融合注意力机制的木薯叶病害分类方法  

A Classification Method with Fusion Attention Mechanism for Cassava Leaf Diseases

作　　者：王文涛[1,2] 张根 陈大江 徐菡廷 WANG Wentao;ZHANG Gen;CHEN Dajiang;XU Hanting(School of Computer Science,South-Central Minzu University,Wuhan 430074,China;Hubei Engineering Research Center of Intelligent Manufacturing Management for Manufacturing Enterprises,Wuhan 430074,China)

机构地区：[1]中南民族大学计算机科学学院,湖北武汉430074 [2]湖北省制造企业智能管理工程技术研究中心,湖北武汉430074

出　　处：《徐州工程学院学报（自然科学版）》2023年第3期40-48,共9页Journal of Xuzhou Institute of Technology(Natural Sciences Edition)

摘　　要：文章针对木薯叶病害图像因病斑区域小,部分疾病特征相似和易受背景干扰而导致识别准确率低的问题,在研究ResNeXt神经网络的基础上,设计了3种方法以提高模型的准确性和鲁棒性:1)增加了一个数据不连续的掩膜层,以缓解神经网络的过拟合问题;2)引入了多头自注意力机制模块,将卷积神经网络和注意力机制结合起来,对图像的局部和全局特征进行分析,提高了相似疾病之间的可分离性;3)使用焦点损失函数缓解木薯叶病害数据集中的类别不平衡问题.仿真实验表明,改进后的算法在木薯叶病害数据集上具有较好的准确率,同时具有较好的泛化能力,适用于其他植物叶片数据集.In view of the low accuracy of disease recognition caused by the small size of diseased areas,similarities among some disease features,and susceptibility to background interference in cassava leaf disease images,this paper designs the following methods based on the study of the ResNeXt neural network to improve the accuracy and robustness of the model:1)adding a data-discontinuous masking layer to alleviate the overfitting problem of the neural network;2)introducing a multi-head self-attention mechanism module to combine convolutional neural networks and attention mechanisms,analyzing the local and global features of the image,and improving the separability of similar diseases;3)using focal loss functions to mitigate the class imbalance problem in the cassava leaf disease dataset.Simulation experiments show that the improved algorithm has better accuracy on the cassava leaf disease dataset,as well as good generalization capabilities,which is suitable for other plant leaf data sets.

关 键 词：植物病理 图像分类 ResNeXt CNN 多头自注意力机制 焦点损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]