摘要
利用深度神经网络对小儿肺炎图片进行识别分类,以提高诊断的准确性和自动性。本研究利用融合了注意力机制和残差机制的预训练模型DenseNet121对特征进行训练。对网络结构加入全局平均池化层和Dropout层以防止过拟合,采用交叉熵损失函数以避免学习速率降低及梯度弥散问题,利用迁移学习减少训练参数从而节省训练时间,同时对训练数据做了数据增强。该成果提高了小儿肺炎诊断的识别率。
In order to improve the accuracy and automaticity of diagnosis,deep neural network is used to recognize and classify the pictures of pneumonia in children.In this study,densenet121,a pre training model combining attention mechanism and residual mechanism,is used to train features.The global average pooling layer and dropout layer are added to the network structure to prevent over fitting,the cross entropy loss function is used to avoid the reduction of learning rate and gradient dispersion,the transfer learning is used to reduce training parameters to save training time,and the training data is enhanced.The results improve the recognition rate of pneumonia diagnosis in children.
作者
章广能
ZHANG Guang-neng(School of network education,Southwest University,Chongqing 400715;Haimani Pharmaceutical Co.,Ltd.,Chongqing 401121)
出处
《电脑与信息技术》
2021年第2期31-33,共3页
Computer and Information Technology
关键词
池化层
过拟合
交叉熵
梯度弥散
Pooling layer
over fitting
cross entropy
gradient dispersion