Research On Citrus Disease Recognition Based On Deep Learning

With the development of artificial intelligence technology,researchers in the field of computer vision and agricultural engineering began to combine artificial intelligence technology with agricultural production knowledge to accelerate the development of intelligent agricultural technology,so as to achieve the purpose of automation,intelligent management and production of agricultural production process.In the field of citrus,in recent years,many researchers have proposed many citrus disease recognition methods based on machine learning and deep learning,and most of them have achieved good results.However,in the current citrus disease recognition methods,most of them are aimed at the citrus disease leaves captured in the experimental environment,ignoring the problems of occlusion,different shooting angles and different light intensity in the natural environment,which affect the recognition effect in the actual planting environment,reduce the recognition accuracy of the algorithm for citrus disease,and can’t be well applied to the actual planting activities.In addition,if the symptoms of citrus disease are found in time in the early stage,it has important practical significance in controlling the aggravation of the disease,preventing the spread of the disease and reducing the loss of citrus yield.Therefore,in this paper,the methods of citrus disease recognition and disease degree recognition in complex natural environment was studied and improved.The main research contents and achievements of this paper are as follows:(1)Aiming at the problem of low recognition accuracy of citrus disease image in complex natural environment,this paper proposes a multi-classification citrus disease image recognition method based on resnet34 deep learning model.An improved model named S-Res Net,which is more sensitive to positive features,was proposed by discarding the identity mapping in part of the residual structure of Res Net34.In SRes Net,the low layers features of the citrus disease images are further extracted,and the negative features are reduced.Compared with Res Net34 model,the average recognition accuracy of improved model is improved by 3.9%,and the recognition accuracy of citrus disease image in natural environment has been improved.At the same time,a model named M-Res Net,which can obtain more representative feature,was proposed by replacing the original convolution kernel with several fine-grained convolution kernels to extract deep features in citrus disease images.To avoid the problems of bad training approximation,generalization and robustness in single model,S-Res Net and M-Res Net are fused by model fusion method to obtain a model called FRes Net,which is more adaptable to new samples.The experimental results show that the accuracy rate of F-Res Net can reach 93.6%,and the model fusion method can improve the classification performance of neural network for citrus disease images in complex natural environment.(2)In order to solve the problem that fine-grained features of different classifications of images have subtle different and difficult to distinguish in deep learning,this paper proposes a citrus disease recognition method based on unsupervised clustering image segmentation,deep separable convolution and transfer learning.This paper uses unsupervised clustering image segmentation to extract the salient features of citrus disease image,uses deep separation convolution method to compress the model parameters and volume,and uses transfer learning method to reuse model parameters and weight.The accuracy of the final model is 88.79%,which can meet the needs of disease identification in the actual orchard,and has a good prospect.