ABSTRACT

A novel deep learning architecture called Attention-Residual block-Embedded Adversarial Networks (AREANs) is proposed in this letter, which can give the robust pixel-wise scene understanding in remote sensing images without any post-processing and additional data. The generator of AREANs, a novel designed encoder-decoder structure network, takes full advantage of Attention-Residual block to learn local-to-global contextual information through semantic and position information enhanced aggregation. To further improve the performance, a patchGAN-based discriminator is applied to train the generator. This training method can not only promote to mine the distinguishable and inherent features from data, but also boost the feature extraction performance of the generator through fine-tuning its parameters. Moreover, the multipath composite loss is proposed as an auxiliary loss in the generator training stage to cope with the class imbalance problem. The comparative experimental results demonstrate that our proposed AREANs can achieve better performance on both Vaihingen and Potsdam datasets.

摘要

在这封信中，提出了一种新颖的深度学习架构，称为注意力残留块嵌入式对抗网络（AREAN），该架构可以在不进行任何后处理和附加数据的情况下，在遥感图像中提供鲁棒的逐像素场景理解。AREAN的生成器是一种新颖设计的编码器-解码器结构网络，它充分利用“注意力残差”块通过语义和位置信息增强的聚合来学习局部到全局上下文信息。为了进一步提高性能，将基于patchGAN的鉴别器用于训练生成器。这种训练方法不仅可以促进从数据中挖掘出可区分的固有特征，而且可以通过微调其参数来提高生成器的特征提取性能。而且，提出将多径复合损耗作为发电机训练阶段的辅助损耗，以解决类不平衡问题。对比实验结果表明，我们提出的AREAN在Vaihingen和Potsdam数据集上都可以实现更好的性能。