基于小波-空间高阶特征聚合网络的遥感图像场景分类

Scene Classification of Remote Sensing Images Based on Wavelet-Spatial High-Order Feature Aggregation Network

提出一种小波-空间高阶特征聚合网络(WHFA-Net),该网络可分为小波域特征提取和空间域特征提取两个分支。首先,将Harr小波变换嵌入卷积神经网络(CNNs),保留深度卷积特征的低频分量作为小波深度特征;其次,利用最大池化进行深度特征学习,并将其输出作为空间深度特征;将两分支的深度特征进行向量化后,获取其自相关和互相关高阶深度特征向量,并依次进行特征正规化、特征聚合和特征归一化操作;最后,引入交叉熵损失函数进行端到端网络训练。在NWPU45(NWPU-RESISC45 Dataset)和AID(Aerial Image Dataset)数据集上的实验结果表明:相较于基准网络(VGG-16),本文所提WHFA-Net的场景分类准确率有5.13%~12.12%的提升;与DCCNN、APDC-Net、GBNet、LCNN-BFF、MSCP和Wavelet CNN相比,WHFA-Net的场景分类准确率均有不同程度的提升;通过消融实验验证了各模块和分支的有效性及其性能差异。因此,WHFA-Net可有效且稳定地抽取遥感场景图像不同特征域的高阶聚合特征,并提升场景分类准确率。

This paper proposes a wavelet-spatial high-order feature aggregation network(WHFA-Net) which can be divided into two branches: wavelet domain feature extraction and spatial domain feature extraction. Firstly, a Harr wavelet transform is embedded into convolutional neural networks(CNNs), and low-frequency components of the depth-wise convolutional features are retained as wavelet depth features. Secondly, depth feature learning is performed by the max pooling, and then the output is used as spatial depth features. In addition, the wavelet domain and spatial domain depth features are vectored, and their auto-correlation and cross-correlation high-order depth feature vectors are obtained.Feature regularization, feature aggregation, and feature normalization are then performed in sequence. Finally, a crossentropy loss function is utilized for end-to-end network training. The experimental results on NWPU45(NWPU-RESISC45 Dataset) and AID(Aerial Image Dataset) show that compared with that of the benchmark network(VGG-16), the accuracy of the proposed WHFA-Net in scene classification is improved by 5. 13%-12. 12%.Furthermore, compared with DCCNN, APDC-Net, GBNet, LCNN-BFF, MSCP, and Wavelet CNN, the accuracy of WHFA-Net in scene classification is higher. Additionally, the effectiveness and performance differences of each module and branch are verified through the ablation experiments. Therefore, WHFA-Net can effectively and stably extract the high-order aggregated features of different feature domains in remote sensing scene images and accurately perform scene classification.