Research On Citrus Leaf Disease Identification Based On Convolutional Neural Network

With the vigorous development of artificial intelligence technology,researchers in the field of deep learning and agricultural engineering have also begun to combine artificial intelligence technology with agricultural engineering technology to help promote the development of smart agricultural technology,thereby achieving the informatization and intelligent management and control of agricultural production processes.Purpose.In recent years,many scientific researchers in the field of agriculture have begun to propose many solutions based on traditional machine learning and deep learning for citrus disease classification.These methods can achieve good recognition results.However,most scientific researchers are based on photographing citrus diseased leaves for identification and detection in the laboratory environment,but ignore the different environmental factors that may exist in the natural environment:object occlusion,light intensity,noise interference,shooting angle and other issues,resulting in the obtained results.The experimental results are difficult to apply to actual citrus cultivation and production.In the actual citrus production and planting process,if the occurrence of various citrus diseases can be prevented in time in the early stage of citrus disease,it is very important to help fruit farmers control the deterioration of diseases in time,prevent the spread of diseases and avoid the loss of citrus yield.realistic meaning.(1)According to the problem that the recognition speed of citrus diseases is too slow in complex natural environment,this paper takes the leaves of citrus diseases as the main research object,and provides a citrus disease image detection method based on binary Faster R-CNN.The convolutional neural network VGG-16 is the initial type of network.In this network,pooling and normalization are performed by embedding the RPN layer and the ROI pooling layer.The improved model converts the original fully connected layer neural network into a binary full convolution.Neural network,and then use the greedy algorithm to obtain the local optimal solution,and finally form the binarized Faster R-CNN network model.The test results show that the accuracy of the model for citrus black spot,canker,Huanglongbing,scab,and healthy leaves is87.2%,87.6%,89.8%,86.4%,and 86.6%,respectively,and the m AP is 87.52%.The number of parameters of the latter algorithm is greatly reduced,about 15.3M,the model recognition time is increased by 0.53s,the detection time of each image is 0.31s,and the FLOPs of the algorithm is 2.8×10~9.(2)In order to help fruit farmers more effectively prevent the occurrence and deterioration of Citrus Diseases,this paper classifies the citrus disease grade,and proposes a citrus disease grade recognition method based on unsupervised clustering image segmentation,convolution normalization and transfer learning.The K-means unsupervised clustering image segmentation technology is used to extract the salient features in the citrus disease map,then the model is simplified through fine-tuning strategy,and the parameters and weights obtained from the pre training of transfer learning are reused.Finally,the improved NC-Res Net50 model is obtained.The experimental results show that the accuracy of the improved model is improved by2.28%,and the accuracy of the citrus disease grade identification model is 85.45%,which can meet the needs of disease identification in the actual orchard and timely control the disease.