Image Recognition Of Crop Diseases Based On Convolution Neural Network

Agriculture is an indispensable basic industry in the construction and development of national economy.As the main product of agriculture,crop development and its importance,but the annual crop disease problem will cause serious loss of crop yield.At present,artificial diagnosis is the most common method to deal with crop diseases.Due to the limitation of time and space in the process of implementation,the diagnosis and treatment of crop diseases in a large area is not timely.Because of this problem,the diagnosis method based on crop disease image recognition is also developing rapidly.In recent years,most of the models used by the leading participants in the ILSVRC competition are based on the convolution neural network(CNN)model and optimization model.Because of the advantages of CNN such as sharing weights and self-selection of feature training weights,the effect of CNN in the field of image recognition is remarkable.In order to improve the accuracy and generalization of crop disease image recognition,the research of crop disease image recognition based on CNN is carried out.The main work is as follows:(1)Aiming at crop disease image recognition under the condition of small samples and unbalanced samples,the pre-training models such as Inception and ResNet are trained by using the auxiliary data set ImageNet,and then the network structure and characteristic parameters of the pre-training model are migrated by using the migration learning method to train the method of crop disease recognition.Finally,focal loss method and group normalization algorithm optimization model are integrated.Improve the recognition rate.Then,10 test sets of crop diseases were tested,and the results were compared with those obtained by the unoptimized model and traditional classification methods.Finally,the feasibility and validity of this classification method were verified.(2)To solve the problem of crop disease image recognition with high similarity between categories,the convolution feature matrix of each inception block in Inception-v3 structure is extracted and fused,and the multi-feature fusion model is trained.Then,the model was tested in different crop disease datasets,and the recognition accuracy was improved by about 1%-10%.It also proved that the classification method has good generalization ability and robustness.