Identification Of Citrus Canker Based On Convolutional Neural Network

Citrus is an important economic crop in modern agriculture.It is cultivated in a wide range of areas and full of nutrients.Many people are very fond of it.However,the development of citrus cultivation has been constrained by citrus canker for so many years.So the citrus canker has been listed as an important quarantine target by countries around the world.Image recognition technology can achieve high-efficiency recognition of target images.It is also low-cost,environmentally-friendly,and free from the constraints of time and geographical.So it is of great significance for the automatic identification and detection of citrus canker.In this paper,we build a citrus canker disease data through the cooperation with plant protection experts.But there is an imbalance in data by considering the inherent hazards of citrus canker disease,the strictness of methods for preventing and controlling diseases,and other factors.So we preprocessed the dataset through data enhancement and amplification.Then Using the Alexnet model of the convolutional neural network made experiments by comparing with the existing algorithms and traditional image recognition algorithms,which show that the TPR and FPR of the convolutional neural network are superior to those existing algorithms and traditional classification algorithms in the unbalanced data set.In addition,for dealing with the small and medium samples the transfer learning method was introduced.Finally we successfully improved the TPR from 0.9725 to 0.9922,and decreased the FPR from 0.0077 to 0.0017.When the convolutional neural network model is used in the relatively simple application scenario,there will exit the problem of feature redundancy.To solve the problem,this paper throws three methods to improve it.Such as simplifying the network,changing the network structure and pruning of network weights automatically.The experimental results show that the three methods all effectively reduce the network parameters,the disk space occupied by the network and do not reduce the classification ability of the network.The three methods start from different perspectives and do not affect each other.In practical applications,they can be used in combination to achieve the best results.Finally,this paper studies the problem of the recognition effect of fuzzy images on citrus canker and then puts forward the combination of image quality evaluation system IQA-CNN and automatic recognition system to reduce the bad effects of fuzzy image on automatic system.