摘要
残差网络能够有效地解决卷积神经网络出现的梯度消失问题,应用于高光谱图像分类取得了良好的效果,但简单地堆积残差单元并不能很好地提高模型性能。通道注意力机制能够有区别地处理卷积层输出的特征图,更好地利用对分类有用的特征通道。为了充分利用残差网络及通道注意力机制的特征提取能力,设计适用于高光谱图像分类的残差通道注意力网络。在残差单元中结合卷积层和通道注意力机制,实现对特征通道的重新调整,并在模型中实现局部残差学习和全局残差学习,促进信息传递,增强模型稳定性。实验结果表明,该方法用于Indian Pines数据和University of Pavia数据能够分别取得98.78%和 99.22%的分类精度,在有限数量训练样本的情况下,能够达到较高的分类精度。
The problem of vanishing gradients in convolutional neural network can be solved effectively by residual network, which has achieved good results in hyperspectral image classification. However, simply stacking many residual blocks will fail to achieve better performance. Channel attention mechanism can deal with the feature map of convolution layer flexibly and make better use of useful feature channels for classification. In order to make full use of the feature extraction ability of residual network and channel attention mechanism, a residual channel attention network for hyperspectral image classification is designed. The feature channels are readjusted in the residual block which combining convolution layer and channel attention mechanism. Local residual learning and global residual learning are realized to promote information transmission and network stability. The experimental results show that the proposed method can achieve 98.78% and 99.22% classification accuracies for Indian Pines data and University of Pavia data respectively. High classification accuracy can also be achieved with limited number of training samples.
作者
魏祥坡
余旭初
管凌霄
WEI Xiangpo;YU Xuchu;GUAN Lingxiao(Information Engineering University, Zhengzhou 450001, China;61618 Troops, Beijing 100094, China)
出处
《测绘科学技术学报》
北大核心
2019年第2期161-166,172,共7页
Journal of Geomatics Science and Technology
基金
国家自然科学基金项目(41801388)
河南省科技攻关计划项目(152102210014
182102210148)
关键词
高光谱图像
分类
残差网络
通道注意力
残差通道注意力网络
hyperspectral image
classification
residual network
channel attention
residual channel attention network
