基于阶段聚焦损失和并行增广策略的遥感图像场景分类

Remote sensing image scene classification via stage-based focal loss and parallel data augmentation strategy

随着深度学习的不断普及,卷积神经网络已成为遥感图像场景分类的主要手段,然而当前的研究主要集中于多网络主干的信息融合以及注意力机制等领域,在提高分类精度的同时也带来更高的计算复杂度。针对上述问题,分别从改进卷积神经网络损失函数和设计新的样本训练策略两个角度出发,在不增加计算复杂度的前提下,提升卷积神经网络的分类性能。首先,在对传统交叉熵和Focal loss损失函数进行分析的基础上,提出一种阶段聚焦损失函数,该损失函数可以在训练阶段对卷积网络进行有侧重的性能挖掘。其次,设计了一种并行样本训练策略,将采用Gridmask算法增广后的样本图像和原始样本图像,分为两路输入卷积网络进行并行训练,进一步提升卷积网络的分类性能。实验结果表明,所提出的算法分别在AID和NWPU-RESISC45两个大规模数据库上取得了96.72%和93.95%的检测精度,可以显著提升遥感图像场景分类的性能。

With the continuing popularity of deep learning techniques,convolutional neural network(CNN)has become the main tool to solve the remote sensing image scene classification tasks.However,current research interests are highly focused on the topic of how to fuse multi-branch-based CNN and how to apply attention models.Despite that these approaches enhance the classification accuracy markedly;it leads to high computational complexity.In this paper,the above problems are addressed by means of introducing a modified loss function and designing a novel data augmentation strategy,which can significantly improve the classification performance of CNN without increasing the computational complexity.First,a stage-based focal loss function is presented to adaptively mining the hard sample during the training process.Second,a parallel training strategy is conducted to feed the original image samples and samples after Gridmask operation into the sharing CNN separately.Experimental results show that the proposed algorithm achieves 96.72%and 93.95%detection accuracy on two large-scale databases of AID and NWPU-RESISC45,respectively,and can significantly improve the performance of remote sensing image scene classification.