Recognition Of Tomato Disease Degree Based On Visual Transformer

At present,tomato disease is one of the important factors affecting the high and stable yield of tomato.Accurate identification of the degree of disease is crucial for selecting appropriate treatment methods and monitoring the occurrence trend of the disease.However,the existing research mainly focuses on the identification of tomato disease types,and there are relatively few studies on the identification of the severity of the disease.There are two main difficulties in realizing the accurate identification of tomato disease degree:(1)In the disease data sample,due to the labeling error in the actual labeling process,there is a lot of noise in the sample library,which reduces the robustness of the disease degree identification model;(2)The feature similarity between different disease degrees of the same disease is high,and existing methods are often difficult to distinguish the fine-grained differences,resulting in low disease degree recognition accuracy.Therefore,this study focuses on these two difficulties.First,aiming at the problem of data noise,an improved confidence learning framework based on consistency regularization is proposed,which improves the robustness of the model in noise samples and provides support for the next step of disease degree identification.A multi-granularity feature extraction model based on visual Transformer proposed to accurately identify the fine-grained differences between disease degrees.The specific work is as follows:(1)An improved confidence learning framework based on consistency regularization is proposed for the problem of noise processing in the disease sample library.Traditional belief learning has the problems of insufficient use of noisy data information and non-iterative training methods.Therefore,this paper improves confidence learning based on consistency regularization and designs an iterative noise learning framework.Experiments show that the improved noise learning framework reduces the impact of noisy data on the model,and further improves the recognition accuracy of the model on noisy data.The method in this paper achieves 87.76%accuracy on the noisy tomato disease data set,which is 1.2%higher than that before improvement.(2)Aiming at the problem that the fine-grained difference in disease degree is difficult to distinguish,a multi-granularity feature extraction model based on visual Transformer proposed.The Transformer Encoder with multi-granularity feature scale and multi-layer enhanced residual used to extract rich feature semantic information of the image,and a feature selection module and contrast loss added to help distinguish fine-grained image differences.Full comparison and ablation experiments carried out on tomato disease data to verify the superiority of the method.The accuracy of the method in this paper for tomato disease reached 88.24%.(3)Integrate the noise learning framework with the multi-granularity feature extraction model to construct a tomato disease degree identification model.By identifying the fine-grained differences between the disease degrees and improving the robustness of the model to noise data,the tomato disease degree recognition model is finally realized.The recognition accuracy is 90.82%,which is better than that of existing methods.The method studied in this paper has been demonstrated and applied in the Shandong Yellow River Delta Agricultural High Area and Modern Agriculture Demonstration Area.