Multi-Level Alignment Network for Cross-Domain Ship Detection

Ship detection is an important research topic in the field of remote sensing. Compared with optical methods, Synthetic Aperture Radar (SAR) ship can penetrate clouds to detect hidden ships all-day and all-weather. Currently, state-of-the-art methods exploit convolutional neural networks train detectors, which require a considerable labeled dataset. However, it difficult label SAR images because expensive labor well-trained experts. To address above limitations, this paper explores cross-domain task, adapts detector from unlabeled images. There significant visual difference between achieve detection, multi-level alignment network, includes image-level, convolution-level, instance-level, proposed reduce large domain shift. First, image-level exploits generative adversarial generate Then, generated real are used detector. further minimize distribution shift, integrates convolution-level instance-level alignment. Convolution-level trains classifier on each activation features, minimizes distance learn domain-invariant features. Instance-level reduces shift features extracted region proposals. The entire network trained end-to-end its effectiveness proved multiple datasets.