Agricultural Disease Classification Based On Improved Convolutional Neural Network

With the increasingly serious environmental and climatic pollution,the problems of agricultural diseases are becoming more and more prominent,Leaf diseases and trunk diseases have become the focus of attention.At present,agricultural diseases problems usually only rely on eye observation by professionals and subjective experience judgment,which leads to the disease control is prone to misdiagnosis and slow treatment and resulting in a large number of economic losses.Therefore,the most critical part of disease control is to rely on scientific means to quickly judge the disease species,so as to make scientific and effective countermeasures.With the rapid development of computer vision technology,traditional machine learning and deep learning methods have been widely used in the classification and recognition of crop diseases.On the other hand,the agricultural disease images collected by field capture and collection will have high time complexity,difficulty of feature extraction,and poor robustness for the recognition of diversified changes due to the complex growth background environment,unstable illumination and different poses,which makes the classification and recognition more difficult and limits the recognition effect of traditional machine learning methods.And deep learning approach to geometry,illumination and deformation have a certain degree of adaptability,can be extracted in combination with the characteristics of the description,has the advantages of better flexibility and generalization ability.In this thesis,based on experiment and comparative analysis on deep learning framework,main content and research innovations include the following two aspects:1.An improved algorithm based on convolutional neural network is proposed to achieve efficient classification and recognition of original images.Aiming at the problems of long training convergence time and excessive model parameters,this thesis improves the traditional convolutional neural network,and proposes a kind of Inception module integration feature fusion,which combines the SE block structure and the global pooling layer together with the new convolutional neural network recognition model.The multi-scale fusion of the feature data is carried out to improve the accuracy of the disease data set.At last,the globalaverage pooling layer is used to replace the full connection layer to reduce the number of model parameters.The accuracy was 91.7% on the test data set,the number of model parameters was reduced to 57.3MB,and the model parameters and training time were greatly reduced.2.For the data set of apple tree trunk disease images collected in the field,the convolution layer was used to extract the advantages of disease region information,so as to realize the accurate classification of different disease categories.For data collection of the second ring spot images,and shows the characteristics of disease area it is difficult to identify,distinguish between information characteristics,using the training model and combining the adaptive learning training mode,to achieve equilibrium convergence quickly and a single category(tree ring spot)in classification accuracy increased by about 6%,the final results in94.5% accurate data set on the trunk.