基于OctConv的DCNN在遥感图像场景分类中的应用

Application of DCNN based on OctConv in scene classification of remote sensing images

传统卷积神经网络(convolutional neural network,CNN)模型在遥感场景分类中存在大量的空间特征信息冗余,这极大的影响了模型的分类精度和运行效率,针对此问题提出一种基于Octave卷积(octave convolution,OctConv)的深度卷积神经网络(DCNN)模型.首先将卷积层输出的特征图根据频率分解为高低频两部分,并采用全局平均池化将特征映射信息量较少的低频部分压缩为当前尺寸的四分之一,然后使用OctConv替换传统卷积操作,实现高低频特征的自我更新和信息交互,最后引入迁移学习用于提升模型的鲁棒性以及弥补训练样本不足的问题.实验证明该方法在UC_merced_Land_Use公开数据集下能够达到99.25%的分类精度,相较于同类型方法提高了2个百分点,表明该方法的优越性以及有效性。

The traditional convolutional neural network(CNN) model has a large amount of spatial feature information redundancy in remote sensing scene classification, which greatly affects the classification accuracy and operational efficiency of the model. In view of this problem, the paper proposes a DCNN model based on octave convolutional(OctConv). Firstly, the feature map outputted by the convolutional layer is decomposed into two parts according to the frequency, and using global average pooling to compress the low-frequency part with less feature mapping information into a quarter of the current size, then using OctConv to replace the traditional convolution operation, to achieve high-low frequency feature self-renewal and information interaction, finally, introducing transfer learning to improve the robustness of the model and making up for the lack of data. The experimental results show that the proposed method can achieve 99.25% classification accuracy under the UC_merced_Land_Use public data set, which is 2% higher than the same type method, which shows the superiority and effectiveness of the method.