Detection And Severity Estimation Of Tea Diseases In Natural Scene Images Based On Deep Learning

Tea is a beverage crop with great nutritional value and economic value,and China has a long and splendid history of tea drinking culture.With the expansion of tea planting scale,the prevention and control of tea diseases are facing great challenges.Promoting tea garden disease control to intelligence,precision and efficiency is the key to solve the problem.Deep learning,as an efficient intelligent data and image processing method,can cross spatial and temporal constraints to achieve real-time and off-site detection of tea diseases and play an important role in obtaining tea disease information and monitoring the status of tea plantations.This study takes tea diseases in natural scene images as the research object,and applies deep learning and image processing techniques to tea disease detection and severity estimation in natural scene images,which can achieve accurate detection and localization of tea diseases while identifying tea disease categories and estimating the severity of tea diseases,which is beneficial to the construction of smart tea plantations.The main work and research of this paper are as follows.1.Using Canon EOS 80D SLR camera camera,Iphone 7 and Iphone 11 cell phones to collect images of tea diseases in Miaozhu Village tea garden in Chizhou City,Anhui Province,Tianjing Mountain tea garden in Hefei City,Anhui Province,and Anhui Agricultural University tea garden several times,the images were collected without artificial control of the captured background,and the lens of the collection equipment was about 0.3 m away from the tea leaves.in order to facilitate subsequent processing and meet practical needs,the diseased tea images are preprocessed initially,and the images are labeled using the open source image annotation tool Labelme to construct the original dataset.2.A method for detection and severity estimation of tea leaf blight based on improved Faster R-CNN and VGG16 is proposed to solve the problem of low accuracy of tea disease detection caused by the existing methods under the influence of factors such as light change,shadow,shape change,scale transformation and mutual leaf occlusion in natural environment.Firstly,Retinex algorithm is used to enhance the natural scene tea disease images to reduce the effects of light changes and shadows.Then an improved Faster R-CNN algorithm is given to improve the detection performance of blurred,curled and small diseased leaves for the characteristics of tea disease in natural scenes.The detected diseased leaves are fed into the trained VGG16 network to achieve severity estimation.Experimental results show that the detection accuracy and severity grading accuracy of the method are significantly improved compared with existing depth models and methods.3.A method for detection and severity estimation of multiple tea diseases based on improved Mask R-CNN and ShuffleNetv2 is proposed.The tea background in natural scenes is very complex,with tea leaves shading each other and variable lighting conditions,which poses a great challenge to the accurate detection and segmentation of tea diseases.To be able to accurately detect tea diseases and estimate tea diseases severity in natural scenes,this method introduces the second generation deformable convolution into the feature extraction network of Mask R-CNN instance segmentation algorithm,which can adaptively extract diseased tea features for effective detection of tea diseases and segmentation of diseased leaves.Meanwhile,in order to quickly and accurately grade the severity of multiple tea diseases after segmentation,the method uses ShuffleNetv2 network to train a tea diseases severity estimation model.The experimental results show that the average detection accuracy and severity grading accuracy of the method are significantly improved compared with the existing depth models and methods.