摘要
提出了一种基于改进自适应矩估计(adaptive moment estimation,Adam)算法优化器的卷积神经网络(convolutional neural networks,CNN)电子显微镜(电镜)医学图像分类方法。该方法根据卷积神经网络数据迭代的特点,采用具有下降趋势的幂指数学习率改进策略,通过添加修正因子,将上一阶段的梯度值与当前梯度值进行对比、调节,通过梯度值衰减来逐次更新学习率的大小,实现优化器学习率的自适应变化,改善CNN网络模型的收敛性能,实现医学电镜图像的分类。实验结果表明,相比经典的Adam优化器分类方法,改进方法能提高电镜医学图像分类算法的精度,最大分类精度可以到达92%,同时减小图像样本在分类时出现的迭代振荡、分类稳定性不足等现象。
A new method of convolution neural network electron microscope images classification method based on improved adaptive moment estimation(Adam)optimizer is proposed.The method combines with the characteristics of convolution neural network data iteration,and proposes an improved strategy of power exponential learning rate with a decreasing trend,and updates the learning rate successively through gradient value decay,so that the learning rate of the optimizer can realize adaptive change and electron microscope images can be classified.In the improved method,by adding correction factors,combining the gradient value of the previous stage to compares and adjusts with the current gradient value,so that the target requirements can be realized and the convergence performance of the network model is changed.The experimental results show that the new algorithm can reduce the problems of the original algorithm in image classification compared with the improved algorithm,the maximum classification accuracy of the new algorithm can reach about 92%,the oscillation of the iterative curve and the insufficient classification stability is reduced significantly.
作者
汪友明
徐攀峰
Wang Youming;Xu Panfeng(School of Automation,Xi'an University of Posts and Telecommunications,Xi'an 710121,China)
出处
《西安邮电大学学报》
2019年第5期26-33,共8页
Journal of Xi’an University of Posts and Telecommunications
基金
陕西省重点研发计划资助项目(2019GY-086)
关键词
电镜医学图像
Adam优化器
幂指数学习率
electron microscope image
Adam optimizer
power exponential learning rate
