Research On Tomato Disease Recognition Based On Deep Learning

As one of the research areas in precision agriculture,the timely and effective detection of leaf diseases and their diagnosis and prevention will significantly improve the quality and yield of agricultural products,reduce losses and promote the development of green agriculture.In the field of tomato research,many researchers have proposed many identification methods based on machine learning and deep learning for tomato disease identification problem in recent years,and most of them have achieved good identification results.However,existing deep learning-based tomato disease recognition still has the problem of many training parameters and insufficient accuracy,which greatly limits the practical application and deployment of the model;secondly,disease recognition in natural scenes still has certain challenges,ignoring the problems of shading,varying shooting angles,and different light intensities that exist in natural environments,thus affecting the recognition effect of the algorithm in the actual growing environment.Most of the existing research on tomato disease recognition is based on hyperspectral image technology,and most of the research has been conducted on the degree of disease of a single category of tomato diseases.Therefore,in this thesis,we first investigate tomato diseases in simple backgrounds and complex natural scenes,and then investigate tomato disease degree classification,and the main contents and results of the study are as follows:(1)To address the problems that existing simple background tomato disease identification models have high computational complexity,require many parameters,have high memory requirements and limit model deployment.This study proposes a lightweight tomato disease identification model based on knowledge distillation(KD).Firstly,the teacher model and the student model are trained separately,and the most suitable teacher model is selected to guide the student model using the knowledge distillation method.At the same time,a mixed loss function was introduced to uncover the hidden information contained in the information discarded by the teacher model to ensure the accuracy of the model while compressing it as much as possible.The experimental results demonstrate that the number of parameters of the knowledge distillation model with the introduction of the hybrid loss function is only 5.23 M and the number of floating point operations is 0.34 G,and the average recognition rate is still98.87%.(2)To address the problem that the recognition of tomato diseases in natural scenes is not very satisfactory.This study constructs a multi-scale tomato disease recognition model incorporating an attention mechanism.We first incorporate a selective convolutional kernel attention mechanism to the residual module of the teacher network Res Net101,which eliminates useless features and highlights key local regions by attenuating environmental interference noise.The multi-scale Inception structure is then introduced to extract disease features from different visual fields using its multi-scale feature,thus further improving the performance of the model.The experimental results show that the multi-scale tomato disease recognition model incorporating the attention mechanism achieves an average recognition rate of 92.51%,which is 8.86% higher than the original Res Net101 model.(3)To address the problem that disease extent image classification studies require high differentiation of key features in the sample images.In this study,disease images are first pre-processed using the unsupervised K-means algorithm to enhance the disease features in the sample images.Then use the transfer learning method to pre-train large public datasets,and use the obtained weights and parameters to indirectly increase the number of samples during training,break the limitation of insufficient training images,and solve the problem of tomato suffering caused by too few sample images.The problem of low precision or overfitting caused by the low sensitivity of the disease severity classification recognition model to fine-grained features.The experimental results show that the average recognition rate of the model-based migration learning method on the validation set ends up being 88.15%.