SDA-Net: a detector for small, densely distributed, and arbitrary-directional ships in remote sensing images

摘要

Ships in remote sensing images are usually arranged in arbitrary direction, with small objects and dense distribution, and are easily interfered by background noise, which makes the existing object detection methods have a certain missed detection rate and false detection rate in this kind of complex scene. In order to solve the above problems, a neural network for ship detection is proposed. The network is composed of symmetrical deformation feature pyramid network, multi-scale attention network, multi-scale local context network and rotation branch. Firstly, in order to fully extract the features of ships with arbitrary directions and small objects in remote sensing images, an adaptive deformable convolution unit is designed and a symmetrical deformation feature pyramid network is constructed with it as core. Secondly, a multi-scale attention module is designed to guide the network to focus on ship areas of different scales and suppress background noise. Then, a multi-scale local context network is designed to capture the correlation between ships and local nearby objects, and to learn multi-scale co-occurrence features. Finally, a rotation bounding box is generated by the rotation branch, which is used to mark ships in arbitrary direction. The experimental results on two remote sensing public datasets DOTA, HRSC2016 show that the proposed method achieves state-of-the-art accuracy.