摘要
基于机器学习和深度学习的叶分类图像模型过于关注某一类特征而忽视了模型的泛化能力,导致叶片的分类准确度不高、时间开销大.针对这一问题,在研究密集卷积神经网络的基础上,以苹果叶为对象,通过以下几种方法来提高检测效果:(1)增加一个数据不连续掩模层,以缓解训练神经网络时的过拟合现象;(2)使用广义平均池化改造原有池化方法,以增大输入特征的对比度,专注于输入特征图突出的部分,更好地利用来自卷积层输出张量的信息;(3)使用基于标签平滑(LableSmoothing)损失函数防止模型训练时过度拟合.仿真实验表明:改进后的算法不仅可发现原有数据集中存在的同一种病叶的错误标签分类问题,同时提高了整个苹果叶的检测效果.
In plant pathology research,the image classification model of plant leaves based on machine learning and deep learning pays too much attention to a certain kind of features and ignores the generalization ability of the model,which leads to low classification accuracy and high time cost of plant leaves. Aim to this problem,based on the study of dense convolutional neural network,apple leaves are used as models to improve the detection effect by using the following methods :(1)A data discontinuity mask layer is added to alleviate the over fitting phenomenon when training neural network;(2)The generalized average pooling is used to improve the original pooling method to increase the contrast of the input features,focus on the prominent part of the input feature map,and make better use of the information from the output tensor of the convolution layer;(3)Lable smoothing loss function is used to prevent over fitting during model training. Simulation results show that the improved algorithm can not only find the wrong label classification problem of the same kind of diseased leaves in the original data set,but also improve the detection effect of the whole apple leaves.
作者
王文涛
柳鸣
赵志伟
王嘉鑫
WANG Wentao;LIU Ming;ZHAO zhiwei;WANG Jiaxing(College of Computer Science,South-Central University for Nationalities,Wuhan 430074,China)
出处
《中南民族大学学报(自然科学版)》
CAS
北大核心
2022年第1期71-78,共8页
Journal of South-Central University for Nationalities:Natural Science Edition
基金
教育部产学研合作协同育人项目(201902214013)。
关键词
植物病理
图像分类
数据增强
池化策略
标签平滑
plant pathology
image classification
data augmentation
pooling strategy
label smoothing