Research Of Plant Disease Recognition Based On MobileNeXt

Plant disease recognition is an important prerequisite for effective disease control and plant protection.The rapid development of apple planting in our country is a pillar industry of agricultural economy,which has made great contribution to increasing farmers’ income and promoted the process of poverty alleviation consolidation and rural industry revitalization.However,the yield and quality of apple are easily affected by apple diseases,which will cause serious economic losses and even hinder the healthy development of the apple industry.At present,deep learning methods have been widely applied in the field of plant disease recognition,but the traditional convolutional neural network is difficult to meet the real-time requirements in practical applications due to its large number of parameters,difficulty in storage and high computational complexity.Moreover,apple leaf disease recognition under outdoor conditions has the problems of complex recognition background and low recognition accuracy.Therefore,an apple leaf disease recognition method based on improved lightweight convolutional neural network Mobile Ne Xt was proposed in this paper,and the recognition of four kinds of apple leaf diseases images with complex background was completed.The main contents of this paper are as follows:(1)Apple leaf disease images collected by Northwest A & F University were screened,and four diseases,namely spotted deciduous leaf disease,gray spot disease,Mosaic disease and rust spot disease,were selected as research objects.Each disease contained both simple background images and complex background images,which were collected under different weather conditions,and the image information was more abundant and diverse.In order to prevent the phenomenon of network overfitting caused by insufficient training samples,the data enhancement method was used to expand the number of samples and increase the diversity of training samples,so as to construct the apple leaf disease data set simulating the real environment.(2)Based on the lightweight convolutional neural network model,an improved method on Mobile Ne Xt for apple leaf disease recognition was proposed.In order to improve the feature learning ability of the model on the complex background image,the coordination attention mechanism CA was introduced to help the model distinguish and recognize the target from the complex background by capturing channel information and location information at the same time.The Inception structure was combined to increase the depth and width of the network to improve the accuracy of model identification.Mixed use of depth separable convolution and Ghost convolution,as far as possible to reduce the improvement of the number of parameters;Finally,based on the core structure of Mobile Ne Xt network sandglass,two kinds of SCI(Sandglasss-CA-Inception)structures are designed based on the above three points.(3)Shuffle Net series and Mobile Net series lightweight models were selected for comparative experiments to verify the performance ability of Mobile Ne Xt on apple leaf disease data set with complex background;SE,CBAM and CA attention modules were added to Mobile Ne Xt respectively to verify the superiority of coordinated attention mechanism.Multiple groups of ablation experiments were conducted for the improved part.The results showed that the Top-1 accuracy and average accuracy of the model using SCI＿2 structure were respectively increased by 1.23% and 1.18%,compared with the basic model,while the number of parameters was only increased by 0.62 M,which proved the effectiveness of the improved model.It provides a solution for apple leaf disease recognition in real environment.