The accuracy of blade matching based on SIFT features is 90.9% when identifying the same blade.In disease monitoring,the false alarm comprehensive elimination rate(Matthews correlation coefficient)is reached.84.3%.In terms of disease saliency detection,this paper constructed a 4×5 grid neural networks,which detected leaf disease significantly.The results showed that F? is 0.916,the absolute difference MAE is 0.048 in simple background.In complex background F? is 0.754,the absolute difference MAE is 0.126.