Research On Image Recognition Of Agricultural Disease Based On Deep Learning

Crop diseases and pests are the hidden dangers of food safety in agricultural production for a long time,and over-use of agricultural chemicals will cause excessive residue,and which will seriously affect the health of people.With the application of artificial intelligence in agriculture,the problem of slow efficiency and low accuracy brought by artificial identification of crop diseases has been gradually solved.The best time to prevent and control agricultural diseases in the early stage of agricultural disease development has become an important goal.Deep learning technology has been applied to the identification of agricultural diseases,but there are still problems.For example,imbalance of samples and difficult classification of difficult samples.In view of these problems,this paper proposes several improvement measures of algorithms,establishes crop disease identification network,and uses transfer learning technology.The target recognition technology is applied to the self-made data set,and the task demand of agricultural disease identification is realized.The main research work of this paper is as follows:First,the data set of agricultural diseases is expensive and consuming manpower.In the early stage of the research,small-scale data sets are collected by network collection and field shooting.Then the data is expanded by scaling,horizontal rotation,vertical rotation,angle rotation,brightness adjustment,contrast adjustment and color adjustment.Due to the inconsistent size of the collected images,it is necessary to normalize the size of the image to256 × 256.Secondly,after obtaining the “AI Challenger 2018” data set and self-made data set,qualified pictures are selected to adapt to the training of the network.In addition,we use“Label Img” software to mark two data sets manually,among which 10639 photos are taken in the pre-training data set “AI Challenger 2018”,1555 pictures are made in the self-made data set.By manually marking the upper left corner,label of the target and lower right corner.We can generate the XML format file to train.Thirdly,the paper combines Retia Net and FCOS network,aiming at solving the problem of sample imbalance and difficulty in sample classification,and proposes the improvement of the algorithm.The improved algorithm is based on FCOS network,which is built by the technology of Feature Pyramid Networks and Dilated/Atrous convolution.The network framework can improve the ability of network framework to deal with the problems such as category imbalance and difficult sample classification,improve the sense field and reduce the calculation based on free-anchor algorithm,and improve recall rate and difficult classification of difficult samples.Finally,the algorithm is trained on the “AI Challenger 2018” datasets,and then transfer learning on the self-made datasets to make the algorithm adapt to the identification of agricultural diseases.Through experiments,I found that the experimental results have achieved good expectations,which shows that the improved algorithm can be better adapted to the identification of agricultural diseases.In conclusion,the data set of agricultural diseases and the pretreatment of the data set are made.In addition,the targeted improvement measures are put forward to build the crop disease identification network,and the transfer learning technology is combined.The target recognition technology can be used in the identification of agricultural diseases.It also shows that the transfer learning technology can be used in their research fields,and it has better mobility and universality.