Research On Crop Diseases Image Classification Based On Deep Learning

Crop diseases often lead to significant reductions in crop yield and quality.Therefore,quickly and accurately identifying crop diseases can help to improve crop yield and quality by taking the correct and effective treatment measures.Traditional visual observation methods and machine learning-based crop disease classification methods suffer from subjectivity dependence and unsatisfactory accuracy.Currently,convolutional neural networks,as a representative algorithm of deep learning algorithms,can automatically extract disease features,which avoids the problems associated with relying on subjective experience to extract features,and achieves better classification accuracy in crop disease image classification tasks.However,convolutional neural networks usually focus on intra-class distinctive features,ignoring subtle inter-class differential features,which leads to decreased accuracy of disease classification.In addition,deep learning-based methods require large amounts of labeled data to achieve excellent generalization performance,while the sparse labeled data of crop disease images in real-world environments leads to decreased generalization of the model.To address these two problems,this thesis investigates a deep learning-based crop disease image classification method using crop leaf diseases collected under controlled environments and in real environments in the field.The main contents are as follows:(1)To address the problem that convolutional neural networks ignore inter-class subtle differential features leading to reduced classification accuracy,a multi-granularity feature aggregation method based on self-attention mechanism and spatial reasoning is proposed for fine-grained crop disease classification.The method introduces a pixel-level feature selfattention module to capture fine-grained discrimination cues of disease classes,uses a blocklevel feature self-attention module to improve discrimination of different crop species features,and uses a spatial reasoning module to further improve discrimination of disease and species features.Experimental results on the PDR2018,FGVC8 and Plant Doc datasets show that the method is not only able to improve classification accuracy,but also has low complexity.(2)To address the problem of sparse labeled crop disease image data in real-world environments leading to reduced model generalization ability,a crop disease classification method based on self-supervised pre-training is proposed and eventually used for crop disease severity assessment.The method introduces a self-supervised pre-training model to improve model generalization and uses an Asymmetric loss function to optimize the model in order to achieve improved crop disease classification accuracy.Experimental results on Plant Village,PDR2018,FGVC8 and Apple datasets show that the method is effective in improving the model classification accuracy and generalization ability.Taking apple rust severity as an example,the method achieved high accuracy.In summary,the crop disease image classification method proposed in this thesis can effectively improve the accuracy and model generalization ability of crop disease classification under the actual growing environment,and has achieved high accuracy in disease severity assessment.The experimental results show that the method can provide theoretical support for crop disease image classification,and have certain theoretical significance and research value.