基于深度森林的多级特征融合SAR目标识别

大多数传统的合成孔径雷达(SAR)目标识别方法仅仅使用了单一的幅度特征，但是由于斑点噪声的存在，仅仅使用幅度特征会限制识别的性能。为了进一步提高SAR目标识别的性能，该文提出了一个基于深度森林的多级特征融合SAR目标识别方法。首先，在特征提取阶段，提取了多级幅度特征和多级密集尺度不变特征变换(Dense-SIFT)特征。幅度特征反映了目标反射强度，Dense-SIFT特征描述了目标的结构特征。而多级特征可以从局部到全局表征目标。随后，为了更完整、充分地反映SAR目标信息，借鉴深度森林的思想对多级幅度特征和多级Dense-SIFT特征进行联合利用。一方面通过堆叠的方式不断将多级幅度特征和多级Dense-SIFT特征进行融合，另一方面通过逐层的特征变换挖掘深层信息。最后利用得到的深层融合特征对目标进行识别任务。该文在MSTAR数据集上进行对比实验，实验结果表明所提算法在性能方面取得了提升，且其性能对超参数设置具有一定的鲁棒性。

Multi-level Feature Fusion SAR Automatic Target Recognition Based on Deep Forest

In most of Synthetic Aperture Radar (SAR) target recognition methods, only the amplitude feature, i.e., intensity of pixels, is used to recognize targets. Nevertheless, due to the speckle noise, only using the amplitude feature will affect the recognition performance. For further improving the recognition performance, in this paper, a novel multi-level feature fusion target recognition method based on deep forest for SAR images is proposed. At First, in the feature extraction step, two kinds of features, i.e., the multi-level amplitude feature and the multi-level Dense Scale-Invariant Feature Transform (Dense-SIFT) feature are extracted. The amplitude feature describes intensity information and the Dense-SIFT feature describes structure information. Furthermore, for each feature, its corresponding multi-level features are extracted to represent target information from local to global. Then, for reflecting target information more comprehensive and sufficient, the multi-level amplitude feature and the multi-level Dense-SIFT feature are jointly utilized profiting from the idea of deep forest. On the one hand, the cascade structure can fusion multi-level amplitude feature and the multi-level Dense-SIFT feature steadily. On the other hand, the deep feature representation can be mined by layer-by-layer feature transformation. Finally, the fusion feature is used to recognize targets. Experiments on the moving and stationary target acquisition and recognition data show that the proposed method is an effective target recognition method, and the recognition performance is robust to the hyper-parameters.