Research On Image Recognition Of Crop Diseases And Model Light-weighting

It is of great research significance and application value to accurately identify and timely prevent crop diseases.At present,image recognition technology based on deep learning has developed rapidly and improved the performance of crop disease recognition,but there are two common problems.On the one hand,in image-based crop disease identification,the shape,region and color of leaf lesions are the main basis for distinguishing different diseases.How to extract more accurate and rich color features and spatial features of leaf lesions has become the key to solving the problems of rough fine-grained classification of disease severity and low recognition accuracy.A large amount of memory and the need for highperformance computing unit support make deployment severely limited.This thesis explores the above-mentioned crop disease image recognition problem,and the main research results include:1.Aiming at the problem of low recognition accuracy,this paper proposes a disease recognition method that combines color mask network and self-attention mechanism(Fusion Color Mask and Self-Attention Network,FCMSAN).FCMSAN consists of two parts,a color mask network and a self-attention network with fusion channel adaptation.The color mask network improves the model’s ability to extract disease color features by learning leaf disease color region information;the self-attention network fused with channel adaptation extracts the lesions features in the global scope,and at the same time adds the location features and channel adaptive features of the lesions to accurately and comprehensively locate Leaf lesions.The outputs of the two networks are fused through the feature transformation fusion module to finally predict the classification output.Experiments show that in the fine-grained identification of 61 types of crop diseases,FCMSAN has a Top-1 classification accuracy of87.97% and an average F1-Score of 0.84.Finally,through visual analysis,the effectiveness of the proposed FCMSAN in disease identification is verified.2.Aiming at the problem of model lightweighting,a disease identification method based on multi-level feature transfer and knowledge distillation(Multi-level Feature Transfer Knowledge Distillation,MFT-KD)is proposed,which consists of three parts: teacher network FCMSAN,improved lightweight convolution The kernel design obtains a lightweight student network and multi-level feature transfer.Multi-level feature transfer includes mid-level knowledge distillation based on attention feature map transfer and deep knowledge distillation based on similarity transfer of predicted values.The knowledge distillation of the middle layer enhances the attention feature map through migration,so that the student network imitates the internal spatial structure of the middle layer of the teacher network;the deep knowledge distillation generates a prediction function by comparing the similarity between the predicted output value of the deep teacher network and the real value of the sample,Dynamically guide students to learn the dark knowledge of teachers’ network online,and improve the accuracy of disease prediction and classification of students’ networks.The top-1 classification accuracy of the student network in the MFT-KD algorithm reaches 87.89%,which is only 0.08% lower than the teacher network FCMSAN recognition performance,but the model parameters are reduced by 86.55%,and the detection time is reduced by 52.60%.It can be seen that the MFT-KD algorithm takes into account good crop disease identification performance and identification efficiency,and is easy to deploy and apply in real-time detection.