Attention-Based Convolutional Networks for Ship Detection in High-Resolution Remote Sensing Images

Environmental information, like sea-land distribution, plays an important role in detecting ships from remote sensing images. However, the huge scale difference between environments and ship targets makes current CNN-based detection models hard to learn large-scale geographical information and focus on small targets at the same time. We propose an attention-based method by adding a Fully Convolutional Networks (FCN) to a detection networks as an attention branch to extract environmental features. Within a detection phase, the target detection branch is guided by the attention branch so as to focus on the potential target locations while in a training phase, the losses of other locations are simply ignored. We test our method on a public available remote sensing target detection dataset: LEVIR. By taking the classical Single Shot MultiBox Detector (SSD) as baseline, our method improves its detection accuracy in ship detection task while with an acceptable computational overhead.