Research And Implementation Of Vision Transformer-based Algorithm For Plant Disease Detection In Complex Environments

In modern agricultural production,plant diseases can seriously reduce crop yields and thus threaten food security.In recent years,with the development of artificial intelligence technology,many scholars have used computer vision combined with deep learning solutions applied to detect plant diseases.However,there are still some problems in agricultural production application scenarios.In this paper,we take plant disease detection in complex environments as the core problem,and improve the accuracy and robustness of plant disease detection in com-plex environments by building a deep learning detection model.The main work of this paper is summarized as follows:1.Research on plant disease detection in complex environments.In this paper,the Convo-lutional Swin Transformer(CST)series model and ACmixed Alex Net network structure are de-signed for disease detection in complex environments.Very high robustness is obtained by com-bining convolutional neural networks and self-attentive mechanisms.Among them,the highest detection accuracy of 0.909 was obtained for the CST series model and 0.904 for ACmixed Alex Net in the cucumber dataset.The highest detection accuracy of 0.922 was obtained by the CST series model and 0.940 by ACmixed Alex Net in the banana dataset.2.Research on the effect of image noise on plant disease detection.To address this prob-lem,the CST series models are proposed in this paper,and their performance in detecting plant diseases under the influence of noise is tested in experiments.Firstly,10%,20% and 30% of salt noise are added to the image data obtained by shooting in real environment.The CST series models were able to maintain the detection accuracy of 0.8 under the cucumber dataset with increased noise.In detecting banana dataset with increased noise,even with 30% increased salt noise still could obtain 0.9 detection accuracy.3.Research on the effect of light intensity on plant disease detection.To address this problem,the ACmixed Alex Net model is proposed in this paper.First,the brightness of the images in the banana dataset is increased and decreased by 10%,20% and 30% to obtain a total of six different image brightnesses.The highest detection accuracy of 0.956 is obtained by ACmixed Alex Net with 30% brightness increase under this data.At the same time,the detection accuracy of 0.9 was still achieved under the condition of reducing the brightness by 30%.4.Usability study in mobile devices.In this paper,the algorithm model will be deployed in a mobile device in the form of Android software.First,the trained CST series model and ACmixed Alex Net model are packaged into the Android software using Py Torch Mobile.Sub-sequently,the performance of the deployed software in detecting plant diseases is tested in ex-periments,and the overall detection accuracy remains above 0.8 after the model is pruned and optimized in mobile.