Research On Rice Leaf Disease Recognition Algorithm Under Complex Field Conditions

Rice is a widely grown cereal crop throughout the world and is an important source of human nutrition.However,Rice Disease is one of the major factors causing production decline,yield reduction and even famine in some areas.Therefore,the automatic recognition of rice disease under natural field conditions is important for the accurate control of rice growth.At present,although many CNN methods have been proposed by scholars at home and abroad for rice disease,the scale variation and complex background interference existing in real scenes have not been effectively addressed,which affects the recognition accuracy of classification models for disease recognition in natural fields.To address this problem,this thesis proposes a multiscale voting mechanism and a directed self-attention mechanism approach to solve the scale problem and complex background interference.The specific research works are as follows:(1)In response to the limitations of the existing public Rice Leaf Disease(RLD)dataset collection environment,this thesis collected datasets for RLD identification under natural field conditions.It includes 1046 images of rice leaf blight,1053 images of rice blast,1542 images of rice brown spot,823 images of rice sheath blight and 1582 images of healthy leaves.This dataset has the richness and diversity to meet practical scenarios.(2)For the problem of RLD scale variation under field conditions,a multiscale voting strategy for RLD identification is proposed.In this thesis,Efficient Net-b0 is used as the base model and a feature pyramid network is embedded for feature fusion.To further reduce the negative impact of different scale feature layers on the classification results,a multiscale voting strategy in terms of probability distribution was proposed to integrate the decisions from different scales.Each proposed module was carefully validated through an ablation study to demonstrate its effectiveness,and the proposed method was compared with current state-of-the-art algorithms,including SSD,FPN,PANet and Bi FPN.Experimental results have shown that the classification accuracy of our model can reach 90.24%,which is 4.48%higher than that of the original Efficient Net-b0 model and 1.08% higher than that of existing multiscale networks.(3)A directional self-attentive mechanism based on Grad-Cam is proposed for problems such as complex backgrounds that exist in the natural environment of the field.In this thesis,using Efficient Net-b0 as the base model,we first used Grad-Cam to generate the heat maps of all under different labels.Based on this,a directed self-attention mechanism is proposed to reduce the negative impact of complex background on the classification accuracy of the network model.The experimental results demonstrate that the proposed method achieves an accuracy of 89.42%,which is the highest recognition accuracy compared with the mainstream methods such as CBAM,DANet,ECANet,and Sim AM.Finally,this thesis combines the multiscale voting strategy with the directed self-attention method,and the experimental results demonstrate further improvement in the recognition performance of the combined network model,which achieves an accuracy of 90.65%.The research results of this thesis extend the adaptability of deep learning classification models in complex environments in the field and lay the foundation for automatic recognition of RLD under natural field conditions,which can provide decision information for precise control of rice growth and has wide application prospects.